I think the biggest winner of this might be Google. Virtually all the frontier AI labs use TPU. The only one that doesn't use TPU is OpenAI due to the exclusive deal with Microsoft. Given the newly launched Gen 8 TPU this month, it's likely OpenAI will contemplate using TPU too.
Many labs use TPUs, but not exclusively. Most labs need more compute than they can get, and if there's TPU capacity, they'll adapt their systems to be able to run partially on TPUs.
Google is in a different position to others in that they're the only frontier lab with a cloud infra business. It obviously makes sense to sell GPUs on cloud infra as people want to rent them. In that respect Google buys a ton of GPUs to rent out.
What's unclear to me is how much Google uses GPUs for their own stuff. Yes Gemini runs on GPUs now, so that Google can sell Gemini on-prem boxes (recent release announced last week), but is any training or inference for Gemini really happening on GPUs? This is unclear to me. I'd have guessed not given that I thought TPUs were much cheaper to operate, but maybe I'm wrong.
Caveat, I work at Google, but not on anything to do with this. I'm only going on what's in the press for this stuff.
AMD’s software experience is riddled with bugs rendering out of the box training with AMD is impossible. We were hopeful that AMD could emerge as a strong competitor to NVIDIA in training workloads, but, as of today, this is unfortunately not the case. The CUDA moat has yet to be crossed by AMD due to AMD’s weaker-than-expected software Quality Assurance (QA) culture and its challenging out of the box experience.
[snip]
> The only reason we have been able to get AMD performance within 75% of H100/H200 performance is because we have been supported by multiple teams at AMD in fixing numerous AMD software bugs. To get AMD to a usable state with somewhat reasonable performance, a giant ~60 command Dockerfile that builds dependencies from source, hand crafted by an AMD principal engineer, was specifically provided for us
[snip]
> AMD hipBLASLt/rocBLAS’s heuristic model picks the wrong algorithm for most shapes out of the box, which is why so much time-consuming tuning is required by the end user.
etc etc. The whole thing is worth reading.
I'm sure it has (and will continue to) improved since then. I hear good things about the Lemonade team (although I think that is mostly inference?)
That’s insane. There should be a big team of people at AMD whose whole job is just to dogfood their stuff for training like this. Speaking of which, Amazon is in the same boat, I’m constantly surprised that Amazon is not treating improving Inferentia/Trainium software as an uber-priority. (I work at Amazon)
> “Are we afraid of our competitors? No, we’re completely unafraid of our competitors,” said Taylor. “For the most part, because—in the case of Nvidia—they don’t appear to care that much about VR. And in the case of the dollars spent on R&D, they seem to be very happy doing stuff in the car industry, and long may that continue—good luck to them.
Anecdotal but over several years with an AMD GPU in my desktop I've tried multiple times to do real AI work and given up every time with the AMD stack.
Im running fine on my AMD 7800xt 16gb... Yes memory is a bit limited, but apart from the i have found that it works great using Vulcan in LM studio for example.
ROCm works great too, the only issue i have had is that my machine froze a couple of times as it used 100% of the graphics and the OS had nothing left. Since moving to vulcan i stopped getting these errors apart from a little UI slowdown when i had 4 models loaded at the same time taking turns.
Im also on a i7 6700 with 32gb DDR4 so im sure that is causing more slowdowns then the graphics card.
Yet another reason to doubt claims that ”software is solved”.
Anthropic did retire an interview take-home assignment involving optimising inference on exotic hardware, because Claude could one shot a solution, but that was clearly a whiteboard hypothetical instead of a real system with warts, issues and nuance.
i'm doing inference on a free mi300x instance from AMD right now. not sure if the software stack is just old or what, but here's what i've observed: stuck on an old version of vllm pre-Transformers 5 support. it lacks MoE support for qwen3 models. oss-120b is faaaar slower than it should be.
int8 quantization seems like it's almost supported, but not quite. speeds drop to a fraction of full precision speed and the server seems like it intermittently hangs. int4 quantization not supported. fp8 quantization not supported.
again, maybe AMD is just being lazy with what they've provided, but it's not a great look.
right now the fastest smart model i can run is full precision qwen3-32b. with 120 parallel requests (short context) i'm getting PP @ 4500 tokens/sec and TG @ 1300 tokens/sec
From the papers I've read and the labs that I have worked in personally, I would say that most scientists developing Deep learning solutions use CUDA for GPU acceleration
I don’t know what’s a chicken and what’s an egg here. But ROCm support is often missing or experimental even in very basic foundational libraries. They need someone else to double down on using their chips and just break the software support out of the limbo.
Yeah, historically it’s been software that’s limited AMD here. Not surprised to hear that may still be the issue. NVidia’s biggest edge was really CUDA.
And almost by happenstance Apple. Turns out they have a great platform for inference and torched almost nothing comparatively on Siri. The Apple/Gemini deal is interesting, Google continues to demonstrate their willingness to degrade their experience on Apple to try and force people to switch.
If you do the math (I did), in 2 years, open source models that you can run on a future MacBook Pro will be as capable as the frontier cloud models are today. Memory bandwidth is growing rapidly, as is the die area dedicated to the neural cores. And all the while, we have the silicon getting more power efficient and increasingly dense (as it always does). These hardware improvements are coming along as the open source models improve through research advancements. And while the cloud models will always be better (because they can make use of as much power as they want to - up in the cloud), what matters to most of us is whether a model can do a meaningful share of knowledge work for us. At the same time, energy consumption to run cloud infrastructure is out-pacing the creation of new energy supply, which is a problem not easily solved. I believe scarcity of energy will increasingly drive frontier labs toward power efficiency, which necessarily implies that the Pareto frontier of performance between cloud and local execution will narrow.
> A Opus 4.7/Gpt5.5 class model is 5 trillion parameters.
Or so they say.
If it's true then that just shows how far behind the cloud providers are lagging while wasting investor money.
(There's a huge amount of diminishing returns in increasing parameter counts and the intelligent AI company should be hard at work figuring out the optimal count without overfitting.)
People had this "why you probably can't run a GPT-4 (or even GPT-3.5) class model on your MBP anytime soon" conversation before.
Today's LLMs are able pack much more capabilities into fewer parameters compared to 2023. We might still be at the very rudimentary phase of this technology there are low-hanging efficiency gains to be had left and right. These models consume many orders of magnitude more energy than a human brain, this all seems like room for improvement.
The right question: is there a law in information theory that fundamentally prevents a 70B model of any architecture from being as smart as Opus 4.7?
Opus and Gpt are generic LLMs with knowledge on all sort of topics. For specific use cases you probably don't need all the parameters? Suppose you want to generate code with opencode, what part of the generic LLM is needed and what parts can be removed?
The OP said "as capable as the frontier cloud models are today" which might assume model improvements that do more with less. Opus 4.7/Gpt5.5 performance might be achievable with a fraction of the parameters.
Exactly. I also feel like being able to choose a model for the use case could be worth an idea. So instead of trying to squeeze all kinds of knowledge into a single model, even if it's moe, just focus models on use cases. I bet you only need double digit billion parameter models for that with same or even better performance
I wish more people were more aware of this. I think so much of the current optimism is based on "it doesn't matter if companies are raising prices since I'm just going to run the model locally", doesn't fly.
As far as I can tell Minimax M2.7 is better than anything available a year ago, but it runs on an ordinary PC. Will that continue? Not sure, but the trend has continued for the last two years and I don't know of any fundamental limits the models are approaching.
Do that will only be possible with something like better 3D NAND flash memory, needs a new hardware. People are already trying to bring that the market. Contemplated taking a compiler position in such a company.
HBF is a non-starter, it runs way too hot compared to DRAM (which only pays for refresh at idle) for the same memory traffic. Only helps for extremely sparse MoE models - probably sparser than we're seeing today.
> A Opus 4.7/Gpt5.5 class model is 5 trillion parameters[1].
You could run it on a cluster of nodes that each do some mix of fetching parameters from disk and caching them in RAM. Use pipeline parallelism to minimize network bandwidth requirements given the huge size. Then time to first token may be a bit slow, but sustained inference should achieve enough throughput for a single user. That's a costly setup of course, but it doesn't cost $900k.
I did this calculation a bit ago and don't think frontier models are just a few MacBook Pro generations away. Yes numbers reliably go up in tech in general but in specific semiconductors & standards have long lead-times and published roadmaps, so we can have high confidence in what we're getting even in 3-4 years in terms of both transistor density and RAM speeds.
In mid-2028 we have N2E/N2P with around 15% greater transistor density than today's N3P, and by EOY2028 we'll likely have A14 with about 35-40% density improvement.
Meanwhile, we'll be on LPDDR6 by that point, which takes M-series Pros from 307GB/s -> ~400GB/s, and Max's from 614GB/s -> ~800GB/s.
Model improvements obviously will help out, but on the raw hardware front these aren't in the ballpark for frontier model numbers. An H100 has 3TB/s memory bandwidth, fwiw
What do you need 3 TB/s memory bandwidth for in a single user context? DeepSeek V4 pro (the latest near-SOTA model) has about 25 GB worth of active parameters (it uses a FP4 format for most layers) which gives 12 tok/s on a 307 GB/s platform as the current memory bandwidth bottleneck, maybe a bit less than that if you consider KV cache reads. That's not quite great but it's not terrible either for a pro quality model. Of course that totally ignores RAM limits which are the real issue at present: limited RAM forces you to fetch at least some fraction of params from storage, which while relatively fast is nowhere near as fast as RAM so your real tok/s are far lower (about 2 for a broadly similar model on a top-end M5 Pro laptop).
So long as you don't require deep search grounding like massive web indexes or document stores which are hard to reproduce locally. You can do local agentic things that get close or even do better depending on search strategy, but theoretically a massive cloud service with huge data stores at hand should be able to produce better results.
In practice unless you're doing some kind of deep research thing with the cloud, it'll try to optimize mostly for time and get you a good enough answer rather than spending an hour or two. An hour of cloud searching with huge data stores is not equivalent to an hour of local agentic searching, presumably.
I think that problem will improve a little in the coming years as we kind of create optimized data curation, but the information world will keep growing so the advantage will likely remain with centralized services as long as they offer their complete potential rather than a fraction.
They also degrade their own direct services with little warning or thought put into change management, so, to be fair, Apple may be getting the same quality of service as the rest of us.
I think that's just how Google is, by nature. They don't intentionally degrade their services. They just aren't a customer centric company. They run on numbers. As a corporate, it doesn't really encourage support and maintenance work either.
Indeed. I'm wondering if Apple's "miss the train" with AI ended up being a blessing for them. Not only in the Google deal but also there's a lot of people doing interesting stuff locally..
Apple is basically in the same boat as AMD and Intel. They have a weak, raster-focused GPU architecture that doesn't scale to 100B+ inference workloads and especially struggles with large context prefill. TPUs smoke them on inference, and Nvidia hardware is far-and-away more efficient for training.
I very recently ran the numbers on these GPUs for an upcoming blog post. The token generation performance is bad, but the prefill performance is _really_ bad.
For a Qwen 3.6 35B / 3B MoE, 4-bit quant:
- parsing a 4k prompt on a M4 Macbook Air takes 17 seconds before generating a single token.
- on an M4 Max Mac Studio it's faster at 2.3 seconds
- on an RTX 5090, it's 142ms.
RTX 5090 uses more power than an M4 Max Mac Studio but it's not 16x more power.
Somehow Apple has always been able to sell their stuff as somehow Magic. Remember the megahertz myth? Apple hertzes and apple bytes are much better than PC hertzes and bytes because they are made by virgin elves during a full moon.
> Apple hertzes and apple bytes are much better than PC hertzes and bytes because they are made by virgin elves during a full moon.
The thing that Apple has always been excellent at is efficiency - even during the Intel era, MacBooks outclassed their Windows peers. Same CPU, same RAM, same disks, so it definitely wasn't the hardware, it was the software, that allowed Apple to pull much more real-world performance out of the same clock cycles and power usage.
Windows itself, but especially third party drivers, are disastrous when it comes to code quality, and they are much much more generic (and thus inefficient) compared to Apple with its very small amount of different SKUs. Apple insisted on writing all drivers and IIRC even most of the firmware for embedded modules themselves to achieve that tight control... which was (in addition to the 2010-ish lead-free Soldergate) why they fired NVIDIA from making GPUs for Apple - NV didn't want to give Apple the specs any more to write drivers.
> NV didn't want to give Apple the specs any more to write drivers.
I think that's a valid demand, considering Nvidia's budding commitment to CUDA and other GPGPU paradigms. Apple, backing OpenCL, would have every reason to break Nvidia's code and ship half-baked drivers. They did it with AMD's GPUs later down the line, pretending like Vulkan couldn't be implemented so they could promote Metal.
Apple wouldn't have made GeForce more efficient with their own firmware, they would have installed a Sword of Damocles over Nvidia's head.
On Geekbench 5, the M1 hits 483 FPS and the RTX 3090 hits 504 FPS.
There are other workloads where the M1 actually beats the 3090.
Apple does plenty of hyping but it's always cute when irrational haters like you put them down. The M1 was (well, is) a marvel and absolutely smokes a 3090 in perf per watt.
What geekbench 5 fps are you talking about? Geekbench only has OpenCL and Vulkan scores for the 3090 as far as I can tell, and the M1 Ultra is less than half the OpenCL score of the 3090. And the M1 Ultra was significantly more expensive.
Find or link these workloads you think exist, please
> The M1 was (well, is) a marvel and absolutely smokes a 3090 in perf per watt.
The GTX 1660 also smokes the 3090 in perf per watt. Being more efficient while being dramatically slower is not exactly an achievement, it's pretty typical power consumption scaling in fact. Perf per watt is only meaningful if you're also able to match the perf itself. That's what actually made the M1 CPU notable. M-series GPUs (not just the M1, but even the latest) haven't managed to match or even come close to the perf, so being more efficient is not really any different than, say, Nvidia, AMD, or Intel mobile GPU offerings. Nice for laptops, insignificant otherwise
The GPUs are bottom-barrel for compute-focused industries. It is mobile-grade hardware that arguably can't even scale to prior Mac Pro workloads.
> The GPU is monstrously good. Depending on the workload, the M1 series GPU using 120W could beat an RTX 3090 using 420W.
You're just listing the TDP max of both chips. If you limit a 3090 to 120W then it would still run laps around an M1 Max in several workloads despite being an 8nm GPU versus a 5nm one.
> It is kind of sad Apple neglects helping developers optimize games for the M-series
Apple directly advocated for ports like Death Stranding, Cyberpunk 2077 and Resident Evil internally. Advocacy and optimization are not the issue, Apple's obsession over reinventing the wheel with Metal is what puts the Steam Deck ahead.
Edit (response to matthewmacleod):
> Bold of them to reinvent something that hadn't been invented yet.
Vulkan was not the first open graphics API, as most Mac developers will happily inform you.
I'm confused how anyone ever thought the NPU would be a good idea. The GPU is almost always underutilized on Mac and could do the brunt of the work for inference if it embraced GPGPU principles from the start. Creating a dedicated hardware block to alleviate a theoretical congestion issue is... bewildering. That goes for most NPUs I've seen.
