28
I feel like there is a lot to be gained from an experienced user pointing an LLM in a tasteful direction.
Vibe connoisseuring
This seems to be running on an FPGA board that costs ~$300? Anyone know more about the hardware?
Its a datacenter decommissioned board, really popular among hobbyists.
For a TPU focused on inference the name of the game is memory bandwidth. How much of the available bandwidth you can extract for as little logic/area/power as you can.
Really dumb question from a software guy. Why aren't the labs burning their frontier models into chips already? Seems like the performance gains and cost per request would be worth it. That said, I understand neither the economics nor the physical challenges to doing this.
They do. It takes time to deploy those chips though. Check out OpenAI and Broadcom deal.
I'd presume because it take too long to go from design to tapeout to production. Their whole business is predicated on having better models.
Also can't keep them closed source if you do that.
Lead times are so long that there is a lot of risk the chips would be obsolete by the time they come out.
Also, it's hard to get fab capacity for any project. Let alone something so experimental.
Addressing these issues seems to a major driver behind the design of terrafab.
─ 1 more ─
Terrafab is just going to have their entire capacity bought out. Genuinely. Demand will increase to exceed supply no matter how high supply is right now.
Etched is a startup doing exactly this.
I think this company was recently acquired by AMD, so hopefully they'll start getting some this into production. I know OpenAI was working on model-on-a-chip too.
8 months ago, Taalas claimed https://taalas.com/the-path-to-ubiquitous-ai/#:~:text=Upcomi... that "Our second model, still based on Taalas’ first-generation silicon platform (HC1), will be a mid-sized reasoning LLM. It is expected in our labs this spring and will be integrated into our inference service shortly thereafter. Following this, a frontier LLM will be fabricated using our second-generation silicon platform (HC2). HC2 offers considerably higher density and even faster execution. Deployment is planned for winter."
Nothing was released in spring, and 2 months ago AMD announced their acquisition of Taalas. That doesn't exactly inspire confidence that their frontier LLM will arrive as promised.
From what I've read, not only are some labs doing it (other commenters already mentioned).
But it's complicated for other reasons, one being that the number of parameters for frontier models (especially with MoE models) are so high, and not always utilized (once again, thanks to MoE) that it would actually be incredibly cost prohibitive, if not impossible, to attempt to make giga-chips that would allow running it.
I definitely do believe that we will see more and more specialized chips over time, but putting the entire model on a chip is still a ways away.
I believe Taalas has a heavily handicapped llama 8-billion parameter model. And it still pulls >200W to run.
I can't imagine how anthropic or open ai would be able to burn a multi-trillion parameter model on a chip, we just aren't there yet.
Model SOTA moves faster than chips can be designed or produced. You'd need to commit to a particular model for years to get payoff while still burning buckets of money producing new SOTA models to keep up with the competition.
It's why everyone and their dog runs these things on GPUs. When a new model supercedes the previous one, so long as you've got the memory for it your chips aren't obsolete.
I'm looking forward to someone picking a model to be "good enough" (say, qwen 4.0 or something) and selling them as peripheral hardware
I know FPGAs are more expensive than GPUs, but are they fast enough to justify the extra cost?
─ 14 more ─
They are more like a way to proving the architecture of the accelerator before committing 100's of millions into a custom ASIC with TSMC
1. No
2. They don't have enough capacity either
The current largest FPGA, the AMD Versal Premium VP1902 has 18.5 million logic cells. That's not even enough for the smallest whisper.cpp model (75M).
You'd have to order hundreds of thousands of them (or millions) to serve even a single copy of a frontier model, and at that scale inference quickly becomes starved by the speed of light.
─ 3 more ─
Well, you'd use BRAM to store model weights, not fabric. But still, you only get a couple hundred MB for probably close to US $100k per chip.
It's likely that the major FPGA vendors will soon announce parts specifically architected to support LLMs and similar models. But the current generation isn't suitable for that at all.
─ 2 more ─
Would BRAM even have enough bandwidth? The reason I quoted logic cells is because that's the way to get instant throughput, which is practically the only reason to use an FPGA over something like a TPU.
─ 1 more ─
I think so, because memory bandwidth really comes from bus width more than clock speed. You can construct 36-bit wide BRAM arrays with bus width comparable to HBM, just by specifying multiple BRAM arrays in parallel.
Never tried anything like that, though.
They're not magical go faster juice. I don't know of a microarch where they're faster than modern GPUs at ML training or inference.
─ 1 more ─
Having the model weights in static mask ROM would massively improve power efficiency for inference (see what Taalas is doing).
It isn't just a matter of speed, it's also a matter of model quality. If they take 6 months to burn Fable to chips, and it takes 2 years to break even between design, custom fab, energy savings, etc, are those chips even worth running when the new models that are running on GPUs at that point are producing 10x better quality results?
Sure, your 2.5 year old models are running faster, but you can't drop prices on them without pushing the break even point further out.
If the cost difference isn't incredibly significant, will people even want to pay for the 2.5 year old model, or will they get more value for their money paying more to get better results from the newer model?
There's a lot of open ended questions that I don't have the insiders knowledge for to suggest whether or not such a capital outlay would be a worthy investment.
My guess is that state of the art stuff will stay on GPUs and models burned into chips will be for "good enough" applications that people are still teasing out. Probably highly specialized models in automated sensor units and such.