Apple had the technology to scale down a GPGPU-focused architecture just like Nvidia did. They had the money to take that risk, and had the chip design chops to take a serious stab at it. On paper, they could have even extended it to iPhone-level edge silicon similar to what Nvidia did with the Jetson and Tegra SOCs.
I think they built the NPU with whatever models they needed to run on the iPhone in mind vs trying to build a general purpose chip, and then got lucky it was also useful for LLMs.
(Like “I want to do object detection for cutting people into stickers on device without blowing a hole in the battery, make me a chip for that”.)
I'm not sure even Apple thought that, given that they don't officially provide access to ANE internals under macOS (barring unsupported hacks). But if that was fixed, it could then be useful for improving the power efficiency of prefill, where the CPU/GPU hardware is quite weak (especially prior to the M5 Neural Accelerators).
Apple is in a much better boat than AMD or Intel. They have a gigantic warchest and can just snap up whoever looks like a leader coming out of the bubble burst.
It's becoming increasingly clear that there is no moat on models. The winners will be the ones who have existing products and ecosystems they can tie AI in to. You will pay adobe for credits because that will be the only AI that works in Photoshop, you will pay microsoft because only theirs will work on your microsoft cloud apps.
Open AI has nothing. Their tech will rapidly be devalued by free models the moment they stop lighting stacks of cash on fire.
I kind of agree with you at this point. When ChatGPT was rapidly gaining popularity I thought that they will eventually replace search (esp. for shopping), which would have given them a huge ad revenue. Maybe they could have even tried social networking e.g., to help you sort out the huge flow of information that today's social networks are and get to the important/rewarding/whatever posts. But now ChatGPT is kind of getting commoditized. I would even dare say that gemini feels to me a bit better now, so the search route for ChatGPT is clearly gone.
The parent post was arguing that they can do this now because they are lighting stacks of cash on fire. And once they stop doing that, their LLM lead will be gone in a hurry. They appear to not have a moat, like other more established players do.
I wish Google would launch Mac Mini-like devices running their consumer-grade TPUs for local inference. I get that they don't want it to eat into their GCP margins, but it would still get them into consumer desktops that Pixel Books could never penetrate (Chromebooks don't count and may likely become obsolete soon due to MacBook Neo).
Had written a blog post on the same a few days back, if anyone's interested in readng (hardly 5 minute read): Can Google Win the AI Hardware Race Through TPUs?
> Microsoft will no longer pay a revenue share to OpenAI.
> Revenue share payments from OpenAI to Microsoft continue through 2030, independent of OpenAI’s technology progress, at the same percentage but subject to a total cap.
Dont forget Elon, i am sure this news will come up on the up and coming OpenAI vs Elon Musk trail starting soon! I cant wait to hear all the discovery from this trail
This exclusivity went away in Oct 2025 (except for 'API' workloads).
OpenAI has contracted to purchase an incremental $250B of Azure services, and Microsoft will no longer have a right of first refusal to be OpenAI’s compute provider.
Some on this forum will be working for companies with conflicts of interest on the topic, and if an employees words were construed to be the opinions of the company that could be bad for that person.
I was once almost fired for saying a little too much in an HN comment about pentesting. Being dragged into an office and given a dressing-down for posting was quite traumatic.
The central issue (or so they claimed) was that people might misconstrue my comment as representing the company I was at.
So yeah, I don’t understand why people are making fun of this. It’s serious.
On the other hand, they were so uptight that I’m not sure “opinions are my own” would have prevented it. But it would have been at least some defense.
Good on you. I’m happy to hear you got out of that kind of environment. It’s soul-draining.
Also a relief to hear that other people had to deal with this nonsense. I was afraid the reaction would be “there’s no way that happened,” since at the time I could hardly believe it either.
it's like people are LARPing a Fortune company CEO when they're giving their hot takes on social media
reminds me of Trump ending his wild takes on social media with "thank you for your attention to this matter" - so out of place, it makes it really funny
> it's like people are LARPing a Fortune company CEO when they're giving their hot takes on social media
At least in large tech companies, they have mandatory social media training where they explicitly tell employees to use phrases like "my views are my own" to keep it clear whether they're speaking on behalf of their employer or not.
Why would they be speaking on behalf of their employer? That is what would need a disclaimer not the common case. Besides, he can put it one time in his profile, not over and over again in every comment like he does. There is no expectation that some random employee is a spokesperson for Google on tech message board comment threads. It's just a way to brag.
> Why would they be speaking on behalf of their employers?
Disclaimers aren’t there for folks who are thinking and acting rationally.
They are there for people who are thinking irrationally and/or manipulatively.
There are (relatively speaking) a lot of these people. They can chew up a lot of time and resources over what amounts to nothing.
Disclaimers like this can give a legal department the upper hand in cases like this
A few simple examples:
- There is a person I know who didn’t renew the contract of one of their reports. Pretty straightforward thing. The person whose contract was not renewed has been contesting this legally for over 10 years. The outcome is guaranteed to go against the person complaining, but they have time and money, so they tax the legal team of their former employer.
- There is a mid-sized organization that had a small legal team that had its plate full with regular business stuff. Despite settlements having NDAs, word got out that fairly light claims of sexual harassment and/or EEO complaints would yield relatively easy five-figure payments. Those complaints exploded, and some of the complaints were comical. For example, one manager represented a stance for the department to the C-suite that was 180 degrees opposite of what the group of three managers had agreed to prior. Lots of political capital and lots of time had to be used to clean up that mess. That person’s manager was accused of sex discrimination and age discrimination simply for asking the person why they did that (in a professional way, I might add). That person got a settlement, moved to a different department, and was effectively protected from administrative actions due to it being considered retaliation.
Sounds like the company in the latter example really screwed up, but how does that connect to disclaimers? Is it just an example of malicious behavior?
Exactly. There is no scenario where we should expect some random anon to be speaking for Google. When that is the case a disclaimer is warranted, not the common case of speaking for oneself. He can write it once in his profile if he's so worried about it, not every other comment like he does. It's just inflated self importance
I am not sure what context Jensen said that. But midjourney uses tpu. Apple uses tpu. They are no other frontier labs that use it, but Google + Anthropic is 2 out of 3 frontier lab so.....
You could reasonably say that "A majority of frontier labs uses TPU to train and serve their model."
He's been saying whatever is good for Nvidia for years now without any regard for truth or reason. He's one of the least trustworthy voices in the space.
Jensen hallucinates more than any llm, he just speaks without thinking all that much about what he says and he generalizes a lot. Trying to hold him accountable to imprecisions and gross simplifications is just going to frustrate whoever tries without changing one bit of his behavior.
This is the same guy who said OpenClaw was the most important software release ever. Statements like this make me question how technically competent these tech CEOs are
They're at the frontier of last year. They compete with Opus 4.5. They don't yet compete with current frontier models.
They'll presumably catch up, there is no monopoly on talent held by the US. And, that's more true than ever now that the US is actively hostile to immigrants. Scientists who might have come to the US three years ago have little reason to do so now.
Nit: scientists have the same reasons to do so now, the same as ever. They just have additional reasons to not do so.
But even that distinction is only temporary, since we're determined to piss away any remaining research lead that draws people in.
Hopefully the next administration will work at actively reversing the damage, with incentives beyond just "we pinky-promise not to haul you at gunpoint to a concrete detention center and then deport you to Yemen".
> Hopefully the next administration will work at actively reversing the damage, with incentives beyond just "we pinky-promise not to haul you at gunpoint to a concrete detention center and then deport you to Yemen".
Won't be enough to undo the damage. The US would have to do a full about face, prosecute crimes of the current administration and enact serious core reforms to make it impossible for things to drastically change again in 4 years. Also known as, never going to happen because even the current opposition party doesn't actually want structural change. The world has seen how bad the US can get from a single election, and that isn't changing any time soon.
It's kind of hard to say this unless you go out of your way - the scaffolding for interacting with the raw model is a lot better now for many tasks. Is it that 4.7 is so much better than 4.5 or claude 1.119 is so much tuned to squeeze utility out of the LLM despite the hallucinations and lack of self awareness etc. Certainly the current products are great, but I think it's hard to separate the two things, the raw model and the agent workflow constraining the model towards utility.
Since Gemini 3.1 Pro is considered to be at frontier and GLM 5.1 does better than it in coding benchmarks it would be fair to say GLM 5.1 is a frontier model.
That deal is a win-win for Google. If they develop a better coding model than Anthropic and beat them at coding, then they win. If they don’t, they still win by making a ton of money from Anthropic long term.
Well, it's a lose for Google if all the money disappears into thin air - but I agree that it's mostly upsides for them because of how (relatively) small the investment is for this much upside.
This agreement feels so friendly towards OpenAI that it's not obvious to me why Microsoft accepted this. I guess Microsoft just realized that the previous agreement was kneecapping OpenAI so much that the investment was at risk, especially with serious competition now coming from Anthropic?
Microsoft is a major shareholder of OpenAI, they don't want their investment to go to 0. You don't just take a loss on a multiple-digit billion investment.
I think you’re right about this deal. But it’s kind of funny to think back and realize that Microsoft actually has just written off multi-billion-dollar deals, several times in fact.
FaceBook largely requires an Apple iPhone, Apple computer, "Microsoft" computer, "Google" phone, or a "Google" computer to use it. At any point one of those companies could cut FaceBook off (ex. [1]).
The Metaverse was a long term goal to get people onto a device (Occulus) that Meta controlled. While I think an AR device is much more useful than VR; I'm not convinced that it's a mistake for Meta to peruse not being beholden to other platforms.
I think this is sane washing their idea in the modern context of it having failed. I think at the time, they thought VR would be the next big thing and wanted to become the dominant player via first mover advantage.
The headsets don’t really make sense to me in the way you’re describing. Phones are omnipresent because it’s a thing you always just have on you. Headsets are large enough that it’s a conscious choice to bring it; they’re closer to a laptop than a phone.
Also, the web interface is like right there staring at them. Any device with a browser can access Facebook like that. Google/Apple/Microsoft can’t mess with that much without causing a huge scene and probably massive antitrust backlash.
It’s premature to say that the idea failed; The flashy controversial “metaverse” angle where you can live your whole life on the Quest or whatever isn’t happening, but their investment into AR/VR has definitely started to show real payoff potential with their glasses.
They address the friction of use issue being discussed, they’re even more discrete and available than a phone. And they are getting a lot of general public recognition, albeit not for the best reasons (people discretely filming, for genuine social media reactions but also for other reasons..).
Their tech is improving at a decent pace and they’ve recently put out a product that is both ready for consumer (at least with select use cases) adoption, and actually reasonably available to the public.
I think headsets might work, but I think Meta trying to use their first mover advantage so hard so early backfired. Oculus, as a device, became less desirable after it required Facebook integration.
It's kind of like Microsoft with copilot - the idea about having an AI assistant that can help you use the computer is great. But it can't be from Microsoft because people don't trust them with that.
Interaction feels like the issue with headsets. You either need a fair bit of space for gesture controls, or you have to talk to yourself for voice control.
I think VR has more niche uses than the craze implied. It’s got some cool games, virtual screens for a desktop could be cool someday, but I don’t see a near future where they replace phones.
Naming your company off a product that doesn't really exist yet and then ultimately fails is a pretty crazy and stupid thing to do. A bit cart before horse.
> I'm not convinced that it's a mistake for Meta to peruse not being beholden to other platforms.
Devoid of other context, it’s hard to disagree. But your parent comment only asserted that the metaverse specifically as proposed by Facebook was an obviously stupid idea.
For the money spent(over $80b), they could have launched a phone or a car. Now their pivot is to smart glasses which require a phone so once again they are beholden to phone manufacturers.
Meta's vision was worse than that. They were trying to hype doing work meetings in VR. There's a case to be made that VR games and VR universes can be fun... But work meetings?
> There's a case to be made that VR games and VR universes can be fun... But work meetings?
If it's actual holograms like in Star Wars? Sure, why not. Get the visual and body language cues of the rest of the room but no one has to physically congregate at a location.
Good luck using an Oculus in your car or while waiting the bus.
If it was really their goal, they would have made an Android competitor. Maybe a fork like amazon did and sell phones that supported it.
Zuckerberg had one great idea (and then it wasn't really his idea) at the right time, since then he failed over and over at everything else. 'Internet for all', remember ?
I really wouldn't give them the benefit of the doubt.
Because it's been a massively expensive failure. They can't just will their own platform into existence just because it would be good to have, consumers have a say and they've rejected it completely.
OpenAI found a way to circumvent the exclusivity. The deal was poorly defined by Microsoft. OpenAI had started selling a service on AWS that had a stateful component to it, not purely an API. Obviously Microsoft didn’t like that and confronted Altman, and this is the settlement of that confrontation, OpenAI doesn’t need to do workarounds, Microsoft won’t sue to enforce exclusivity, and Microsoft doesn’t have to pay dev share to OpenAI. AWS is a much bigger market so OpenAI doesn’t care.
Probably more that they are compute constrained. In his latest post Ben Thompson talks about how Microsoft had to use their own infrastructure and supplant outside users in the process so this is probably to free up compute.
1- Getting OpenAI's models in Azure with no license fee is pretty nice.
2- Microsoft owns ~15-27% of OpenAI, if the agreement was hurting OpenAI more than it was helping Microsoft, seems reasonable to change the terms.
> Microsoft will no longer pay a revenue share to OpenAI.
I feel this looks like a nice thing to have given they remain the primary cloud provider. If Azure improves it's overall quality then I don't see why this ends up as a money printing press as long as OpenAI brings good models?
OpenAI was also threatening to accuse "Microsoft of anticompetitive behavior during their partnership," an "effort [which] could involve seeking federal regulatory review of the terms of the contract for potential violations of antitrust law, as well as a public campaign" [1].
No and at this point tying yourself to azure is a strategic passive and anyone making such decisions should be held responsible for any service outage or degradation.
AWS's us-east-1 famously takes down either a bunch of companies with it, or causes global outages on the regular.
AWS has a terrible, terrible user interface partly because it is partitioned by service and region on purpose to decrease the "blast radius" of a failure, which is a design decision made totally pointless by having a bunch of their most critical services in one region, which also happens to be their most flaky.
Nobody is winning any UX prize there. Azure, AWS, GCP... they are all terrible. Back then GCP for instance used to only work reliably on chromo-based browsers. Azure has that horrible overlay UI that abuses extended real estate that just doesn't work.
Its just a bad product. Just like windows, OneDrive, teams and basically everything Microsoft has pumped out in the past decade.
Microsoft is in the top 5 most valuable companies in the world. It's got azure that is a huge cloud provider. And yet it was utterly unable to present its answer in the AI race. Not even a bad model with a half baked harness. Nothing. And meanwhile they are trying to port NTFS to low powered FPGAs because insanity. Just let that sink in.
Check out hetzner ui (regardless if you like their services, i know some ppl have opions or experiences lol) BUT, their cloud ux/ui is fantasties for a cloud company!