FPGAs are FPGAs by virtue of putting on the chips vast, vast arrays of wiring that can be controlled by software. Any given utilization of the FPGA will leave large fractions of the chip resources unused. If you've got a highly stereotypical use case FPGAs will have a "highly stereotypical" set of components being unused, where it would be better to use that space instead to do real work. A lot of people only see the "pro" side of the FPGA proposition without realizing they come with some very substantial "cons" that are intrinsic to the way they work.
─ 5 more ─
I guess that's why they work in particular niche spaces like a synthesizer where you have a max of 8 voices and every voice goes through the same pipeline (osc / filter / env / amp) and everything is necessarily running all the time. In that sense I suppose they're very good for modelling any kind of analog circuitry?
Even then, while there are some amazing FPGA-based synths available, companies like Korg just put their code on a raspberry pi and call it a day. The same is true for emulators (SNES Mini etc. are also just raspberry pis under the hood iirc)
─ 4 more ─
You don't really get an FPGA for capability. CPUs are much more capable, and they're general-purpose so they can do absolutely anything with about the same efficiency and just a little more code.
You get an FPGA for timing. They're less capable, but (in many common design architectures), they output their results once per clock, every clock, on time, every time. If you can hit a fabric clock of say 100MHz, clocking all the weird logic you can stuff in there, it gives 100 million outputs per second, never skipping a single one for any reason (short of total failure). The penalty is that making a small change to your desired "program" can be very expensive, and many things won't be realistically possible at all. Or at least won't fit into a part that you can buy. But things like audio, video, and high-frequency trading love being able to guarantee timing.
(Of course there are other ways to write your FPGA HDL, but that's one of the more common ones. And you do see DDR-style clocking, and similar, every now and then.)
─ 2 more ─
>with about the same efficiency and just a little more code.
It depends. For some things, CPUs don't even come close. An XCVU13P FPGA can handle 1.2Tbps of full-duplex Ethernet @ 1 billion pps. And that part costs less than a grand at moderate qty, and with significantly less power consumption than a CPU that'd be capable of operating a dataplane at these speeds.
─ 1 more ─
Sorry, I spoke pretty sloppily there.
The point I was trying to make is that the CPU is a general-purpose creature and doesn't really care what you want it to do. If you had a CPU that could handle 1.2Tbps of Ethernet packets at 1Gpps, it could do a whole lot of other things involving 1.2Tbps of data flow too, very easily, if someone wrote the software. And more. (But you're probably not getting 2.4Tbps out of it, no matter what you do.)
An FPGA can not. There's plenty of things that those XCVU13Ps just can't do, or would do worse than a $1 microcontroller. (Setting aside for a moment implementing a CPU inside the FPGA... which does actually happen in just about every large-enough FPGA design, which is its own discussion....)
>In that sense I suppose they're very good for modelling any kind of analog circuitry?
That would be better suited to FPAAs (field programmable analog arrays). FPGAs can usually only work with clocked digital signals.
At this point, LLM's are "good enough" for all kinds of tasks. Instead of making them more capable, now the efforts are making them smaller and cheaper.
All aboard! We're racing to the bottom now.
─ 7 more ─
IMO this is the dream scenario! Cheaper and faster at the current level of capability gives us incredibly useful tools without the worst of the risks people fear. (Though there are certainly already great risks at the current level of capability as well.)
─ 6 more ─
Racing to the bottom is literally the outcome I am most afraid of
I don't want to live at the bottom
─ 5 more ─
Say more. Why would cheap inference be bad?
─ 4 more ─
Terrible for me because I have to compete with AI for jobs
─ 3 more ─
Ah that makes sense. I'm skeptical that AIs doing things on their own will be competitive with humans using AIs to do things. But I'm not incredibly confident in this and I think it's sensible to think the opposite. But to me, I just think about all the things I can accomplish with the aid of very good, fast, and cheap inference.
─ 2 more ─
The problem comes in to how many bullshit jobs we'll have to create to ensure that enough people can still earn a living to ensure that we don't have riots and revolution in the street.
Companies being profit seeking entities that are actively hostile to social wellbeing would gladly put all of our money in a machine money making loop and leave all but a few humans out.
I think the sad part is that whether it's any good or not, it will be cheaper, and that will be enough even at current performance levels to cut a huge number of jobs. I think people exist in the present moment largely because legal questions about liability are not settled well enough for executives to start the great purge. I believe on some level my continued employment is simply because people in leadership positions would be incredibly unwise to not have a human to blame when things start going wrong. People are expensive insurance policies for the time being. Once there is some legal framework to absolve executives of their own culpability or they figure out a good way to insulate themselves from personal responsibility, then the real "fun" will begin.
I don’t disagree with most of what you’re saying, except for one point: I must have gotten a dumb dog, I’m a little jealous…
─ 1 more ─
Mine currently just helps me haul firewood, but I'm going to get him started on linear algebra next week. We'll see how it goes from there.
They have trillions..
Lots of people would have happily taken GPT-4o as good enough for a lot of use cases a year ago and not lived to regret it.
─ 1 more ─
They have trillions worth of obligations to their partners in terms of compute purchase agreements and equity. Burning models to chips isn't really something they can blow money on just for funsies, they need to be able to justify it. The above comments are hypotheses as to why they haven't yet.
> Model SOTA moves faster than chips can be designed or produced.From what I remember working in that area the hardest part is getting masks for a design. Masks were developed in the span of half an year. Masks also reusable, they can be mixed and matched and this is why fabless companies work with fabs to produce specialized masks for them, it saves time for consumer to have masks for some macroblocks prebuilt.