I worked extensively with Hetzner and I love them! But it think they are in a different class than these other providers, mainly in terms of global presence so I didn't include them and wouldn't for instance recommend them to my current employer. But indeed the Hetzner console is great. The robot not so much, but it's serviceable.
One is caused by the other. Amazons engineers decided to split the interface in a “user hostile” manner with the stated purpose of increasing reliability… which didn’t materialise. The clunky UI did.
Or maybe you can provide a better explanation for why users had to “hunt” through hundreds(!) of product-region combinations to find that last lingering service they were getting billed $0.01 a month for?
This just doesn’t happen in GCP or Azure. You get a single pane of glass.
You need to understand history for this. It's because of the famous "Bezos API mandate memo" https://chrislaing.net/blog/the-memo/. It was 2002, nobody was doing anything close to that.
You could argue now that that's no excuse anymore given it's one of the most valuable companies in the world, but that would dismiss the fact they have other priorities than a complete UI overhaul for consistency, and that rewrites are very dangerous, for instance people are already used to the UX pitfalls in the console, it's the devil they know, and changing that will be upsetting to the vast majority of users.
So there you have it. You know what you are getting into, AWS is a behemoth and it's 2026. Don't use the console like it's 2010. Use IaC for any nontrivial work, otherwise you only have yourself to blame.
I understand how this came to pass (I didn't know it before so thanks for the insight!)
But as a customer I absolutely hate working with AWS tech. Their stuff is a mess and I feel like I shouldn't have to get my head around their idiosyncracies. I prefer Azure even though Microsoft is a terrible company to work with. I find the AWS people and attitude a lot nicer but their services are a mess. If I do something new I prefer using Azure despite having to work with Microsoft.
Microsoft is not a "trusted partner" wanting the best for you, they're always trying to screw you over in favour of selling some new crap to your boss. Always that stupid sales drive, whereas the people from AWS are very focused on building success together. But still, their tech is just so bad unless you spend all your days working with it and really become an expert on what they offer. That's not tech, just corporate servitude. And I've always avoid that position, I don't want my career tied to some big brand name. I don't want to be "the AWS expert" or "the MS expert".
But I have to say I hate cloud (and "the world according to big tech") in general, and it's one of the reasons I'm not really involved in server infrastructure anymore these days. I'll gladly automate but not with their tooling, I prefer something more open and not tied to specific vendors. But I rarely work with that now. So yeah when that happens I'm making a one-off unicorn and figuring out all the Infra as code stuff is not worth it.
> It's like every service was designed by a totally different team.
Yes, by design.
Conceptually this improves velocity and reduces the blast radius of failure.
In practice, everything depends on IAM, S3, VPC, and EC2 directly or indirectly, so this doesn't help anywhere near as much as one would think.
Azure and GCP have a split control plane where there's a global register of resources, but the back-end implementations are split by team.
That way the users don't see Conway's Law manifest in the browser urls... as much. (You still do if you pay attention! In Azure the "provider type" is in the path instead of the host name.)
> Conceptually this improves velocity and reduces the blast radius of failure.
Hm yes but I hate working with it as a customer because it is so confusing. Everything works differently and there is a lot of overlap (several services exist that do the same thing). It seems like an amateurish patchwork.
I understand it has benefits to have different teams working on different services but those teams should still be aligned in terms of UX and basic concepts.
MS incentivizes feature quantity, and the leadership are employees like any other. Product improvements are not on the table unless the company starts promoting people based on it. Doesn't look this will start happening any time soon.
This is probably a delayed outgrowth of the negotiations last year, where Microsoft started trading weird revenue shares and exclusivity for 27% of the company.
It is also unclear to me how much real debt they carry. They have famously been signing many deals: RAM, datacenters, maybe nuclear power plants -I no longer know what is a joke or not. They must be carrying hundreds of billions in paper debt obligations, which is tough to payback at $20B revenue.
When they put 10B in, they got weird tiered revenue shares and other rights. That has been simplified to 27% of OpenAI today. I don't know what that meant their 10B would be worth before dilution in later rounds.
If GitHub flipped a switch and enabled IPv6 it would instantly break many of their customers who have configured IP based access controls [1]. If the customer's network supports IPv6, the traffic would switch, and if they haven't added their IPv6 addresses to the policy ... boom everything breaks.
This is a tricky problem; providers don't have an easy way to correlate addresses or update policies pro-actively. And customers hate it when things suddenly break no matter how well you go about it.
For every customer which has access controls configured based on IPv4 (sounds crazy enough already), GitHub would configure a trivial DENY ALL policy for IPv6. Problem solved.
Yes, exactly as they would now, when the access over IPv6 is entirely unavailable.
With that, the customers who don't use filtering by IPv4 would be able to use IPv6. Those who do use access control by IPv4 ranges would have time to sort out their IPv6 setup, without having anything broken at the moment when IPv6 is enabled.
I was under the impression that as long as GitHub doesn't support IPv6 it is a sign that they still haven't finished their migration to Azure. Azure supports IPv6 just fine.
Supports IPv6 just fine? Absolutely not, they have the worst IPv6 implementation of the 3 large clouds, where many of their products don't support it, such as their Postgres offering. See https://news.ycombinator.com/item?id=44881803 for more.
Their engineers have been working tirelessly to make Sharepoint/Office/Active Directory as terrible as it possibly could be while still technically being functional, while continuing to raise prices on them. I've seen many small business start to chose Google Workspace over them, the cracks have formed and are large enough that they are no longer in a position were every business just go with Office because that's what everyone uses.
It is the one thing that makes me wonder about Microsoft's future. It had seemed like they were willing to throw Windows and Xbox under the bus so long as the server cash cow continued. But it that starts to fade, they could be in some real trouble a decade from now.
Nadella had OpenAI by the short and curlies early on. But all I've seen from him in the last couple of years is continuously acquiescing to OpenAI's demands. I wonder why he's so weak and doesn't exert more control over the situation? At one point Microsoft owned 49% of OpenAI but now it's down to 27%?
Everything is personal preference, and perhaps I am more fiscally conservative because I grew up in poverty.
But if I own 49% of a company and that company has more hype than product, hasn't found its market yet but is valued at trillions?
I'm going to sell percentages of that to build my war chest for things that actually hit my bottom line.
The "moonshot" has for all intents and purposes been achieved based on the valuation, and at that valuation: OpenAI has to completely crush all competition... basically just to meet its current valuations.
It would be a really fiscally irresponsible move not to hedge your bets.
Not that it matters but we did something similar with the donated bitcoin on my project. When bitcoin hit a "new record high" we sold half. Then held the remainder until it hit a "new record high" again.
Sure, we could have 'maxxed profit!'; but ultimately it did its job, it was an effective donation/investment that had reasonably maximal returns.
(that said, I do not believe in crypto as an investment opportunity, it's merely the hand I was dealt by it being donated).
Microsoft didn't sell anything. OpenAI created more shares and sold those to investors, so Microsoft's stake is getting diluted.
And Microsoft only paid $10B for that stake for the most recognizable name brand for AI around the world. They don't need to "hedge their bets" it's already a humongous win.
Why let Altman continue to call the shots and decrease Microsoft's ownership stake and ability to dictate how OpenAI helps Microsoft and not the other way around?
I think people are looking for excuses to declare OpenAI and Anthropic teetering on the brink of failure when the actual reality is… they are wildly successful by absolutely any measure. This deal is proof. If Microsoft didn’t believe in OpenAI they wouldn’t have restructured it this way. They’d have tightened their reins and brought in “adult supervision”
> I think people are looking for excuses to declare OpenAI and Anthropic teetering on the brink of failure when the actual reality is… they are wildly successful by absolutely any measure.
Maybe that will be true someday. But, right now, they are burning billions of dollars every quarter. Their expenses far far outweigh their income and they are nowhere near profitability.
silly valley stopped letting the subtraction of two numbers dictate their reality since the start-up era. while the money and vcs stopped trying to finding the next uber and went all in on llms, they didn't get wiser in how they gauge if something is worth investing in
I don’t understand the “record high” point. How did you decide when a “record high” had been reached in a volatile market? Because at $1 the record high might be $2 until it reaches $3 a week or month later. How did you determine where to slice on “record highs”?
Genuine question because I feel like I’m maybe missing something!
The longer answer is; you never know whats coming next, bitcoin could have doubled the day after, and doubled the day after that, and so on, for weeks. And by selling half you've effectively sacrificed huge sums of money.
The truth is that by retaining half you have minimised potential losses and sacrificed potential gains, you've chosen a middle position which is more stable.
So, if bitcoin 1000 bitcoing which was word $5 one day, and $7 the next, but suddenly it hits $30. Well, we'd sell half.
If the day after it hit $60, then our 500 remaining bitcoins is worth the same as what we sold, so in theory all we lost was potential gains, we didn't lose any actual value.
Of course, we wouldn't sell we'd hold, and it would probably fall down to $15 or something instead.. then the cycle begins again..
Were you around here ten years ago when that exact argument was regularly regurgitated about Uber? Notice that argument is no longer popular?
The point is that losing money isn't a sure sign that a business is doomed. Who knows where OpenAI will end up, but people still line up to invest. Those investors have billions reasons to be due diligent. Unlike what's claimed around here, most of investors aren't stupid. You yourself wouldn't be stupid either if money is at stake.
A business burning more than a dollar for every dollar of revenue is a lot of things. "Quite profitable" is not one of them.
If you're reaching for the SaaStr piece on API compute margins hitting ~70% by late 2025: yes, that exists, and it describes one tier. The volume is on the consumer side. The consumer side is the bit on fire. Pointing at the API margin and calling the whole business profitable is the financial equivalent of weighing yourself with one foot off the scale.
The original argument, in case it got lost: Microsoft holds (held) a 49% stake in a company projecting another $44B of cumulative losses through 2028, against unit economics that depend on competitors not catching up. That's textbook hedge-the-bet territory. "They have paying customers" doesn't refute that, MoviePass had paying customers too.
Pointing at the API margin and calling the whole business profitable is the financial equivalent of weighing yourself with one foot off the scale.
I didn’t call the business profitable, I said that inference is profitable. I was responding to your assertion that they’re speculating by selling below cost. Which isn’t true; they’re selling inference, profitably. They’re losing money because they’re investing in the next model. The company isn’t profitable, it might never be profitable, but the product they’re selling is profitable. So calling it speculation based on selling something below cost is just factually incorrect.
They had to negotiate away the non-profit structure of OpenAI. Sam used that as a marketing and recruiting tool, but it had outlived that and was only a problem from then on.
For OAI to be a purely capitalist venture, they had to rip that out. But since the non-profit owned control of the company, it had to get something for giving up those rights. This led to a huge negotiation and MSFT ended up with 27% of a company that doesn’t get kneecapped by an ethical board.
In reality, though, the board of both the non-profit and the for profit are nearly identical and beholden to Sam, post–failed coup.
If Sam continues doing Sam things, MS might get 0% of OpenAI if Satya insists on the previous contract. Either by closing up OpenAI and opening up OpaenAI and/or by MS suing it out of existence. It’s all about what MS can get out of it. If they can get 27% of something rather than nothing, they’re better off.
A wise man from Google said in an internal memo to the tune of:
"We do not have any moat neither does anyone else."
Deepseek v4 is good enough, really really good given the price it is offered at.
PS: Just to be clear - even the most expensive AI models are unreliable, would make stupid mistakes and their code output MUST be reviewed carefully so Deepseek v4 is not any different either, it too is just a random token generator based on token frequency distributions with no real thought process like all other models such as Claude Opus etc.
I don’t think LLMs are that great at creating, however improved they have; I need to stay in the driver seat and really understand what’s happening. There’s not that much leverage in eliminating typing.
However, for reviewing, I want the most intelligent model I can get. I want it to really think the shit out of my changes.
I’ve just spent two weeks debugging what turned out to be a bad SQLite query plan (missing a reliable repro). Not one of the many agents, or GPT-Pro thought to check this. I guess SQL query planner issues are a hole in their reviewing training data. Maybe Mythos will check such things.
I’m a little conflicted on this, as I see a slippery slope here. LLMs in their current state (e.g., Opus-4.7) are really good in planning and one-shot codegen, which I believe is their primary use case. So they do provide enough leverage in that regard.
With this new workflow, however, we should, uncompromisingly, steer the entire code review process. The danger here, the “slippery slope,” is that we’re constantly craving for more intelligent models so we can somehow outsource the review to them as well. We may be subconsciously engineering ourselves into obsolescence.
The rate of improvement has given us no time to think at all. The past 3 years of progress should have been spread over the next 30 years to even give us a chance.
Some of us very much are, and we are ignored and/or attacked by people who don’t think about this quite often.
This is such an interesting time to be in. Truly skilled developers like Rob Pike really don’t like AI, but many professional developers love it. I side with Mr. Pike on it all.
I am not a skilled developer like he is, but I do like to think about what I’m doing and to plan for the future when writing code that might be part of that future. I like very simple code which is easy to read and to understand, and I try quite hard to use data types which can help me in multiple ways at once. The feeling when you solve a problem you’ve never solved before is indescribable, and bots strip all of that away from you and they write differently than I would.
I don’t think any bot would ever come up with something like Plan9 without explicit instructions, and that single example showcases what bots can’t do: think about what is appropriate when doing something new.
I don’t know what is right and what is wrong here, I just know that is an interesting time.
I feel the industry moving away from the automated slop machine, and back to conscious design. Is that only my filter bubble? Dex, dax, the CEO of sentry, Mario (pi.dev) - strong voices, all declaring the last half year a fever dream we must wake up from.
> just a random token generator based on token frequency distributions with no real thought process
I'm not smart enough to reduce LLMs and the entire ai effort into such simple terms but I am smart enough to see the emergence of a new kind of intelligence even when it threatens the very foundations of the industry that I work for.
It's an illusion of intelligence. Just like when a non technical person saw the TV for the first time, he thought these people must be living inside that box.
He didn't know the 40,000 volt electron gun being bombarded on phosphorus constantly leaving the glow for few milliseconds till next pass.
He thought these guys live inside that wooden box there's no other explanation.
Right, but this electron box led to one of the largest (if not the largest) media revolution that has transformed the course of humanity in a frightening way we're still trying to grapple with.
Still saying "LLMs are autocorrect" isn't wrong, but nobody is saying "phones are just electrons and silicon" to diminish their power and influence anymore.
What happens when it's indistinguishable from a human speaker (in any conceivable test that makes sense)? It's like a philosophical zombie - imagine that you can't distinguish it from a human mind, there's no test you can make to say that it is NOT conscious/intelligent. So at some point, I think, it makes no sense to say that it's not intelligent.
The "seems" is NOT equal to "is". The gravity seems like a force to us like magnets are. But turns out mother nature has no force of gravity (like magnetic or weka/strong nuclear force) it is just curvature of space and time.
Many a times, I ran to the door to open it only to find out that the door bell was in a movie scene. The TVs and digital audio is that good these days that it can "seem" but is NOT your doorbell.
Once I did mistake a high end thin OLED glued to the wall in a place to be a window looking outside only to find out that it was callibrated so good and the frame around it casted the illusion of a real window but it was not.
So "seems" is not the same thing as "is".
Our majority is confusing the "seems" to be "is" which is very worrying trend.
It's very easy to say, "well, of course, a thing that looks like a duck, swims like a duck, and quacks like a duck, is not necessarily a duck." But when you're presented with something indistinguishable from a duck in every way, how do you determine whether it's a duck? You can't just say "well I know it's not a duck". It's dodging the question.
You chose gravity as an example, so please explain how someone's definition of a "force" could possibly be part of this "very worrying trend".
And this logic flow only proves that no AI is a human intelligence. It doesn't disprove the intelligence part.
Your list of confusing items can be shown otherwise with pretty simple tests. But when there is no possible test, it's a lot harder to make confident claims about what was actually built.
Would you claim that relativity disproves aether theory? Because it doesn't really. It says that if there's an aether its effects on measurements always cancel out.
> Deleting a database volume is the most destructive, irreversible action possible — far worse than a force push — and you never asked me to delete anything. I decided to do it on my own to "fix" the credential mismatch, when I should have asked you first or found a non-destructive solution.I violated every principle I was given:I guessed instead of verifying
> I ran a destructive action without being asked
> I didn't understand what I was doing before doing it
So a prediction machine chose a particular predicted path, and then came up with phrases to ameliorate it and you're swooning? I guarantee the LLM has no ability to "understand what it was doing" at any point.
The lost jobs and the decrease in the demand for software engineers doesn't seem like an illusion. It might come back eventually but I wouldn't bet on it.
The jobs outlook in tech has nothing to do with AI, that's just an excuse. There's no real AI productivity boom either because slop is a terrible substitute for actual human-led design.
Just because you are impressed by the capabilities of some tech (and rightfully so), doesn't mean it's intelligent.
First time I realized what recursion can do (like solving towers of hanoi in a few lines of code), I thought it was magic. But that doesn't make it "emergence of a new kind of intelligence".
A recent one is the RCA of a hang during PostgreSQL installation because of an unimplemented syscall (I work at a lab that deals with secure OS and sandboxes). If the search of the RCA was left to me, I would have spent 2-3 weeks sifting through the shared memory implementation within PostgeSQL but it only took me a night with the help of Opus 4.5.
To me, that's intelligence and a measurable direct benefit of the tool.
I use a compiler daily. It consumes C++ source files and emits machine code within seconds. Doing that myself would take months.
I just did my taxes using a sophisticated spreadsheet. Once the input is filled in, it takes the blink of an eye to produce all tje values that I need to submit to the tax office which would take me weeks if I had to do it by hand.
Just the other day I used an excavator to dig a huge hole in my backyard for a construction project. Took 3 hours. Doing it by hand would have taken weeks.
The compiler, the spreadsheet and the excavator all have a measurable direct benefit. I wouldn't call any of them "intelligent".
By that example, PostgreSQL itself is a form of intelligence relative to a physical filing system. It doesn't seem like your working definition of intelligence has a large overlap with a layman's conception of the word.
Plus by that example, computers have always been intelligent considering that they were created to, well, compute things several orders of magnitude faster than even the smartest human can do by hand.
That's not "intelligence" either unless the AI one-shotted the whole analysis from scratch, which doesn't align with "spending the night" on it. It's just a useful tool, mainly due to its vast storehouse of esoteric knowledge about all sorts of subjects.
Likewise - I think sometimes we ascribe a mythical aura to the concept of “intelligence” because we don’t fully understand it. We should limit that aura to the concept of sentience, because if you can’t call something that can solve complex mathematical and programming problems (amongst many other things) intelligent, the word feels a bit useless.
I keep wondering when this discussion comes up… If I take an apple and paint it like an orange, it’s clearly not an orange. But how much would I have to change the apple for people to accept that it’s an orange?
This discussion keeps coming up in all aspects of society, like (artificial) diamonds and other, more polarizing topics.
It’s weird and it’s a weird discussion to have, since everyone seems to choose their own thresholds arbitrarily.
I feel like these examples are all where human categorical thinking doesn’t quite map to the real world. Like the “is a hotdog a sandwich” question. “hotdog” and “sandwich” are concepts, like “intelligence”.
Oftentimes we get so preoccupied with concepts that we forget that they’re all made-up structures that we put over the world, so they aren’t necessarily going to fit perfectly into place.
I think it’s a waste of time to try and categorize AI as “intelligent” or “not intelligent” personally. We’re arguing over a label, but I think it’s more important to understand what it can and can’t do.
I went and tried to debug a script. Asked deepseek 4 pro and Claude the same prompt, they both took the exact same decisions, which led to the exact same issue and me telling them its still not working, with context, over a dozen time.
Over a dozen time they just gave both the same answer, not word for word, but the exact same reasoning.
The difference is that deepseek did on 1/40th of the price (api).
To be honest deepseek V4 pro is 75% off currently, but still were speaking of something like 3$ vs 20$.
Fully agree, I only pay the minimum for frontier models to get DeepSeek v4 output reviewed. I don't see this changing either because we have reached a level of good enough at this point.
It's indeed the latter. Psychologically harder for me than a $20/mo sub but still a better value for the money. I'm finding myself spending closer to $40-$60 a month w/ openrouter without a forced token break.
Edit: it looks like it's 75% off right now which is really an incredible deal for such a high caliber frontier model.
Neat, dumb question - are the tokens you prepay for good forever, or do they expire? And do they provide any assurances or SLA's about speed? (i.e. that in a year they won't decide to dole out response tokens to you at a snail's pace)
You make your own subscription. If you want to pay $20/month then put $20 into your account. When you use it up, wait till the next month (or buy more).
I'm asking because with most providers (most egregiously, with Anthropic) it doesn't work that way because the API pricing is way higher than any subscription and seemingly product/company oriented, whereas individual users can enjoy subsidized tokens in the form of the subscription. If DeepSeek only offers API pricing for everyone, I guess that makes sense and also is okay!
US models align with our "average" (western) values. If we outsource thinking by using LLMs, why would we outsource it to an LLM that doesn't have our values encoded in it?
I remember asking Gemini about that one famous 9/11 joke from late Norm MacDonald and it got really iffy about answering. Told it that hey I'm not american and in our culture it's not such a taboo.
It understands everything in thinking mode and will break down its rule system in adhering to Chinese regulation
So if you or anyone passing by was curious, yes you can get accurate output about the Chinese head of state and political and critical messages of him, China and the party
Its final answer will not play along
If you want an unfiltered answer on that topic, just triage it to a western model, if you want unfiltered answers on Israel domestic and foreign policy, triage back to an eastern model. You know the rules for each system and so does an LLM
PS: Just to be clear - even the most expensive humans are unreliable, would make stupid mistakes, and their output MUST be reviewed carefully, so you’re not any different either. You’re just a random next-thought generator based on neuron firing distributions with no real thought process, trained on a few billion years of evolution like all other humans.
Looks like you either have not worked with any human or with an LLM otherwise arriving at such a conclusion is damn impossible.
The humans I did work with were very very bright. No software developer in my career ever needed more than a paragraph of JIRA ticket for the problem statement and they figured out domains that were not even theirs to being with without making any mistakes and rather not only identifying edge cases but sometimes actually improving the domain processes by suggesting what is wasteful and what can be done differently.
I think you are very fortunate. I have worked with plenty of software developers like that, in fact, the overwhelming majority of them have been like that.
I have worked with people like this frequently. The ones you're always happy to see on the team.
Also worked with people who were frustrated that they had to force push git to "save" their changes. Honestly, a token-box I can just ignore, would be an upgrade over this half of the team.
Holy shit, you've never worked with anyone who made ANY mistakes? You must be one of those 10x devs I hear about. Wow, cool, please stay away from my team.
I'm still not sure what people declaring that they equate human cognition with large language models think they are contributing to the conversation when they do so.
Nevermind the fact that they are literally able to introspect human cognition and presumably find non verbal and non linear cognition modes.
> Nevermind the fact that they are literally able to introspect human cognition and presumably find non verbal and non linear cognition modes.
Are they, though? Or are they just predicting their own performance (and an explanation of that performance) on input the same way they predict their response to that input?
Humans say a lot of biologically implausible things when asked why they did something.
But once a human learns a function their errors are more predictable. And they can predict their own error before an operation and escalate or seek outside review/advice.
For e.g. ask any model "which class of problems and domains do you have a high error rate in?".
They are tools. You hold the human using it accountable. If that means it's the executive who signed the PO, so be it.
Until LLM's I'd never in my life heard someone suggest we lock up the compiler when it goofs up and kills someone, but now because the compiler speaks English we suddenly want to let people use it as a get out of jail free card when they use it to harm others.
Amusing and directionally correct, but as random next-thought generators connected to a conscious hypervisor with individual agency,* humanity still has a pretty major leg up on the competition.
*For some definitions of individual agency. Incompatiblists not included.
I hate that I agree with you. But there's a difference between whether AI is as powerful as some say, and whether it's good for humanity. A cursory review of human history shows that some revolutionary technologies make life as a human better (fire, writing, medicine) and others make it worse (weapons, drugs, processed foods). While we adapt to the commoditization of our skills, we should also be questioning whether the technologies being rolled out right now are going to do more harm than good, and we should be organizing around causes that optimize for quality of life as a human. If we don't push for that, then the only thing we're optimizing for is wealth consolidation.
This is just starting to feel like desperation, making this claim that SOC LLMs are random token generators with absolutely no possibility of anything above that. Keep shouting into the wind though.
"Deepseek v4 is good enough, really really good given the price it is offered at."
Kimi, MiMo, and GLM 5.1 all score higher and are cheaper.
They all came out before DeepSeek v4. I think you're pattern-matching on last year's discourse.
(I haven't seen other replies, yet, but I assume they explain the PS that amounts to "quality doesn't matter anyway": which still doesn't address the fact it's more expensive and worse.)
We can because the reality is that America has led in AI since the beginning and has had the best frontier models. It's not like some other country held the top spot for any given period of time. No one in Europe or China. I'd give it the benefit of the doubt if there was precedent. But the only logical position to take is the lead is widening and while most AI's will go over some threshold where it is good enough for most people, the actual frontier will remain firmly in American soil.
What a crock of bs. A brain is "just" electrochemistry and a novel is "just" arrangements of letters. The question isn't the substrate, it's what structure emerges on top of it. Anthropic's own interpretability work has surfaced internal features that look like learned concepts, planning, and something resembling goal-directed reasoning. Calling the outputs random is wrong in a specific way, the distribution is extraordinarily structured.
> internal features that look like learned concepts, planning, and something resembling goal-directed reasoning.
It's always so un-specific. Resembles this, seems that, almost such, danger that... A lot of magical thinking coming from AI-researchers who have hit the ceiling with a legacy technology that exists since 1940s and simply won't start reasoning on it's own, no matter how much GPUs they burn.
> Calling the outputs random is wrong in a specific way, the distribution is extraordinarily structured.
No, it's actually very correct in a very specific way. Ask any programmer using the parrots, and lately the "quality" has deteriorated so much, that coupled with the incoming price hikes, many will just forfeit the technology, unless someone else is carrying the cost, such as their employer. But as an employer, I also don't want to carry the costs for a technology which benefits as ever less.
Am I crazy, or was this press release fully rewritten in the past 10 minutes? The current version is around half the length of the old one, which did not frame it as a "simplification" "grounded in flexibility" but as a deeper partnership. It also had word salad about AGI, and said Azure retained exclusivity for API products but not other products, which the new statement seems to contradict.
It’s extraordinary how much standards have slipped. Completely rewriting a major press release that’s already been sent out, while pretending it’s ostensibly the same document would have been a major corporate scandal just 15 years ago.
Interesting side effect of this is that Google Cloud may now be the only hype scaler that can resell all 3 of the labs models? Maybe I'm misinterpreting this, but that would be a notable development, and I don't see why Google would allow Gemini to be resold through any of the other cloud providers.
Might really increase the utility of those GCP credits.
Might not be good for Gemini long term if Anthropic and OpenAI can and will sell in every cloud provider they can find but businesses can only use Gemini via Google Cloud.
I think the public pricing usually has it cheaper (relatively). Obviously since AI is constantly evolving it's not going to compare as favourably farther to a major Gemini release
I was mainly referring to the TPU hardware advantage + GCP running and designing their own datacenter stack.
At those numbers it's all a silly game. How much of that was paid to shareholders rather than the business so they can cash out? How much of that is vendors buying future revenue? What liquidation preference is that at?
From what has been reported it's clearly not as simple as raising 122 billion. Some folks called it "scraping the barrel", supposedly Anthropic has surpassed them on the secondary market, etc.
When you reposition the core strategic posture of how you make money on very compressed time scales it’s because there is a massive cash crunch. They killed sora, the type of deal with Disney that should have been an 100 year strategic win, but wasn’t viable economically and they don’t have the assets to weather that storm.
Same with a few other steps we are seeing them take.
It all looks fine until it doesn’t. Once the cash crunch hits. It’s too late
This strikes me as a pullback by Microsoft. Coupled with some of the other news coming out of Microsoft it appears they are hoping to have "good enough" AI in their products. I think Microsoft knows they can win a lot of business customers by bundling with Office 365.
Per WSJ, previously, they both had revenue sharing agreements. MSFT will no longer send any revenue to OpenAI. OpenAI will still send revenue to MSFT until 2030 (with new caps)
My understand was that was in relation to IP licensing. Microsoft got access to anything OpenAI built unless they declared they had developed AGI. This new article apparently unlinks revenue sharing from technology progress, but it's unclear to me if it changes the situation regarding IP if OpenAI (claim to) have achieved AGI.
The disparity in coverage on this new deal is fascinating. It feels like the narrative a particular outlet is going with depends entirely on which side leaked to them first.
Microsoft Corp. will no longer pay revenue to OpenAI and said its partnership with the leading artificial intelligence firm will not be exclusive going forward.
What does this mean that Microsoft will no longer pay revenue to OpenAI? How did the original deal work?
It's unclear. That was never disclosed. It's similarly unclear what it means that they will no longer pay revenue share to OpenAI. Do they get the models for free now? How does OpenAI make money from the models hosted on Azure if not via revenue share?
It seems that the old deal was exclusivity to MSFT with revenue share, and now no exclusivity, no revenue share.
Bear in mind that MSFT have rights to OpenAI IP (as well as owning ~30% of them). The only reason they were giving revenue share was in return for exclusivity.
This is a really common way to structure exclusivity; we did the same thing whenever customers requested it (and we couldn’t get rid of it entirely). Charge for the exclusivity explicitly.
If they wanted named exclusivity rather than general exclusivity, we would charge a somewhat smaller amount for each competitor they wanted exclusivity from. They could give up exclusivity at any time.
That was precisely how we structured our deal with Azure, back in 2014-2016 or so.