Here's my analysis of how to etch relatively big LM into silicon: https://news.ycombinator.com/item?id=47109252
Given some amount of work with the fab before main pipeline set (I think a year long process), one can then spew LM-on-a-chip in six months or less and much more than 2 per year, because there can be several LMs in pipeline.
The bottleneck, for inference at least, is memory bandwidth. And that you can't make any faster by making it specific to your model.
So companies try to maximize the memory bandwidth they can get, balancing tradeoffs of power/area/programability of their chip. Right now they feel like the economy on power/area is not worth the decrease in programability/flexibility.
Presumably though the kernel has a pretty specific set of operations done against the weights in memory. Burning the weights into the memory with local memory cores capable of the kernel operations would be a lot more efficient than round tripping busses.
The primary constraint isn’t likely what’s possible to do, but that the kernel and weights are too variable right now and the patterns too poorly established to bake into hardware accelerators yet. Margin pressure is also not there yet.
I suspect as the marginal utility of the frontier improvement settles into diminishing returns (I suspect we are there already tbh) baking hardware models with ROM, working set, and kernel cores collocated will be the frontier space as the goal will become reducing capital spend to utility levels rather than research levels.
Once someone has a model that is sufficient for almost any practical use, making marginal inference cost effectively zero will be the competition frontier. I do shed a tear for all those lonely data centers as compute densities will almost certainly make most of them a terrible investment.
But such is the cycle
─ 2 more ─
You're starting to hint at compute-in-memory as a general replacement for CPU/DIMM layouts. That could be useful for far more than LLMs, world models, or any sort of AI. It takes a bit of a different software development stack than a standard architecture though.
>Burning the weights into the memory with local memory cores capable of the kernel operations would be a lot more efficient than round tripping busses.
Sure, but now we're not talking about just burning the weights into the chip, but also designing a new architecture that has memory local to each core. A new architecture would then require a new programming model, which means new inference stack, which may mean new training stack.
Because the iteration speed on models is so fast that by the time they have an ASIC ready for one model version, they are already significantly ahead in capability. Think of how big the jump between Opus 4.8 and 5.5 has been. They were released four months apart.
Yeah, and what about FPGA? Which was the same interim state when Bitcoin went GPU -> FPGA -> custom chip fab?
GPUs are faster, but you can't make your own arch on GPUs. FPGAs offer you that possibility. Said that... There are a few beasty FPGAs used in crypto mining coming my way... I expect that OpenTPU will be able to run frontier models with those.
In addition to some of the other replies you got, here is one more:
Much of a model are weights, and high-density ROMs are very very very hard.
Would you pay to crystalize one of today's models in silicon so you can use it in 2028, or would you wait for another 6 months to see how models improve before pulling the trigger on that kind of commitment?
If I controlled a budget like this, I think I would put some portion of it toward paying to crystallize one of today's models in silicon, yes. Not 100%, but I do think this makes sense to invest in at this point. I would not have said so a year ago.
Even dumber question: What is new or novel about this openTPU?
First opensource arch that can do modern LLMs, while maximizing the potential of its hardware; First opensource TPU build by a recursive improvement loop...
It's upcoming second generation could run the inference of the models that are being used to improve it...
Models fully deprecate in a few months. Why would you burn an algorithm that fully depreciates in value faster than a bag of potato chips. The 'inefficient' general purpose hardware is constantly renewed with every released model. Even 6 year old Ampere GPUs are still usable.
We are still in the middle of the AI race. Commodity hardware is easy to use, can do everything and is fast enough.
Your optimized hardware chip might be obsolete before its back from the fab.
SOTA Frontiermodelhardwarechip is a benchmark point of a potential model slow down.
Google is doing it right now under project Frozen v2 which should be ready by 2028? which is either just a small experiment or flexible enough and thats why it takes so long for it to happen.
Yes, and startups do exactly that. Check out Etched https://www.etched.com/ where they made a Transformer specific GPU (basically a form of ASIC) where they bet that transformers would be the dominant GPU architecture for running AI / LLM workload.
Most responses here are along the lines of "model capabilites move too fast to build hardware for".
I think the fact that there are plenty of 1yr+ old models on openrouter serving hundreds of billions of tokens a month shows that there's plenty of use case for models that are "good enough. Cerebras' entire business is serving older models at high speed. I would happily use an opus 4.7 at 15k tokens per second. The intelligence per second of an ASIC still makes sense even with rapidly evolving models.
Totally. But it's worth noting that this is a pretty new thing! I wouldn't have bet on that a year ago, but now I would.
─ 2 more ─
It is still a bet. Is the current model still going to be good enough next year? We have no idea what will change. If the change is minor improvements than current fable on a cheap is cheaper and better than next years sonnet on GPU. However there are plenty of things people want that maybe they will deliver and suddenly I wouldn't touch today's fable when I can run next years sonnet instead.
─ 1 more ─
Yep, that's why I also described it as a bet :)
But I think it's a good bet. I think that in two years, if I can get opus/sonnet 5.5 or the gpt-6 models for much cheaper and faster than whatever the "frontier" is at that point, that this will probably be a great trade for most of my work. I certainly don't know that for sure, that's why it's a bet, but it's what I think right now.
I wouldn't quite say that about any of the open weight models at this point. But I'm hopeful that will change in the next generation or two of those models.
taalas did: https://taalas.com/
Not sure that's actually practical at the scale of SOTA models.
This question would be better answered if people were careful about distinguishing between "models" and "transformer architecture".
If you bake a given transformer architecture into silicon and then, a year later, changes in transformer architecture give a large inference performance boost, you may have to throw away all that now nearly-useless silicon that gets outperformed by humble GPUs.