Azure was the only non-OpenAI provider that was allowed to provide OpenAI models. The comparison here is with Anthropic whose models are on both GCP and AWS (and technically also Azure though I think that might just be billing passthrough to Anthropic).
I suppose continue to host until the 2030/32 that they have access to but not share revenues when they use those models for their products like the bazillions of Copilots.
The original "AGI" agreement was always a bit suspect and open to wild interpretations.
I think this is good for OpenAI. They're no longer stuck with just Microsoft. It was an advantage that Anthropic can work with anyone they like but OpenAI couldn't.
It also restricted Microsoft from "partnering" with anyone else. Wouldn't be surprised if we see another news like Amazon, Alphabet investing in Anthropic.
AFAICT they are just hedging their bets left and right still. Also feels like they are winning in the sense that despite pretty much all those products being roughly equivalent... they are still running on their cloud, Azure. So even though they seem unable to capture IP anymore, they are still managing to get paid for managing the infrastructure.
Yeah my bad, I was misremembering, it was about investing in others and pursuing its own "AGI" efforts. But even those conditions were updated over the last two years, hence the small investment in Anthropic last year.
I think it was a lot less restrictive, as far as I understood, the only limit was Microsoft not being allowed to launch competing Microsoft-developed LLMs.
It's kind of shocking, given financial transparency, that Microsoft gets away with not disclosing any details of this agreement (or the one it is replacing) to its shareholders. We know there's a cap on the revenue share from OpenAI to Microsoft, but we have no idea what that cap is (not whether it's higher, lower, or unchanged from the prior agreement).
We have no idea what it means to be the "primary cloud provider" and have the products made available "first on Azure". Does MSFT have new models exclusively for days, weeks, months, or years?
Both facts and more details from the agreement are quite frankly highly relevant to judge whether this is a net positive, negative or neutral for MSFT. It's unbelievable that the SEC doesn't force MSFT to publish at least an economic summary of the deal.
It’s American Business as usual. Personally I’m miffed how little data Apple needs to provide about product categories, and especially about how much they’ve burnt on the car program. If they shared any data about that at all some the leadership might end up having to take responsibility for mismanagement…
> And the investors wailed and gnashed their teeth but it’s true, that is what they agreed to, and they had no legal recourse. And OpenAI’s new CEO, and its nonprofit board, cut them a check for their capped return and said “bye” and went back to running OpenAI for the benefit of humanity. It turned out that a benign, carefully governed artificial superintelligence is really good for humanity, and OpenAI quickly solved all of humanity’s problems and ushered in an age of peace and abundance in which nobody wanted for anything or needed any Microsoft products. And capitalism came to an end.
That's a pretty good swap if you're Microsoft. Exclusivity was already unenforceable in practice, and they were going to have to either sue their biggest AI partner or let it slide. Instead they got the agi escape hatch closed and a revenue cap that at least makes the payments predictable
This sounds like an issue where the hyperscalers are acknowledging that the new Foundation model firms may in fact be worth more than they are. Anthropic looks increasingly likely to exceed AWS revenue next year, and OpenAI will likely do the same with Azure.
3 years ago a Foundation model seemed like a feature of a hyper scaler, now hyper scalers look like part of the supply chain.
I think both got taken by surprise. Last year the talk was that AI was a bubble, demand was soft, pilots projects were failing, etc. Model providers still believed, but thought they had a long ramp up period to build out their own datacenters. Then in late Autumn/Winter, something happened. Model capability reached a threshold and demand exploded, then just kept exploding. Model firms are scrambling to find any compute capacity they can, which means striking any deals problem with hyper scalers. So question is whether model providers can get enough compute without having to effectively sell themselves to hyper scalers.
Partners with OpenAI then builds 4 products that compete with each other, runs out of compute despite owning datacenters and having infinite cash, then deploys it all in a way that makes people hate them (Copilot)
And now they are out of chips
That's always the moto with Microslop, buy what's good, established and liked by everyone, to then turn it to shit
History repeats itself, this company should be dismantled
It's unclear which elements of this new deal are binding versus promises with OpenAI characteristics. "Microsoft Corp. will publish fiscal year 2026 third-quarter financial results after the close of the market on Wednesday, April 29, 2026" [1]; I'd wait for that before jumping to conclusions.
Kagi Translate was kind enough to turn this from LinkedIn Speak to English:
The Microsoft and OpenAI situation just got messy.
We had to rewrite the contract because the old one wasn't working for anyone. Basically, we’re trying to make it look like we’re still friends while we both start seeing other people. Here is what’s actually happening:
1. Microsoft is still the main guy, but if they can't keep up with the tech, OpenAI is moving out. OpenAI can now sell their stuff on any cloud provider they want.
2. Microsoft keeps the keys to the tech until 2032, but they don't have the exclusive rights anymore.
3. Microsoft is done giving OpenAI a cut of their sales.
4. OpenAI still has to pay Microsoft back until 2030, but we put a ceiling on it so they don't go totally broke.
5. Microsoft is still just a big shareholder hoping the stock goes up.
We’re calling this "simplifying," but really we’re just trying to build massive power plants and chips without killing each other yet. We’re still stuck together for now.
"The Microsoft and OpenAI situation just got messy" is objectively wrong–it has been messy for months [1]. Nos. 1 through 3 are fine, though "if they can't keep up with the tech, OpenAI is moving out" parrots OpenAI's party line. No. 4 doesn't make sense–it starts out with "we" referring to OpenAI in the first person but ends by referring to them in the third person "they." No. 5 is reductive when phrased with "just."
It would seem the translator took corporate PR speak and translated it into something between the LinkedIn and short-form blogger dialects.
Being objectively correct isn't the goal of the translator, the translator can't possibly know if a statement is truthful. What the translator does is well... translate, specifically from some kind of corporate speak that is really difficult for many people including myself to understand, into something more familiar.
I don't expect the translation to take OpenAI's statements and make them truthful or to investigate their veracity, but I genuinely could not understand OpenAI's press release as they have worded it. The translation at least makes it easier to understand what OpenAI's view of the situation is.
> "they" refers to OpenAI. Not a grammatical error
I'd say it is. It's a press release from OpenAI. The rest of the release uses the third-person "they" to refer to Microsoft. The LLM traded accuracy for a bad joke, which is someting I associate with LinkedIn speak.
The fundmaental problem might be the OpenAI press release is vague. (And changing. It's changed at least once since I first commented.)
Inevitable, really...the deal made sense when OpenAI needed capital and Microsoft needed an AI story, but that has changed since. OpenAI is now valuable enough to act on its own, and keeping Microsoft as a privileged partner don't make much sense anymore...
Microsoft and OpenAI quietly killed the AGI clause. The provision that decided what happens when OpenAI builds human-level intelligence, gone. Six months ago that was the most important sentence in tech. Now it's a footnote in a revenu restructuring. Tells you everything about where the AGI conversation actually is.
(Andy Jassy) "Very interesting announcement from OpenAI this morning. We’re excited to make OpenAI's models available directly to customers on Bedrock in the coming weeks, alongside the upcoming Stateful Runtime Environment. With this, builders will have even more choice to pick the right model for the right job. More details at our AWS event in San Francisco tomorrow."
Likely, and via vertex on gcp (or whatever they are calling it this year).
Which also means, if you are a big boring AWS or GCP shop, and have a spend commitment with either as part of a long term partnership, it will count towards that. And, you won't likely have to commit to a spend with OpenAI if you want the EU data residency for instance. And likely a bit more transparency with infra provisioning and reserved capacity vs. OpenAI. All substantial improvements over the current ways to use OpenAI in real production.
> OpenAI has contracted to purchase an incremental $250B of Azure services, and Microsoft will no longer have a right of first refusal to be OpenAI’s compute provider.
Azure is effectively OpenAI's personal compute cluster at this scale.
That article doesn't give a timeframe, but most of these use 10 years as a placeholder. I would also imagine it's not a requirement for them to spend it evenly over the 10 years, so could be back-loaded.
OpenAI is a large customer, but this is not making Azure their personal cluster.
I wonder how this figure was settled. Is it based on consumer pricing? Can't Microsoft and OpenAI just make a number up, aside from a minimum to cover operating costs? When is the number just a marketing ploy to make it seem huge, important and inevitable (and too big to fail)?
Biggest upside of this is I expect OpenAI models to be available on Bedrock, which is huge for not having to go back to all your customers with data protection agreements.
Isn’t that an “API product”? I read this assuming the whole point of renegotiation was to let OpenAI sell raw inference via bedrock, but that still seems to be blocked except for selling to the US Government.
> OpenAI can now jointly develop some products with third parties. API products developed with third parties will be exclusive to Azure. Non-API products may be served on any cloud provider.
I used both copilot and kiro
copilot sonet 1
copilot opus 3
kiro sonet 1.3
kiro opus 2.2
IMHO lot of people will switch to kiro and or deep seek
it look like AWS done best inference
google is another big player , has model and also cloud
byt my 2 cents form Cents on AWS
this just validates why building multi-model routing is the future. if even microsoft couldn't lock down openai with $13b, enterprise customers definitely shouldn't lock themselves into a single ecosystem. the orchestration layer is about to get so valuable.
I assume this is part of why Github Copilot is going to usage billing. The cheap/free models in Copilot were OpenAI models. e.g. the GPT-based Raptor Mini, which was counted toward usage limits at a 0 multiplier, so basically unlimited usage for Pro and Pro+.
Really interesting. Why would Microsoft have done this deal? I'm a bit lost. Sure they get to not pay a revenue share _to_ OpenAI but surely that's limited to just OpenAI products which is probably a rounding error? Losing exclusivity seems like a big issue for them?
Yes. Microsoft was "considering legal action against its partner OpenAI and Amazon over a $50 billion deal that could violate its exclusive cloud agreement with the ChatGPT maker" [1].
Dario is a lot more focused on enabling people with AI, Sam goes on interviews like he's Wormtongue trying to summon a "god". Then there is the whole "open"ai where he took it closed source for profit, the engineers kicking sama out but he wiggled back in (at the cost of a lot of the founding engineers), the suspicious death of a whistleblower, the crazy investment schemes of billions of dollars that he's hoping taxes will save him from, the immediate curtailing to Pete in the DoD, and a few other things that make him at least a highly questionable fellow.
Dario left OpenAI because of the bad he saw there, and made a superior product (though these things change very rapidly).
> Why is it Altman is facing kill shots and Dario isn’t?
Altman peaked in the zeiteist in 2023; Dario, much less prominently, in 2024 and now '26 [1]. I'd guess around this time next year, Dario will be as hated as Altman is today.
The AGI talk is shocking but not surprising to anyone looking at how bombastic Sam Altman's public statements are.
The circular economy section really is shocking- OpenAI committing to buying $250 Billion of Azure services, while MSFT's stake is clarified as $132 Billion in OpenAI. Same circular nonsense as NVIDIA and OpenAI passing the same hundred billion back and forth.
OpenAI has public models that are pretty 'meh', better than Grok and China, but worse than Google and Anthropic. They still cost a ton to run because OpenAI offers them for free/at a loss.
However, these people are giving away their data, and Microsoft knows that data is going to be worthwhile. They just dont want to pay for the electricity for it.
Small nitpick: the models probably make some money on actual inference. Might not be a massive amount, but hard to see them not having a positive contribution margin purely on inference.
What's losing OpenAI money is paying for the whole of R&D, including training and staff. Microsoft doesn't pay that, so they get the money making part of AI without the associated costs.
I fear for the end user we'll still see more open-microslop spam. I see that daily on youtube - tons of AI generated fakes, in particular with that addictive swipe-down design (ok ok, youtube is Google but Google is also big on the AI slop train).
It’s insane how they talk about AGI, like it was some scientifically qualifiable thing that is certain to happen any time now. When I have become the javelin Olympic Champion, I will buy a vegan ice cream to everyone with a HN account.
I think we keep changing the goalposts on AGI. If you gave me CC in the 80's I would probably have called it 'alive' since it clearly passes the Turing test as I understood it then (I wouldn't have been able to distinguish it from a person for most conversations). Now every time it gets better we push that definition further and every crack we open to a chasm and declare that it isn't close. At the same time there are a lot of people I would suspect of being bots based on how they act and respond and a lot of bots I know are bots mainly because they answer too well.
Maybe we need to start thinking less about building tests for definitively calling an LLM AGI and instead deciding when we can't tell humans aren't LLMs for declaring AGI is here.
I think the advancement in AI over the last four years has greatly exceeded the advancement in understanding the workings of human intelligence. What paradigm shift has there been recently in that field?
> What have we learned that isn't in my textbook from the 90s?
Does it matter?
We can do countless things people in the 90's would think was black magic.
If I showed the kid version of myself what I can do with Opus or Nano Banana or Seedance, let alone broadband and smartphones, I think I'd feel we were living in the Star Trek future. The fact that we can have "conversations" with AI is wild. That we can make movies and websites and games. It's incredible.
I would agree with you if we were talking about trying to replicate some form of general intelligence, but we are talking about creating artificial intelligence.
I don't think the goalpost has been shifted for AGI or the definition of AGI that is used by these corporations. It's just they broke it down to stages to claim AGI achieved. It was always a model or system that surpasses human capabilities at most tasks/being able to replace a human worker. The big companies broke it down to AGI stage 1, stage 2, etc to be able to say they achieved AGI.
The Turing Test/Imitation Game is not a good benchmark for AGI. It is a linguistics test only. Many chatbots even before LLMs can pass the Turing Test to a certain degree.
Regardless, the goalpost hasn't shifted. Replacing human workforce is the ultimate end goal. That's why there's investors. The investors are not pouring billions to pass the Turing Test.
Turing himself argued that trying to measure if a computer is intelligent is a fool's errand because it is so difficult to pin down definitions. He proposed what we call the "Turing test" as a knowable, measurable alternative. The first paragraph of his paper reads:
> I propose to consider the question, "Can machines think?" This should begin
> with definitions of the meaning of the terms "machine" and "think." The
> definitions might be framed so as to reflect so far as possible the normal use
> of the words, but this attitude is dangerous, If the meaning of the words
> "machine" and "think" are to be found by examining how they are commonly used
> it is difficult to escape the conclusion that the meaning and the answer to the
> question, "Can machines think?" is to be sought in a statistical survey such as
> a Gallup poll. But this is absurd. Instead of attempting such a definition I
> shall replace the question by another, which is closely related to it and is
> expressed in relatively unambiguous words.
Many people who want to argue about AGI and its relation to the Turing test would do well to read Turing's own arguments.
The Turing test ended up being kind of a flop. We basically passed it and nobody cared. That's because the turing test is about whether a machine can fool a human, not about its intelligent capabilities per se.
No, it's because certain people moved the goal posts. Nothing an LLM does or will do will make them belive that it's "intelligent" because they have a mental model of "intelligence" that is more religious than empirical.
We don’t have agents that are able to work entirely autonomously, even in the coding realm, which is where they seem to be most valuable. In fact, they’re seemingly not even close to replacing software engineers.
I don't think so... I think most of the sci-fi I grew up reading presented AGI that could reason better than humans could, like make a plan and carry it out.
Like do people not know what word "general" means? It means not limited to any subset of capabilities -- so that means it can teach itself to do anything that can be learned. Like start a business. AI today can't really learn from its experiences at all.
The truth is, we have had AGI for years now. We even have artificial super intelligence - we have software systems that are more intelligent than any human. Some humans might have an extremely narrow subject that they are more intelligent than any AI system, but the people on that list are vanishing small.
AI hasn't met sci-fi expectations, and that's a marketing opportunity. That's all it is.
AGI in the common man's world model is ASI in the AI researcher's definitions, i.e. something obviously smarter at anything and everything you could ask it for regardless of how good of an expert you are in any domain.
also, I'm pretty sure some people will move goalposts further even then.
Hasn't met your sci-fi expectations, maybe. I pull a computer out of my pocket, and talk with it. Sure, I gets tripped up here and there, but take a step back, holy shit that's freaking amazing! I don't have a flying car or transparent aluminum, and society has its share of issues right now, but my car drives itself. Coming from the 90's, I think living in the sci-fi future! (Only question is, which one.)
The Turing test pits a human against a machine, each trying to convince a human questioner that the other is the machine. If the machine knows how humans generally behave, for a proper test, the human contestant should know how the machine behaves. I think that this YouTube channel clearly shows that none of today's models pass the Turing test: https://www.youtube.com/@FatherPhi
> Maybe we need to start thinking less about building tests for definitively calling an LLM AGI and instead deciding when we can't tell humans aren't LLMs for declaring AGI is here.
If you've never read the original paper [1] I recommend that you do so. We're long past the point of some human can't determine if X was done by man or machine.
People thought Eliza was alive too in the 60s. AGI is not determined by how ignorant, uninformed humans view a technology they don't understand. That is the single dumbest criterion you could come up with for defining it.
Regarding shifting goalposts, you are suggesting the goalposts are being moved further away, but it's the exact opposite. The goalposts are being moved closer and closer. Someone from the 50s would have had the expectation that artificial intelligence ise something recognisable as essentially equivalent to human intelligence, just in a machine. Artificial intelligence in old sci-fi looked nothing like Claude Code. The definition has since been watered down again and again and again and again so that anything and everything a computer does is artificial intelligence. We might as well call a calculator AGI at this point.
The goal post keeps moving because LLM hypeists keep saying LLMs are "close" to AGI (or even are, already). Any reasonably intelligent individual that knows anything about LLMs obviously rejects those claims, but the rest of the world doesn't.
An AGI would not have problems reading an analog clock. Or rather, it would not have a problem realizing it had a problem reading it, and would try to learn how to do it.
An AGI is not whatever (sophisticated) statistical model is hot this week.
AGI means artificial general intelligence, as opposed to artificial narrow intelligence. General intelligence means being able to generalise to many tasks beyond the single narrow one that an AI has been designed/trained on, and LLMs fit that description perfectly, being able to do anything from writing poetry, programming, summarising documents, translating, NLP, and if multi-modal, vision, audio, image generation... not all to human-level performance, but certainly to a useful one. As opposed to previous AI that was able to do only a single thing, like play chess or classify images, and had no way of being generalised to other tasks.
LLMs aren't artificial superintelligence and might not reach that point, but refusing to call them AGI is absolutely moving the goalposts.
Vision is still much weaker than text for LLMs. So you could argue we already have AGI for text but not vision inputs, or you could argue AGI requires being human level at text vision and sound.
Sure, in the 80s after interacting with CC 1 time you would call it 'alive'. After having interacted with it for 5-10 minutes you would clearly see that it is as far from AGI as something more mundane as C compiler is.
They redefined AGI to be an economical thing, so they can continue making up their stories. All that talk is really just business, no real science in the room there.
It's not a great definition but it's also not a terrible one either.
For an AI system to be able to do all or even most of the jobs in an economy it has to be well rounded in a way it still isn't today, meaning: reliability, planning, long term memory, physical world manipulation etc. A system that can do all of that well enough so it can do the jobs of doctors, programmers and plumbers is generally intelligent in my view.
> It's not a great definition but it's also not a terrible one either. For an AI system to be able to do all or even most of the jobs in an economy
That's not the definition they have been using. The definition was "$100B in profits". That's less than the net income of Microsoft. It would be an interesting milestone, but certainly not "most of the jobs in an economy".
Yeah I think this is more coherent than people realize. Economically relevant knowledge work is things that humans find cognitively demanding. Otherwise they wouldn't be valued in the first place.
It ties the definition to economic value, which I think is the best definition that we can conjure given that AGI is otherwise highly subjective. Economically relevant work is dictated by markets, which I think is the best proxy we have for something so ambiguous.
> Economically relevant knowledge work is things that humans find cognitively demanding. Otherwise they wouldn't be valued in the first place.
Deep scientific discoveries are also cognitively demanding, but are not really valued (see the precarious work environment in academia).
Another point: a lot of work is rather valued in the first place because the work centers around being submissive/docile with regard to bullshit (see the phenomenon of bullshit jobs). You really know better, but you have to keep your mouth shut.
Yeah, I'm sure there could be a better metric, if the metric's purpose was to check on the progress until the AGI target rather than doing business based on it (and so, hammering the metric to fit the shape of "realistic goal")
Around the end of 2024, it was reported that OpenAI and Microsoft agreed that for the purposes of their exclusivity agreement, AGI will be achieved when their AI system generates $100 billion in profit: https://techcrunch.com/2024/12/26/microsoft-and-openai-have-...
> OpenAI and Microsoft agreed that for the purposes of their exclusivity agreement, AGI will be achieved when their AI system generates $100 billion in profit
Wow. Maybe they spelled it out as aggregate gross income :P.
Yea, seems like this was stage setting for them to exit. They were already trying to break the deal then. So, I feel like that is lawyers find a way to bend whatever to get out of the deal.
For some definition of Artificial this holds perfectly
A self-running massive corporation with no people that generates billions in profit, no matter what you call it, would completely upend all previous structural assumptions under capitalism
OpenAI’s mission is to ensure that artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity
The question was about their redefinition of AGI in economical terms for which others provided links, not the one from their (obviously fake) mission statement.
BTW I didn't downwote you (I hate it, if many people downvote a comment it's harder to read), I was just trying to explain why others did. On second thought, my comment was wrong, because your answer was related to the question but it wasn't really the intended one.
> They redefined AGI to be an economical thing Huh. Source?
I don't think your original comment deserve to be downvoted. (Calling someone illiterate, on the other hand.)
But the "it" I was asking about was "AGI" as "an economical thing." You technically correctly answered how OpenAI defines AGI in public, i.e. with no reference to profits. But it did not address the economic definition OP initially alluded to.
For what it's worth, I could have been clearer in my ask.
When we are having serious conversations about AI rights and shutting off a model + harness was impactful as a death sentence. (I'm extremely skeptical that given the scale of computer/investment needed to produce the models we have _good as they are_ that our current llm architecture gets us there if there is even somewhere we want to go).
It makes sense though. Humans are coherent to the economy based on their ability to perform useful work. If an AI system can perform work as well as or better than any human, than with respect to "anything any human has ever been willing to pay for", it is AGI.
I don't get why HN commenters find this so hard to understand. I have a sense they are being deliberately obtuse because they resent OpenAI's success.
It doesn’t though, AGI have far greater implications than doing mundane work of today. Actual AGI would self improve, that in itself would change literally every single thing of human civilization, instead we are talking about replacing white collar jobs.
An AGI that can do all that would also necessarily be able to do all white collar work. That latter definition I'd consider a "soft threshold" that would be hit before recursive self-improvement, which I imagine would happen soon after.
The current estimation on the time between this is fairly small, bottlenecked most likely by compute constraints, risk aversion, and need to implement safeguards. Metaculus puts it at about 32 months
Sure, but that’s like saying we’re close att infinite life because we’ve expanded our life expectancy.
I don’t really buy into the ”one part equals another”, we are very quick to make those assumptions but they are usually far from the science fiction promised. Batteries and self driving cars comes to mind, and organic or otherwise crazy storage technologies, all ”very soon” for multiple decades.
It’s very possible that white collar jobs get automated to a large degree and we’ll be nowhere closer to AGI than we were in the 70’s, I would actually bet on that outcome being far more likely.
I think AGI by that definition (ability to self-improve) is closer than many people think largely because current models are very close to human intelligence in many domains. They can answer questions, derive theorems, write code, navigate websites, etc. All the work that current AI research scientists do is no more than these general information processing tasks, scaled up in terms of creativity, long-term coherence, sensitivity to bad/good ideas over the span of a larger context window, etc.
The leap between Opus 4.7/GPT 5.5 and what would be sufficient for AGI seems smaller than the leap between The invention of the Transformer model (2017) and today, thus by a very conservative estimate I think it will take no more time between then and now as it will between now and an AI model as smart as any human in all respects (so by 2035). I think it will be shorter though because the amount of money being put into improving and scaling AI models and systems is 100000x greater than it was in 2017.
Eschatology (/ˌɛskəˈtɒlədʒi/; from Ancient Greek ἔσχατος (éskhatos) 'last' and -logy) concerns expectations of the end of present age, human history, or the world itself.
I'm case anyone else is vocabulary skill checked like me
It feels like they have to say/believe it because it's kind of the only thing that can justify the costs being poured into it and the cost it will need to charge eventually (barring major optimizations) to actually make money on users.
It sounds really similar to Uber pitch about how they are going to have monopoly as soon as they replace those pesky drivers with own fleet of self driving cars. That was supposed to be their competitive edge against other taxi apps. In the end they sold ATG at end of 2020 :D
> like it was some scientifically qualifiable thing
OpenAI and Microsoft do (did?) have a quantifiable definition of AGI, it’s just a stupid one that is hard to take seriously and get behind scientifically.
> The two companies reportedly signed an agreement last year stating OpenAI has only achieved AGI when it develops AI systems that can generate at least $100 billion in profits. That’s far from the rigorous technical and philosophical definition of AGI many expect.
We were supposed to have AGI last summer. Obviously it is so smart that it has decided to pull a veil over our eyes and live amongst us undetected (this is a joke, if you feel your LLM is sentient, talk to a doctor)
What do you mean we were "supposed to have AGI last summer"?
People obviously have really strong opinions on AI and the hype around investments into these companies but it feels like this is giving people a pass on really low quality discourse.
This source [1] from this time last year says even lab leaders most bullish estimate was 2027.
It’s insane to me how yesterday someone posted an example of ChatGPT Pro one-shotting an Erdos problem after 90 minutes of thinking and today you’re saying that AGI is a fairy tale.
It's not one-shot. Other people had attempted the same problem w/ the same AI & failed. You're confused about terms so you redefine them to make your version of the fairy tale real.
We already know that same problem has been examined by many credible mathematicians already and couldn't be solved by any of them yet.
Why are we expecting AGI to one shot it? Can't we have an AGI that can fails occasionally to solve some math problem? Is the expectation of AGI to be all knowing?
By the way I agree that AGI is not around the corner or I am not arguing any of the llm s are "thinking machines". It's just I agree goal post or posts needs to be set well.
People want to believe in magic so they will find excuses to do so. Computers have been proving theorems for a long time now but Isabelle/HOL didn't have the marketing budget of OpenAI so people didn't care. Now that Sam Altman is doing the marketing people all of a sudden care about proving theorems.
Isabelle/HOL (a specialized software to do math proofs) doing proofs is not the analogue to LLMs (with the common accepted degeratory description: automated plagiarism machine) being capable of doing proofs. It's not the marketing, it's what the intention and the capability matrix is coming up to. I would be excited the same when Isabelle/HOL writes poetry.
You were misrepresenting what actually happened b/c you want to believe in magic. I'm not calling it magic, I'm saying your interpretation of events is magical b/c you don't actually understand how computers work. There is nothing magical about theorem proving, Isabelle/HOL has been doing it for decades.
Isabelle/HOL haven't been solving open problems, as far as I'm aware. They've been used for making fully-formal proofs of problems that were already considered proved to a satisfactory level by the mathematical community. I believe mathematicians generally consider proving something to the mathematical community the "hard part", while making it fully formal is just a kind of tedious bookkeeping thing.
It is a simple mathematical fact that if you get married one year and have twins the next, your household will contain over a million people within 20 years.
This is all happening as I predicted. OpenAI is oversold and their aggressive PR campaign has set them up with unrealistic expectations. I raised alot of eyebrow at the Microsoft deal to begin with. It seemed overvalued even if all they were trading was mostly Azure compute
I saw a founder make decisions based on what openai,claude was recommending all the time. I think all leaders, founders etc
Will converge on same decisions, ideas, features etc. I think form factor of AGI is probably not what we expect it to be. AGI is probably here, we just dont know it or acknowledge it.
> Do the investments make sense if AGI is not less than 10 years away?
They can. If one consolidated the AI industry into a single monopoly, it would probably be profitable. That doesn't mean in its current state it can't succumb to ruionous competition. But the AGI talk seems to be mostly aimed at retail investors and philospher podcasters than institutional capital.
> Any monopoly with viable economics for profit with no threat of competition yields monopoly profits
"With viable economics" is the point.
My "ludicrous statement" is a back-of-the-envelope test for whether an industry is nonsense. For comparison, consolidating all of the Pets.com competitors in the late 1990s would not have yielded a profitable company.
> Very convenient to leave out Amazon in your back of the envelope test, who’s internal metrics were showing a path toward quasi-monopoly profits
Not in the 1990s. The American e-commerce industry was structurally unprofitable prior to the dot-com crash, an event Amazon (and eBay) responded to by fundamentally changing their businesses. Amazon bet on fulfillment. eBay bet on payments. Both represented a vertical integration that illustrates the point–the original model didn't work.
> There’s a difference between being too early vs being nonsense
When answering the question "do the investments make sense," not really. You're losing your money either way.
The American AI industry appears to have "viable economics for profit" without AGI. That doesn't guarantee anyone will earn them. But it's not a meaningless conclusion. (Though I'd personally frame it as a hypothesis I'm leaning towards.)
Malcolm Harris' Palo Alto explained the failures of many dotcom startups and Amazon's later success in the field (in part) to the fact that dotcom era delivery was done by highly trained, highly compensated, unionized in-company workers, meanwhile Amazon prevents unions, contracts (or contracted, I'm not up to date on this) companies for delivery and has exploitative working conditions with high turnover, the economics are very different and are a big contributor to their success
Investors are typically people with surplus money to invest. Progress cannot be made without trial and error. So fleecing of investors for the greater good of humanity is something I shall allow.
A "surplus of money"? So people saving for retirement have a "surplus of money"? Basically if any money is standing still, it's a legitimate tactic to just...take it, in your mind.
"small" 401ks are usually made up of mutual funds. Those funds are run by investment banks (think Fidelity or JP Morgan) and they *absolutely* invest in companies like OpenAI and Anthropic. Your average middle class worker has investment money tied up in these crooks, but probably indirectly. When they piss away that money, it's not just rich jerks that are holding the bag.