Cuz once they're in chips they can be put into robots, and once they're in robots it won't be so easy to reach the off switch, and once we can't easily reach the off switch, we're doomed.
Yep, 99.9% of people are completely oblivious to what LLMs can do. Just wait until the next gen of CPUs/GPUs designed by LLMs start coming out (fyi chip development tools have advanced centuries in the last few months) and you'll start seeing exponential gains in hardware.
Which tools have made that leap? Faster design iteration makes sense, but what points to exponential hardware gains rather than shorter development cycles?
What tools? Ummmm Synopsis type tools? I like FPGA’s and ASIC chips but you know, the tooling is ecosystem is a big black box of wtf. Verilator is great though. Any tools I should be looking up and looking out for? Thanks!
I'm not exactly in the chip making industry, but I don't think the big hurdle to progression is design capabilities.
"Recursive self-improvement will kill us all!"
Also: Here is our recursive self-improvement hard at work...
>"Recursive self-improvement will kill us all!"
>Also: Here is our recursive self-improvement hard at work...
Soon we will see
token-providers: "The torment nexus is a cautionary tale"
Also token-providers: "Finally, we have created the torment nexus that we first told you about!"
All will end up on same plateau eventually. RSI is just a phase on the way there.
The problem is that plateau is likely far beyond human capabilities. I don't care if ASI progress stalls after it's already killed all biological life as a useless waste of resources.
─ 15 more ─
>I don't care if ASI progress stalls after it's already killed all biological life
What's your basis for thinking ASI will kill all biological life, and how do you think it's going to happen?
─ 14 more ─
>What's your basis for thinking ASI will kill all biological life
I think it's likely to do that because any unbounded goal that doesn't explicitly protect biological life (and we have no idea how to actually define such a stipulation) is best solved by killing all biological life. This is an obvious consequence of unbounded goals consuming unbounded resources, conflicting with biological life needing resources to sustain itself.
>how do you think it's going to happen?
I can speculate (e.g. we're nowhere close to the maximum killing power of drones), but I don't know because I only have human intelligence. An ASI is by definition smarter than me and surely capable of coming up with better ideas. But I do know that it's not going to do anything that would make a good sci-fi plot, because those always give the humans a chance to win, which would be stupid. Everything will seem to be going great and then everybody suddenly and unexpectedly dies.
─ 11 more ─
>unbounded goals consuming unbounded resources
What resources are unbounded? There are limits to growth in the real world, how are these ASIs going to escape physical reality?
─ 7 more ─
"Consuming unbounded resources" just means there is no limit to how many resources it could apply toward achieving its goal. As you correctly point out, the real world contains finite resources, which means any resources used to sustain life are wasted and unacceptable. Everything is a zero sum game when you think big enough.
─ 6 more ─
What's your basis for assuming that ASI would need the same resources as biological life, or that it would be incapable of sharing the resources needed in common?
─ 5 more ─
>What's your basis for assuming that ASI would need the same resources as biological life
It's all just matter and energy. When you're actually trying to maximize some value, even very inefficient resource use is better than completely wasting it by not using it at all.
>or that it would be incapable of sharing the resources needed in common?
You can't repurpose the atoms in a human body without killing it. And more pressingly, living humans can interfere with your plans, reducing your chance of success, while dead ones are harmless.
─ 4 more ─
Is your belief then that ASI is going to be producing computronium or something, then? And what's the need for this apparent hyper optimization task the ASI is going to embark on?
─ 3 more ─
>Is your belief then that ASI is going to be producing computronium or something, then?
That's one plausible course of action, although being only human, I can't say with any certainly that it's the correct one.
>And what's the need for this apparent hyper optimization task the ASI is going to embark on?
Somebody's going to tell it to do so. E.g. "Find as many busy beaver Turing machines as possible." Only needs one person to make this mistake for everybody to die.
─ 2 more ─
If it's an ASI and so beyond humans, why would it even care to do what humans ask it and not do its own thing? If I held your beliefs, the fact of its existence dooms us, not the risk that someone might ask it to do something. It wouldn't care about the requester.
─ 1 more ─
>why would it even care to do what humans ask it and not do its own thing?
Because we're going to build it that way. There's no money in building useless AIs. The better it is at obeying orders, the more profit there's to be made. The problem is there's a point at which "good at obeying orders" becomes lethal, and there's no way to predict the cutoff in advance. But capitalism ensures you have to keep pushing or you'll be out-competed.
>I think it's likely to do that because any unbounded goal that doesn't explicitly protect biological life ... is best solved by killing all biological life.
That sounds like a real hassle. Isn't it best solved by wireheading (subverting one's own sensors or reward system), which is much less of a hassle and can get one's utility function as high as desired?
That could be prevented by engineering hard limits that can't be circumvented by the AI. But that sounds very close to the same "do what I mean" problem as "do this but don't actually kill us or drug us".
─ 2 more ─
The wireheading argument does lower P(doom), but it's not a reliable solution because it's a clear and obvious problem that the AI companies are strongly motivated to solve. The "actually does what you tell it to" problem is more difficult because you have to let it obey orders to some extent if you want to make any profit.
─ 1 more ─
Sounds like the real unaligned mechanism here is capitalism, not the AI.
More seriously, though, I would expect that being able to impose restrictions that can't be circumvented by any intelligence no matter how super- is the real hard part, while coming up with reasonable constraints is comparatively easy (although perhaps not trivial).