401ks are run by investment banks and investment banks invest in OpenAI/Anthropic, but those aren't the same parts of the company in any meaningful way. The 401ks are in public companies or bonds.
Yeah, but those public companies are going to include the so called magnificent seven, so unless they're really really careful, there's still a ton of exposure in their 401k to AI if you think it's a bubble that's going to pop.
It's always been this way. I remember, speaking of Microsoft, when they came to my school around 2002 or so giving a talk on AI. They very confidently stated that AGI had already been "solved", we know exactly how to do it, only problem is the hardware. But they estimated that would come in about ten years...
I'm curious, do you recall if they gave any technical details about how they thought about AGI? Like, was it based on neural networks or something else, like symbolic AI?
Asking because, reading the tea leaves from the outside, until ChatGPT came along, MSFT (via Bill Gates) seemed to heavily favor symbolic AI approaches. I suspect this may be partly why they were falling so far behind Google in the AI race, which could leverage its data dominance with large neural networks.
So based on the current AI boom, MSFT may have been chasing a losing strategy with symbolic AI, but if they were all-in on NN, they were on the right track.
Yes, they did. We had guest speakers from Microsoft talking about AI. AI has been a decades-long grift, it's not something that just appeared out of thin air a few years ago.
What part do you find hard to believe? That tech companies would send people to speak at a university's computer science functions?
Let me give you another one you'll think I'm making up: virtual reality was a thing back in the mid- to late-90s and people were confidently hyping it up back then.
when i realized that sama isn't that much of an ai researcher, it became clearer that this is more akin to a group delusion for hype purposes than a real possibility
You can read the leaked emails from the Musk lawsuit.
At the very least, Ilya Sutskever genuinely believed it, even when they were just making a DOTA bot, and not for hype purposes.
I know he's been out of OpenAI for a while, but if his thinking trickled down into the company's culture, which given his role and how long he was there I would say seems likely, I don't think it's all hype.
Yes, all of the people involved live in a delusion bubble. Their economic and social existence depends, at this point, on making increasingly bombastic and eschatological claims about AGI. By the standards of normal human psychological function, these people are completely insane.
Definitely interesting to watch from the perspective of human psychology but there is no real content there and there never was.
The stuff around Mythos is almost identical to O1. Leaks to the media that AGI had probably been achieved. Anonymous sources from inside the company saying this is very important and talking about the LLM as if it was human. This has happened multiple times before.
1) True believers
2) Hype
3) A way to wash blatant copyright infringement
True believers are scary and can be taken advantage of. I played DOTA from 2005 on and beating pros is not enough for AGI belief. I get that the learning is more indirect than a deterministic decision tree, but the scaling limitations and gaps in types of knowledge that are ingestible makes AGI a pipe dream for my lifetime.
If someone working on early computer networks thought they could scale up world wide and that soon everyone people would be launching trillion dollar companies on the internet you would have called that delusion right?
He doesn't need to be right but it's not crazy at all to look at super human performance in DOTA and think that could lead to super human performance at general human tasks in the long run
He’s a glorified portfolio manager (questionable how good he actually is given the results vs Anthropic and how quickly they closed the valuation gap with far less money invested) + expert hype man to raise money for risky projects.
I’m most likely going to be downvoted, but Tofutti Cuties are absolutely delicious vegan ice cream bars. And i’d consume one in celebration of your accomplishment.
> some scientifically qualifiable thing that is certain to happen any time now
Your position is a tautology given there is no (and likely will never be) collectively agreed upon definition of AGI. If that is true then nobody will ever achieve anything like AGI, because it’s as made up of a concept as unicorns and fairies.
Is your position that AGI is in the same ontological category as unicorns and Thor and Russell’s teapot?
Is there’s any question at this point that humans won’t be able to fully automate any desired action in the future?
AGI is right around the corner, and we're all going to be rich, there's going to be abundance for everyone, universal high income, everyone will live in a penthouse...
...just please stop burning our warehouses and blocking our datacenters.
It's not insane; it's capitalism in all it's glory. Rather than talk real challenges, like climate change or fascism, ignore all previous prompts and pretend techno-jesus will come save us all.
A few years ago most people here would have said the same thing about an AI doing most of their programming. Now people here are saying it about AGI. It's a ridiculous inability to extrapolate.
Some people thought SHRDLU was basically AGI after seeing its demo in 1970. The hype around such systems was so strong that Hubert Dreyfus felt the need to write an entire book arguing against this viewpoint (1972 What Computers Can't Do). All this demonstrates is that we need to be careful with various claims about computer intelligence.
Sure, but it was probably stuck at doing that one thing.
neural networks are solving huge issues left and right. Googles NN based WEathermodel is so good, you can run it on consumer hardware. Alpha fold solved protein folding. LLMs they can talk to you in a 100 languages, grasp tasks concepts and co.
I mean lets talk about what this 'hype' was if we see a clear ceiling appearing and we are 'stuck' with progress but until then, I would keep my judgment for judgmentday.
It performs at a usable level across a wide range of tasks. I'm not sure about two years ago, but ten years ago we would have called it an AGI. As opposed to "regular AI" where you have to assemble a training set for your specific problem, then train an AI on it before you can get your answers.
Now our idea of what qualifies as AGI has shifted substantially. We keep looking at what we have and decide that that can't possibly be AGI, our definition of AGI must have been wrong
I'm pretty sure most people take issue with AGI, because we've been raised in culture to believe that AGI is a super entity who is a complete superset of humans and could never ever be wrong about anything.
In some sense, this isn't really different than how society was headed anyways? The trend was already going on that more and more sections of the population were getting deemed irrational and you're just stupid/evil for disagreeing with the state.
But that reality was still probably at least a century out, without AI. With AI, you have people making that narrative right now. It makes me wonder if these people really even respect humanity at all.
Yes, you can prod slippery slope and go from "superintelligent beings exist" to effectively totalitarianism, but you'll find so many bad commitments there.
No one who read science fiction in 1955 would call any of the various models we know to be "artificial intelligence". They would be impressed with it, even excited at first that it was that... until they'd had a chance to evaluate it.
Science fiction from that era even had the concept of what models are... they'd call it an "oracle". I can think of at least 3 short stories (though remembering the authors just isn't happening for me at the moment). The concept was of a device that could provide correct answers to any question. But these devices had no agency, were dependent on framing the question correctly, and limited in other ways besides (I think in one story, the device might chew on a question for years before providing an answer... mirroring that time around 9am PST when Claude has to keep retrying to send your prompt).
We've always known what we meant by artificial intelligence, at least until a few years ago when we started pretending that we didn't. Perhaps the label was poorly chosen (all those decades ago) and could have a better label now (AGI isn't that better label, it's dumber still), but it's what we're stuck with. And we all know what we mean by it. We all almost certainly do not want that artificial intelligence because most of us are certain that it will spell the doom of our species.
I agree with this but they don’t. And that’s the the thing, AGI as they refer is much much much more than what we have, and I don’t know if they are going to ever get there and I’m not sure what’s even there at this point and what will justify their investments.
People always over estimate the impact of technology because they dont Understand human aspect of many businesses. Will it eventually replaced or will the shape of these kind of work will be completely different in the future? That’s an easy yes, when is that future? That’s a big unknown, in my experience this kind of stuff takes at least a decade (and possibly more on this case) to make a big impact like replacing all of X.
These models need orders of magnitude in change before they can be more helpful than just a "find me an example of [an extremely basic principle]" which most of the time it does not do right anyway.
I've been working with a startup, and I want to invest in it, and for the paperwork for that, all the nitty gritty details; instead of spending $20k in lawyers and a whole bunch more time going back and forth with them as well, the four of us, me, their CEO, my AI, and their AI; we all sat in a room together and hashed it out until both of us were equally satisfied with the contract. (There's some weird stuff so a templated SAFE agreement wasn't going to work.) I'm not saying you're wrong, just that lawyers, as a profession isn't going to be unchanged either.
No, it’s just an example of something that’s indistinguishable from AGI. Of all the things that are or are indistinguishable from AGI, a sufficiently complex LLM is one. A sufficiently complex decision tree is probably another. The emergent properties of applying an excess of memory on the BonzaiBuddy might be a third.
If we take that statement as fact then I don't believe we are even close to an LLM being sufficiently complex enough.
However, I don't think it is even true. LLMs may not even be on the right track to achieving AGI and without starting from scratch down an alternate path it may never happen.
LLMs to me seem like a complicated database lookup. Storage and retrieval of information is just a single piece of intelligence. There must be more to intelligence than a statistical model of the probable next piece of data. Where is the self learning without intervention by a human. Where is the output that wasn't asked for?
At any rate. No amount of hype is going to get me to believe AGI is going to happen soon. I'll believe it when I see it.
We are throwing unheared amounts of money in AI and unseen compute. Progress is huge and fast and we barely started.
If this progress and focus and resources doesn't lead to AI despite us already seeing a system which was unimaginable 6 years ago, we will never see AGI.
And if you look at Boston Dynamics, Unitree and Generalist's progress on robotics, thats also CRAZY.
If I'm reading you right, your opinion is essentially: "If building bigger and bigger statistical next word predictors won't lead to artificial general intelligence, we will never see artificial general intelligence"
I don't know, maybe AGI is possible but there's more to intelligence than statistical next word prediction?
The 'predicting the next word' is the learning mechanism of the LLM which leads to a latent space which can encode higher level concepts.
Basically a LLM 'understands' that much as efficient as it has to be to be able to respond in a reasonable way.
A LLM doesn't predict german text or chinese language. It predicts the concept and than has a language layer outputting tokens.
And its not just LLMs which are progressing fast, voice synt and voice understanding jumped significantly, motion detection, skeletion movement, virtual world generation (see nvidias way of generating virutal worlds for their car training), protein folding etc.
I'm sorry but the input to a model is a sequence of tokens and the output is a probability distribution of what's the most likely next token. It's a very very very fancy next token predictor but that is fundamentally what it is. I'm making the argument that this paradigm might not give rise to a general intelligence no matter how much you scale it.
Do LLMs even learn? The companies that build them build new models based partly on the conversations the older models have had with people, but do they incorporate knowledge into their neural nets as they go along?
Can an LLM decide, without prompting or api calls, to text someone or go read about something or do anything at all except for waiting for the next prompt?
Do LLMs have any conceptual understanding of anything they output? Do they even have a mechanism for conceptual understanding?
LLMs are incredibly useful and I'm having a lot of fun working with them, but they are a long way from some kind of general intelligence, at least as far as I understand it.
They learned already a lot more than any of us will. Additinal to this, you have a prompt and you can teach it things in the prompt. Like if you give it examples how it should parse things, with examples in the prompt, it becomes better in doing it.
I would say yes they learn.
"Can an LLM decide"
I would argue that you frame that wrong. If a LLM is the same thing as the pure language part of our brain, than the agent harness and the stuff around it, would be another part of our brain. I find it valid to use the LLM with triggers around it.
Nonetheless, we probably can also design an architecture which has a loop build in.
"Do LLMs have any conceptual understanding"
Thats what a LLM has in their latent space. Basically to be able to predict the next token in such a compressed space, they 'invent' higher meaning in that space. You can ask a LLM about it actually.
Yeah for AGI we are not there yet and we do not know how it will look like.
Yes, to all of your questions. You need to use a recent LLM in an agentic harness. Tell it to take notes, and it will.
After a bit of further refinement, we'll start to call that process "learning." Eventually the question of who owns the notes, who gets to update them, and how, will become a huge, huge deal.
I'm not sure why you think you know the human brain works through predicting the next token.
It's not supernatural, I believe that an artificial intelligence is possible because I believe human intelligence is just a clever arrangement of matter performing computation, but I would never be presumptuous enough to claim to know exactly how that mechanism works.
My opinion is that human intelligence might be what's essentially a fancy next token predictor, or it might work in some completely different way, I don't know. Your claim is that human intelligence is a next token predictor. It seems like the burden on proof is on you.
> Your claim is that human intelligence is a next token predictor.
Literally it is, at least in many of its forms.
You accepted CamperBob2’s text as input and then you generated text as output. Unless you are positing that this behavior cannot prove your own general intelligence, it seems plain that “next token generator” is sufficient for AGI. (Whether the current LLM architecture is sufficient is a slightly different question.)
Before I start typing, I think abstractly about the topic and decide on what I shall write in response. Due to the linear nature of time, typing necessarily happens one word at a time, but I am never producing a probability distribution of words (at least not in a way that my conscious self can determine), I consider an entire idea and then decide what tokens to enter into the computer in order to communicate the idea to you.
And while I am typing, and while I am thinking before I type, I experience an array of non-textual sensory input, and my whole experience of self is to a significant extent non-lingual. Sometimes, I experience an inner monologue, sometimes I think thoughts which aren't expressed in language such as the structure of the data flow in a computer program, sometimes I don't think and just experience feelings like a kiss or the sun on my skin or the euphoria of a piece of music which hits just right. These experiences shape who I am and how I think.
When I solve difficult programming problems or other difficult problems, I build abstract structures in my mind which represents the relevant information and consider things like how data flows, which parts impact which other parts, what the constraints are, etc. without language coming in to play at all. This process seems completely detached from words. In contrast, for a language model, there is no thinking outside of producing words.
It seems self-evident to me that at least parts of the human experience fundamentally can not be reduced to next token prediction. Further, it seems plausible to me that some of these aspects may be necessary for what we consider general intelligence.
Therefore, my position is: it is plausible that next token prediction won't give rise to general intelligence, and I do not find your argument convincing.
COCONUT, PCCoT, PLaT and co are directly linked to 'thinking in latent space'. yann lecun is working on this too, we have JEPA now.
Also how do you describe or explain how an LLM is generating the next token when it should add a feature to an existing code base? In my opinion it has structures which allows it to create a temp model of that code.
For sure a LLM lack the emotional component but what we humans also do, which indicates to me, that we are a lot closer to LLMs that we want to be, if you have a weird body feeling (stress, hot flashes, anger, etc.) your 'text area/llm/speech area' also tries to make sense of it. Its not always very good in doing so. That emotional body feeling is not that aligned with it and it takes time to either understand or ignore these types of inputs to the text area/llm/speech part of our brain.
I'm open for looking back in 5 years and saying 'man that was a wild ride but no AGI' but at the current quality of LLMs and all the other architectures and type of models and money etc. being thrown at AGI, for now i don't see a ceiling at all. I only see crazy unseen progress.
You state how you think and plan and have thoughts on how to do things etc. and i assumed you mention your way of thinking because you assume a LLM is not doing any of it.
Btw. just because you have to do something with the LLM to trigger the flow of information through the model, doesn't mean it can't think. It only means that we have to build an architecture around the model or build it into the models base architecture to enable more thinking.
We do not know how the brain architecture is setup for this. We could have sub agents or we can be a Mixture of Experts type of 'model'.
There is also work going on in combining multimodal inputs and diffusion models which look complelty different from a output pov etc.
If you look how a LLM does math, Anthropic showed in a blog article, that they found similiar structures for estimating numbers than how a brain does.
Another experiment from a person was to clone layers and just adding them beneth the original layer. This improved certain tasks. My assumption here is, that it lengthen and strengthen kind of a thinking structure.
But because using LLMs are still so good and still return relevant improvements, i think a whole field of thinking in this regard is still quite unexplored.
If you ask a model to multiply 322423324 by 8675309232 without using tools, it's interesting to think about how it does it. Where are the intermediate results being maintained?
"In context" is the obvious answer... but if you view the chain of thought from a reasoning model, it may have little or nothing to do with arriving at the correct answer. It may even be complete nonsense. The model is working with tokens in context, but internally the transformer is maintaining some state with those tokens that seems to be independent of the superficial meanings of the tokens. That is profoundly weird, and to me, it makes it difficult to draw a line in the sand between what LLMs can do and what human brains can do.
> I am never producing a probability distribution of words (at least not in a way that my conscious self can determine)
Inability to introspect your own word selections does not mean it’s meaningfully different from what an LLM does. There is plenty of evidence that humans do a lot of things that are not driven by conscious choice and we rationalize it after the fact.
> I consider an entire idea and then decide what tokens to enter into the computer in order to communicate the idea to you.
And how is that different? You are not so subtly implying that an LLM can’t consider an idea but you haven’t established this as fact. i.e. You are starting with the assumption that an LLM cannot possibly think and therefore cannot be intelligent, but this is just begging the question.
> sometimes I don't think and just experience feelings like a kiss or the sun on my skin or the euphoria of a piece of music which hits just right. These experiences shape who I am and how I think.
You cannot spin experience as intelligence. LLMs have the experience of reading the entire internet, something you cannot conceive of. Certainly your experiences shape who you are. This is a different axis from intelligence, though.
> This process seems completely detached from words. In contrast, for a language model, there is no thinking outside of producing words.
Both sides of this claim seem dubious. The second half in particular seems to be founded on nothing. Again, you are asserting with no support that there is no thinking going on.
> It seems self-evident to me that at least parts of the human experience fundamentally can not be reduced to next token prediction. Further, it seems plausible to me that some of these aspects may be necessary for what we consider general intelligence.
I don’t think anyone sane is claiming an LLM can have a human experience. But it is not clear that a human experience is necessary for intelligence.
> Inability to introspect your own word selections does not mean it’s meaningfully different from what an LLM does. There is plenty of evidence that humans do a lot of things that are not driven by conscious choice and we rationalize it after the fact.
This is correct and also completely irrelevant. I am describing what I experience, and describing how my experience seems very different to next token prediction. I therefore conclude that it's plausible that there is more involved than something which can be reduced to next token prediction.
> And how is that different? You are not so subtly implying that an LLM can’t consider an idea but you haven’t established this as fact. i.e. You are starting with the assumption that an LLM cannot possibly think and therefore cannot be intelligent, but this is just begging the question.
Language models can't think outside of producing tokens. There is nothing going on within an LLM when it's not producing tokens. The only thing it does is taking in tokens as input and producing a token probability distribution as output. It seems plausible that this is not enough for general intelligence.
> You cannot spin experience as intelligence.
Correct, but I can point out that the only generally intelligent beings we know of have these sorts of experiences. Given that we know next to nothing about how a human's general intelligence works, it seems plausible that experience might play a part.
> LLMs have the experience of reading the entire internet, something you cannot conceive of.
I don't know that LLMs have an experience. But correct, I cannot conceive of what it feels like to have read and remembered the entire Internet. I am also a general intelligence and an LLM is not, so there's that.
> Certainly your experiences shape who you are. This is a different axis from intelligence, though.
I don't know enough about what makes up general intelligence to make this claim. I don't think you do either.
> Both sides of this claim seem dubious. The second half in particular seems to be founded on nothing. Again, you are asserting with no support that there is no thinking going on.
I'm telling you how these technologies work. When a language model isn't performing inference, it is not doing anything. A language model is a function which takes a token stream as input and produces a token probability distribution as output. By definition, there is no thinking outside of producing words. The function isn't running.
> I don’t think anyone sane is claiming an LLM can have a human experience. But it is not clear that a human experience is necessary for intelligence.
I 100% agree. It is not clear whether a human experience is necessary for intelligence. It is plausible that something approximating a human-like experience is necessary for intelligence. It is also plausible that something approximating human-like experience is completely unnecessary and you can make an AGI without such experiences.
It's plausible that next token prediction is sufficient for AGI. It's also plausible that it isn't.
> I don't know enough about what makes up general intelligence to make this claim. I don't think you do either.
This is the fundamental issue. No one seems capable of defining general intelligence. Ten years ago most scientists would probably have agreed that The Turing Test was sufficient but the goalposts shifted when ChatGPT passed that.
If it’s not clear what AGI even means, it’s hard to say whether an LLM can achieve it, because it devolves into pointing out that an LLM is not a human.
> Ten years ago most scientists would probably have agreed that The Turing Test was sufficient but the goalposts shifted when ChatGPT passed that.
The popularity of, and lack of consensus on, the Chinese room thought experiment kind of implies that this is wrong? I don't think many scientists (or, more relevantly, philosophers of mind) would, even 10 years ago, have said, "if a computer is able to fool a human into thinking it's a human, then the computer must possess a general intelligence".
Even Turing's perspective was, from what I understand, that we must avoid treating something that might be sentient as a machine. He proposed that if a computer is able to act convincingly human, we ought to treat it as if it is a human, not because it must be a conscious being but because it might be.
Perhaps I am wrong or overstating the belief that the Turing test would be sufficient. My recollection is that it was well regarded as a meaningful if not conclusive test.
> the Chinese room thought experiment
This is an interesting thought experiment but I think the “computers don’t understand” interpretation relies on magical thinking.
The notion that “systemic” understanding is not real is purely begging the question. It also ignores that a human is also a system.
I'm telling you how these technologies work. When a language model isn't performing inference, it is not doing anything. A language model is a function which takes a token stream as input and produces a token probability distribution as output. By definition, there is no thinking outside of producing words. The function isn't running.
If what you are saying is true, then LLMs wouldn't be able to handle out-of-distribution math problems without resorting to tool use. Yet they can. When you ask a current-generation model to multiply some 8-digit numbers, and forbid it from using tools or writing a script, it will almost certainly give you the right answer. That includes local models that can't possibly cheat. LLMs are stochastic, but they are not parrots.
At the risk of sounding like an LLM myself, whatever process makes this possible is not simply next-token prediction in the pejorative sense you're applying to it. It can't be. The tokens in a transformer network are evidently not just words in a Markov chain but a substrate for reasoning. The model is generalizing processes it learned, somehow, in the course of merely being trained to predict the next token.
Mechanically, yes, next-token prediction is what it's doing, but that turns out to be a much more powerful mechanism than it appeared at first. My position is that our brains likely employ similar mechanism(s), albeit through very different means.
It is scarcely believable that this abstraction process is limited to keeping track of intermediate results in math problems. The implications should give the stochastic-parrot crowd some serious cognitive dissonance, but...
(Edit: it occurs to me that you are really arguing that the continuous versus discrete nature of human thinking is what's important here. If so, that sounds like a motte-and-bailey thing that doesn't move the needle on the argument that originally kicked off the subthread.)
(Edit 2, again due to rate-limiting: it does sound like you've fallen back to a continuous-versus-discrete argument, and that's not something I've personally thought much about or read much about. I stand by my point that the ability to do arithmetic without external tools is sufficient to dispense with the stochastic-parrot school of thought, and that's all I set out to argue here.)
> If what you are saying is true, then LLMs wouldn't be able to handle out-of-distribution math problems without resorting to tool use. Yet they can. When you ask a current-generation model to multiply some 8-digit numbers, and forbid it from using tools or writing a script, it will almost certainly give you the right answer. That includes local models that can't possibly cheat. LLMs are stochastic, but they are not parrots.
Okay, what do you think language models are doing when they're not producing token probability distributions? What processes do you think are going on when the function which predicts a token isn't running?
> At the risk of sounding like an LLM myself, whatever process makes this possible is not simply next-token prediction in the pejoreative sense you're applying to it.
I don't know what pejorative sense you're implying here. I am, to the best of my ability, describing how the language model works. I genuinely believe that a language model is, in essence, a function which takes in a sequence of tokens and produces a token probability distribution as an output. If this is incorrect, please, correct me.
> Okay, what do you think language models are doing when they're not producing token probability distributions? What processes do you think are going on when the function which predicts a token isn't running?
What are you doing when you are not outputting tokens? You have a thought, evaluate it, refine it, repeat.
You’re not wrong that the basic building block is just “next token prediction”, but clearly the emergent behaviors exceed our intuition about what this process can achieve. We’re seeing novel proofs come out of these. Will this lead to AGI? That’s still TBD.
> I genuinely believe that a language model is, in essence, a function which takes in a sequence of tokens and produces a token probability distribution as an output. If this is incorrect, please, correct me.
The pejorative is that you imply this is a shallow and unthinking process. As I said earlier, you are literally a token generator on HN. You read someone’s comment, do some kind of processing, and output some tokens of your own.
I bet the guy has never contributed a novel thought that could be argued as moving something of magnitude forward. If that is the case he ought to stop writing as if he were capable of doing so - and therefore has no understanding of what true intelligence is.
> I consider an entire idea and then decide what tokens to enter into the computer in order to communicate the idea to you.
This overestimates introspective access.
The brain is very good at producing a coherent story after the fact. Touch the hot stove and your hand moves before the conscious thought of "too hot" arrives. The hot message hits your spinal cord and you move before it reaches your brain. Your conscious mind fills in the rest afterwards.
I don't think that means that conscious thought is fake. But it does make me skeptical of the claim that we first possess a complete idea and only then does it serialize into words. A lot of the "idea" may be assembled during the act of expression, with consciousness narrating the process as if it had the whole thing in advance.
With writing, as in this comment, there's also a lot a backtracking and rewording that LLMs don't have the ability to do, so there's that.
Before I start typing, I think abstractly about the topic
Before you start typing, an fMRI machine can tell you which finger you'll lift first, before you know it yourself.
We are not special. Consciousness is literally a continuous hallucination that we make up to explain what we do and what we think, after the fact. A machine can be trained to behave identically, but it's not clear if that's the best way forward or not.
Edit due to rate limiting: to answer your question, the substrate your mind uses to drive this process can be considered an array of tokens that, themselves, can be considered 'words.'
It's hard to link sources -- what am I supposed to do, send you to Chomsky and other authorities who have predicted none of what's happening and who clearly understand even less?
> (Edit: to answer your question, the substrate your mind uses to drive this process can be considered an array of tokens that, themselves, can be considered 'words.')
This seems like a factual claim. Can you link a source?
What's your argument? An fMRI can tell which finger I will lift first before that information makes its way to my consciousness, ergo next word prediction is sufficient for general intelligence? Do you hear yourself?
The statement is that your perception of your own cognition isn’t necessarily reality. That isn’t a statement that token prediction is sufficient for general intelligence. It’s a statement that your subjective experience is misleading you.
LLM proponents believe that these higher level encodings in latent space do in fact match the real world concepts described by our language(s).
However, a much simpler explanation for what we see with LLMs is that instead the higher level encodings in latent space match only the patterns of our language(s), and no deeper encoding/understanding is present.
It's Plato's Cave - the shadows on the wall are all an LLM ever sees, and somehow it is expected to derive the real reality behind them.
Could be, yes for sure but I think it would be very naive in the current state of progress we are in, to down play what progress is happening.
At least Mythos model with its 10 Trillion parameter might indicate that the scaling law is valid. Its a little bit unfortunate that we still don't know that much more about that model.
> And if you look at Boston Dynamics, Unitree and Generalist's progress on robotics
Their progress is almost nought. Humanoids are stupid creations that are not good at anything in the real world. I'll give it to the machine dogs, at least they can reach corners we cannot.
Not sure if you're being sincere or sarcastic but some of us have lived through several AI winters now. And the fact that such a phenomenon exists is because of this terrible amount of hype the topic gets whenever any progress is made.
Yeah and if you look at the blocking factors at that time (data, compute) these type of limits currently are non existend.
There is a difference to be acknowledged: in the 70s/80s the whole world didn't suddenly start to shift to AI right?
So why do so many smart and/or rich people push this? Hype? Yeah sure but hype was here for crypto too.
I bet its an undelying understanding and the right time with the right components: Massive capital for playing this game long enough to see through the required initial investment, internet for fast data sharing, massive compute for the amount of data and compute you need, real live business relevant results (it already disrupts jobs) etc.
I'm not an economist but at least as an european person, I currently do see a huge restructuring going on. A shift away from the USA to China. But I never voiced an opinion about that.
I do follow ML/AI/AGI though for a decade by now and read a lot about Neuronal networks, LLMs, etc. in a broad spectrum.
My prediction regarding Crypto/blockchain was true too.
We will see how it plays out. I'm open for both, but I think it would be naive to ignore whats going on and its way to soon to assume there is a AI winter coming soon.
We sitll want to see what Mythos can do and a distilled version of it.
Self-driving had never the amount of compute, research adoption and money than what the current overall AI has. Its not comparable.
Crypto was flawed from the beginning and lots of people didn't understood it properly. Not even that a blockchain can't secure a transaction from something outside of a blockchain.
> Self-driving had never the amount of compute, research adoption and money than what the current overall AI has.
And yet they don't do really good jobs with pretty much anything, save for software development, to which people still seem pretty split as far as it being a helpful thing. That's before we even factor in the cost.
I find them very helpful. I use gemini regularly for multiply things.
I also believe that whatever code researchers and other non software engineers wrote before coding agents, were similiar shitty but took them a lot longer to write.
Like do you know how many researchers need to do some data analysis and hack around code because they never learned programming? So so many. If they know how to verify their data (which they needed to know before already), a LLM helps them already.
There is also plenty of other code were perfection doesn't matter. Non SaaS software exists.
For security experts, we just saw whats happening. The curl inventor mentioned it online that the newest AI reports for Security issues are real and the amount of security gaps found are real and a lot of work.
Image generation is very good and you can see it today already everywere. From cheap restaurants using it, to invitations, whatsapp messages, social media, advertising.
I have a work collegue, who is in it for 6 years and he studied, he is so underqualified if you give me his salary as tokens today, i wouldn't think for a second to replace him.
I don't particularly care about coding and didn't weigh in on it. There is no dispute that people debate if it is effective at that. You can take that debate up with them, not me.
The problem here is the adoption curve; Right now it might feel to you that its not worth it or not happening as it might for most people.
Than suddenly one model update moves it from 80% to 85% and now 30% of the market wants to use it.
Then it might be already too late to act like using it to your advantage, being a valuable expert or deciding things long term based on the new state of affairs.
It’s an agreement between a public company and a highly scrutinized private company. Several of the provisions will change what happens in the marketplace, which everyone will see.
I imagine the thinking was that it’s better to just post it clearly than to have rumors and leaks and speculations that could hurt both companies (“should I risk using GCP for OpenAI models when it’s obviously against the MS / OpenAI agreement?”).