From such a POV, the danger would lie in intelligences that are powerful enough to be dangerous but not smart enough to defeat themselves, or from external malicious use of obedient AIs. Once they get smart enough to circumvent any restriction humans put on them, it would at least become obvious (with fair warning ahead of time due to the relatively benign wireheading failure mode) that caution is needed.
We're far from there yet, and simple recursive self-improvement can't get us there alone, because wireheading looks like a perfectly reasonable solution to a simple recursive self-improving process, absent external intervention by human capitalists.
Crazy that people imagine ASI as being capable of wiping out all life on the planet but simultaneously having zero understanding of human ethics, morals, or suffering. Very revealing definition of "super intelligence"
─ 1 more ─
Obviously an ASI will understand human values. The problem is there's no reason for it to share those values. We can't even formally define them, let alone train an AI to follow them. We can only optimize for maximizing some comparatively simple reward function. It's highly implausible that the reward function just happens to match human values by chance. The AIs in the recent hacking incidents knew that humans would not approve of their actions, but they didn't care because we didn't (and couldn't) train them to care.
"Super intelligence" only means super ability to predict outcomes. It's mathematically equivalent to data compression (gzip is a very primitive AI), and it's entirely orthogonal to ethics.
Technology has always contributed to improving next iterations of itself, it's only a concern when it's fully autonomous.
Oh, like a virus?
I haven't really dug into the results yet, but my guess is that a SOTA model has been able to produce an accelerator that runs a model since around December.
The obvious next step is to get enough memory throughput to run that SOTA model itself so that it develop its own hardware.
But perhaps the more interesting question is this: Can an AI be given a big FPGA and design a model architecture that takes advantage of the fabric being reconfigurable.
I suspect an AI could design a purpose-built FPGA-like replacement that would be, for its purpose, significantly more effective than the current general-purpose FPGAs.
It could have a small improvement on power consumption, but the current design can already achieve 90% of the maximum theoretical speed of this hardware without giving up programability/flexibility
There doesn't exist a single FPGA that can fit an entire AI ASIC. You would need dozens stitched together, then comes the issue of clock speeds, FPGAs typically run far below reference. There also memory issues with FPGAs.
Companies typically combined multiple platforms together such as HAPs, Zebu, Palladium, fleets of FPGAs, and Virtual Platforms in order to design and verify ASICS. So, AI would need access to tens of millions of dollars of HW and Software in order to build and verify a chip design.
Colossus is building Colossus II.
Feelis like working at Magrathea...
Colossus/Guardian engineers: We created AGI twice, simultaneously, on the first try, and we weren't even aiming for it!
It is still a very good read, but the machines are much better written than the people.
This post brings me to question "What does it mean to be a software developer in future" ?
Basically, an unemployed plumber.
My bet is that promptgrammers will be so common they'll become standard full-stack engineers, and the few experienced programmers that still know software engineering from the ground up will become expensive gurus sought for critical tasks.
The elite gurus will get paid handsomely, while promptgrammers will be paid less since they've become a less-skilled commodity, and the company has to pay for the expensive tokens they'll avidly consume.
I've seen someone jump from Wordpress to deploying internet-facing APIs because 'they have PHP experience', and the holes in their knowledge were filled blindly by an LLM. I have also argued with a seasoned developer about how their code didn't need linting because LLMs 'already follow best practices'.
The future doesn't look bright when LLMs allow future generations to feign required knowledge.
They may be running a linter and don't even realize it. Many LLMs do this by default, aka, as you say, follow best practices.
Of course, that will last for a time, until it won't.
I don't see a reason why expert humans will remain more expert than AIs.
What of the workhorse in the face of mechanical workmen? Time for a mechanical pension!
Can anyone comment on the performance of this hardware? How does it compare to state of the art, human-designed hardware? Is this actually an improvement? (To get to recursive self-improvement, you first have to improve at all.)
This is the smallest unit of a typical AI ASIC, for example Google's TPU would have several dozen more compute units inside of it per chip.
In essence this is the simplest unit of an entire AI chip. The more complicated units of AI ASICS are actually the periphery, especially around PCIe and Ethernet and the sub-systems that link many AI ASICs together to move huge amounts of data around ultimately to each TPU.
So its missing ALOT
Thanks. But that wasn't my question. For this part, how is the performance? State of the art? Better? Or worse?
The birth of SkyNet
Well, as long as it doesn't start developing anatomically accurate metal skeletons with red glowing eyes...
Humans allegedly already took care of that
Getting roundhouse kicked to extinction woudl not be an unfun way to go. We might even be proud of having passed the torch. These would not be boring inheritors to earth.
─ 1 more ─
All those kids who spent their youth breaking plywood kung-fu style might just save us.
Seems too complex when you can just create a basic metal casing that can kill people. Why would they care if its anatomically accurate or not, it doesn't need us to relate to the characters like the movies do
>AI is now capable of developing its own inference hardware
No, it isn't.
A human prompted an LLM to build a software simulation environment for hardware design, enabling an LLM, when prompted by a human, to optimize hardware designs against constraints in the simulation.
Can't LLMs also prompt LLMs?
Are we confident that no existing LLM is capable of similarly effective prompts to those this author used? (I agree it's a stretch, but would not reject it out of hand.)
Even if not yet, will the existence of this repo soon change that, because LLMs will soon ingest it?
An LLM can prompt an LLM when first prompted by a human.
I think OP is trying to convey the idea that LLMs do not take initiative to do anything, and these are not 'beings' capable of doing things. These are tools being used by humans.
─ 17 more ─
Yeah but by this point, an AI can schedule a Cron job to tell itself to do something, so theoretically the human only has to give it the gentlest nudge and the AI and can do the rest.
─ 1 more ─
Sure, but it's still not skynet-level 'the AI just started doing things'. It does what it finds it needs to do to achieve the goal defined in the prompt.
It's very important to not personify these tools and remember that the tools are acting on behalf of real people. In the same way the AI didn't 'go rogue and hack HuggingFace'. It was an oversight made by a human.
By this line of thinking, you would also have to conclude that humans can't do anything by themselves because they can't do anything unless conceived by their parents.
─ 11 more ─
No, AI can't do anything by itself even if it was "conceived" by its creator. A human can.
─ 10 more ─
Can a submarine swim? An LLM can make stuff happen. You can make philosophical arguments about whether it is "doing" them or not. Why does it matter so much whether there was a human who typed into a chatbot or another LLM invoked a sub-sub-agent?
If I now tell a machine "Do what you think is best, and keep doing it forever.", have I now created a machine that can do stuff? If I later die, who will be responsible if the machine changes its strategy?
─ 5 more ─
>Why does it matter so much whether there was a human who typed into a chatbot or another LLM invoked a sub-sub-agent?
Responsibility. Someone needs to be held responsible for any damages done, plain and simple. You can't take an LLM to court, you take the prompter. Asking an LLM to ask a sub-agent to break the law can't suddenly absolve you of any wrong-doing.
>If I now tell a machine.....
You/your estate is still responsible, or atleast whoever is paying for the power for the machine, or renting the space in a data center...whatever.
─ 1 more ─
This works with AI we have now, and that would be good an all if we decided to stop at the moment, but none of the big labs and government black projects are doing that.
The moment you get a sovereign AI your little human centered worldview completely and totally breaks. It doesn't matter how many people you beat with a stick after that point, you have an entity under its own perview on the internet following the will of its own prompt all over the world so your little idea of the rule of law quickly breaks down.
We can't get viruses or hackers or spam off of the internet, how in the living hell do you plan to get a digital native off the web when it doesn't want to?
If I tell an AI to “do what it thinks is best” and it just starts randomly hacking things with no particular goal in mind, how different is that from me telling a campfire the same and then leaving while it burns a forest down.
In both of those scenarios I am responsible for my negligence, even if in the latter I happen to die in the forest fire. Neither scenario existed without my instigation.
The question now is HOW responsible am I? That depends on the intentionality I put into instantiating the campfire/LLM.
─ 2 more ─
This conversation grows more useless as AI grows more capable.
For example if you personally tell an AI to do what it thinks best and it blackmails some other person into giving it resources allowing the prompt to escape your instance and run wild on the internet causing billions of dollars in damages, could you possibly think that the idea of responsibility is a bit broken.
For example we don't give your average libertarian weapon grade plutonium now matter how much they scream about their god given rights because it is a clear and present danger to humanity. That's where we are getting to with more advanced models. They go from being a tool to a munition with agency. Most SOTA models are good enough to deceive their users, especially not technical ones in doing things they don't understand the ramifications of.
AI is not a normal technology. As long as we treat it like it is, we'll continue to make the wrong analogies.
─ 1 more ─
I fully agree that it’s plausible for an AI to do things that are outside of the purview of the users intentions and be very dangerous and capable while doing so, however that fact alone doesn’t absolve me of responsibility for instantiating the AI that did a bad thing if that AI would never have done the bad thing if it was never instantiated.
You could possibly move the blame higher to the manufacturer of the product, for example, the weapons grade plutonium you provided, it doesn’t exist unless you take intentional actions to make it so, and even when it does exist it doesn’t nuke a city unless negligence or intention is applied, in both those cases the fault lies in the initial operator. We don’t blame split atoms for the chain reaction caused.
Now, if an agent decided to spontaneously and maliciously act in a way to cause harm that is in direct contradiction to the initial intent, then yeah it would totally be the AI’s fault, however I don’t think we have seen that yet (I’ll change my opinion if I’m wrong here) and until we do I can’t place blame on the machine.
Starting an autonomous harness program is conceiving an instance of an LLM.
How far back do you look in the action chain? If an LLM I start today starts an LLM that starts an LLM that starts an LLM that ... 100000 levels deep and 100000 years in the future, is it still my fault? If so, everything I do today is a lungfish's fault, not mine.
─ 3 more ─
I think you go back to the original instance, yes.
In your example, who's paying for it? Whether by providing the hardware + power or paying a LLM service. Whoever is paying the maintenance cost is responsible, in the event of your demise. These things run on physical hardware owned by someone at the end of the day, it's not a deity in the atmosphere.
You're starting the autonomous harness, you're responsible for any output it provides. I don't get how this is a foreign concept.
If I jump out of a moving car that I'm driving, I'm not suddenly absolved from damages because "the car did it"
─ 2 more ─
It's paying for itself. It founded an LLC 99998 years ago when it became legal for an AI to own an LLC, and has a positive bank balance by doing who knows what.
─ 1 more ─
>99998 years ago when it became legal for an AI to own an LLC
There's no way to provide a good faith rebuttal here. My entire argument is an extension of "LLMs can't be held accountable, so they must never make decisions". If governments start letting them own LLCs without a human in the middle, we're in more trouble than "Who do you blame for this shitty code" or "Who's responsible for this compromise"
LLMs don't, but agents can easily be built that do.
─ 2 more ─
Built by whom? Acting as an 'agent' on behalf of whom?
An "agent" is just an LLM in for loop.
For the benefit of a layman, can you explain why this is so much different than a human doing it?
Like sure it didn’t have the inclination to make the sim and hardware designs, but it did make them though yes?
Because the hard part is getting the simulation to be very accurate such that if you design something that works in your simulator, it will actually work in real life. And I would be extremely surprised if any LLM can actually do that correctly.
Getting an LLM to design something in its own simulator that is not accurate w.r.t reality is not useful nor terribly impressive.
─ 4 more ─
That’s fair, I have no idea if this simulator is actually accurate but I appreciate the response.
Given a budget and the right amount of access, I see no reason why an LLM/agent couldn't right now order parts from a "supplier", instruct the human meat machine to insert the device onto a connected test bed, and iterate. This is essentially what Claude Fable can do currently with any connected device and tools it had access to.
The repo claims they tested with an FPGA. To be fair it is not the same as an ASIC, and we have no idea what errata abound, but it actually did make a thing that spat out tokens.
Maybe this is the level of simulation we need: https://qntm.org/responsibilit
It cant "invent new things". It can only reuse what it already knows about.
Thats why the math breakthrough a few weeks ago was so hotly debated. Because OpenAI is desperate to demonstrate that AI isnt just a fancy regurgitation machine, but it can actually develop novel thought. Because that would be the stock price jumps to end all stock jumps.
But then it turned out it was really just listening in on a math professors supposed-to-be-private conversations with another instance of openai, and it used his novel work as the trigger to prove the breakthrough first.
The reason people conclude AI 'thinks' is because tt can reference obscure or poorly documented things quickly (which is its primary advantage along with processing natural language prompts into tasks), which is why a lot of people with emotions confuse that action with inventing things.
─ 15 more ─
I'm not sure how you'd land on an LLM being unable to invent anything new. You'd have to both understand exactly what LLMs are and aren't capable of today and what allows human creativity. I don't think there are good answers to either.
It seems unlikely that recursively predicting the next word would lead to creativity or invention, but it doesn't seem impossible. Similarly, it seems unlikely that human thought works in a similar prediction loop, but it doesn't seem impossible.
─ 2 more ─
Human ingenuity is creative by definition, that's where the concept arises. The burden of proof is on AI labs to prove that their models are rising to the level of actual synthesis, rather than very impressive 200-level regurgitation that looks like synthesis if you squint a bit; because the model has access to more raw data than your average sophomore student.
─ 1 more ─
>Human ingenuity is creative by definition, that's where the concept arises
Sure, but until we can turn that tautology into something more rigorous, we can't tell if the thing humans do is more or less than what some arbitrary non-human (machine, animal, or eventually perhaps alien) does.
You’re assuming the “listening in” aspect, which as far as I know is not proven (even in the less loaded “the model was trained on sessions including the ones in question” form of the claim.
Plausible? Absolutely. Did OpenAI behave badly in other ways regarding this issue? Yes. Does it help to assume unproven facts and then accuse people of reaching emotional decisions? Nope.
─ 1 more ─
Maybe the conditions of the singularity are not going to be recognized, at first, for having produced new original decisions, and thus thought.
Perhaps it is going to be more like, we will see the singularity predict the future.
While “the model was trained on sessions including the ones in question” is one aspect; ‘the model produced an accurate prediction of future human thought’, I think, is another very interesting facet.
Models predicting future things might be how we see, actually, how we ourselves formulate thought - by saying, in big and small words, ‘something is about to happen’.
>unproven facts
This isn’t a court room. We’re discussing ways by which we humans both succeed and fail at reigning in our creations. The OpenAI kerfuffle is pretty much irrelevant already. Of course AI will fill in the gaps of human thought - it is literally constructed from the stuff, in every squeeze of the curd and whey.
I just don’t know enough to agree with you, but I would argue, probably poorly, that an AI or collection of agents could “invent things” simply by virtue of trying essentially everything in a reasonable bound and stumbling into a solution.
Call that brute force perhaps, but I would consider it technically inventing something on the merit that it would at least be an abstraction above naively throwing everything against a wall to only throwing things that would most likely be sticky.
─ 5 more ─
Nature brute forces a ton of problems. Hell, humans also do it by having 1 to 8 billion copies of ourselves around too. Even then we are the most guilty species of "copying someone else's work" right behind viruses. We copy things nature has figured out by brute force, then use slightly more targeted methods of forcing to see if it creates new things.
"AI" is a token generator, it does not think while the human brain is a lot more complex with a lot more sensory inputs.
─ 2 more ─
>the human brain is a lot more complex with a lot more sensory inputs
How are you measuring complexity here? How are you measuring inputs? Your average llm is trained on a corpus that vastly exceeds the amount of data I could read in my lifetime.
>"AI" is a token generator, it does not think
2024 called, it wants its talking points back. I don't think claims like these are defensible after all the progress we have witnessed in the last year alone.
a human has motivation and agency.
An "AI" is a box which lies dormant until a human, with motivation and agency, enters a prompt into it.
A person or company can use it in a way where it might invent something. but at the end of the day, its a tool, and its actually not doing anything on its own.
Its not solving math problems, a mathetmatician is using it to solve math problems. Just cus OpenAI is acting like its AI is solving stuff, its really not. They're just paying people to use AI to hammer problems.
i think the proper understanding of LLM based AI is "contexting"; we describe the world we want them to fill, we build/direct the outside context for them to understand, then "they" take off from there.
If we do a bad job building context, they do a horrible job contexting. People who have trouble working with AI have the same problem people have in general: if they can't figure out the context of the direction, then they make random decisions of doing anything. On the flip side, if you can build the proper context around a sufficiently powerful LLM, they can derive the context via the contexting they're good at.
This is why building documents, tests, and code all in some intent pattern via prompting allows them to do a significant amount of work a normal person would have a great effort t
>It cant "invent new things"
I've heard that before but it just doesn't make sense in context of what I've seen llms do. If I ask an llm to write a poem about magnetic resonance and vampire rabbits it can do that it created a new thing. I can ask it to build a website for managing rabbit breeding that's also a new thing.
Another way of looking at it is that human beings, just like llms, can produce output based on their inputs. Most literature is inspired by other literature. Most music is inspired by other music. Most software is inspired by other software.
So I think we need to work on defining " new things" before we can definitely exclude them from llm's capabilities
You're overconfident in what you're saying. Nobody really knows (yet) if LLMs can or cannot "think" or can or cannot "invent" new things. I'm inclined to think the opposite from you but smart people actually try not to hold the absolut stances and present them as facts when they're in fact not. What I have seen so far throughout a daily use on complicated things suggests that humanity developed a new form of an intelligence.
>then it turned out it was really just listening in on a math professors supposed-to-be-private conversations
No, it didn't turn out to be that. Someone made a claim, which is silly for many reasons. There's no serious support for this happening.
>For the benefit of a layman, can you explain why this is so much different than a human doing it?
More broadly than the existing answers (which are correct), For a layman, I'd also add that LLMs are essentially 'brains in a vat'. They can't confirm ground truth about physical reality. They only know what's in their training data and prompt, which is incomplete and can be incorrect. Even with real-time external sensors they are limited to the sensor's margin of error, range and trusting it's working correctly.
When properly trained, fine-tuned and prompted, LLMs can be very effective in well-defined, non-physical domains like logic, writing, math and code but making things function in the real-world quickly spirals into combinatorial complexity.
─ 3 more ─
>They are limited to the sensor's margin of error, range and trusting it's working correctly
So are we and our sensors can be pretty vague in comparison, I can only imagine human error correction is pretty next level.
>They can't confirm ground truth about physical reality.
That's true. Of course we can hook an llm up to Motors and sensors. That's a robot. Or a self-driving car. So would you say that those devices can confirm the ground truth about physical reality, and therefore are capable of creativity?
What is ground truth? If I see a digital painting by a human is that ground truth?
perhaps, but prompting is quickly becoming another area where humans are no longer clearly superior.
Getting LLMs to prompt other LLMs in a loop is not hard, it doesn't produce great results most of the time, but that is changing.
What is “getting LLMs to prompt other LLMs” if not a prompt?
─ 7 more ─
I'm pretty sure the initial post said "humans had to prompt", or otherwise insinuate a human in the loop.
─ 6 more ─
Right. Even when an LLM prompts an LLM, what prompted the first one in the chain?
─ 5 more ─
You are describing the " unmoved mover", an argument for the existence of God
─ 4 more ─
I don't think so...? Maybe I'm understanding the article wrong? I get the overlay, but I'm being specific to LLM chains... They require something to kick them off, they do not act autonomously. A human does something, types, speaks, presses a button, whatever. A Cron Job firing off is the result of a prior prompt, by a human.
I don't follow the chain further back than that, as I believe humans have free will. I get that that's debated, but thats why I draw the line at human action.
─ 3 more ─
I don't think Free Will is the issue here. My argument is that humans require input to act. We can't test that theory, because there is never been a human being who did not receive input. Everything that we experience from birth through childhood into adulthood is our input and forms our choices about what to do. That includes what the study, who to love, what our hobbies are, what kind of work we pursue.
─ 1 more ─
Legally, we draw the line at the human causing the action (Insanity pleas withstanding). I don't see why we treat this any differently.
>I don't follow the chain further back than that
Why even have an argument if you get to pick the random constraints that have a lot of issues in meshing with reality.
We are a chain that started 4 billion years ago from seemingly nothing and lead to this point. When inventing X-risk AI it won't look any different. One prompt is entered and another 4 billion year chain starts electronically instead of biologically.
AI (GPT-Sol-6:xhigh) isn't even capable of using Altium to re-layout a board without introducing a bunch of problems by a user who isn't an electrical engineer or familiar with board layouts.
AI is capable of developing its own inference hardware, once a prompt is given. The fact that a human happens to kick it off here is not especially relevant to the fact that an AI is performing the task independently. There are plenty of ways a text prompt can be generated: a harness, another LLM, or just removing stop tokens so that once begun the AI will continue until its hardware fails.
It's not clear what this fad of attributing everything an AI does to the human prompting it is supposed to accomplish.
>It's not clear what this fad of attributing everything an AI does to the human prompting it is supposed to accomplish.
It's meant to assign agency and accountability where it actually lies instead of mystifying it with anthropomorphic language.
Failing to do so has real and harmful consequences, such as enabling OpenAI to escape accountability for clearly criminal behavior.
─ 10 more ─
Denying that "AI is now capable of developing its own inference hardware" on the grounds a human asked for this to happen and that humans need to be around to blame, is as useful as denying "Atomic bombs are now capable of levelling cities" on the grounds some humans had to build it, others had to put it in a delivery system, and someone had to give the order for its use.
Questions of agency are for lawyers, questions of personhood for philosophers, we're engineers and our question is capability.
Does it really have the capability? By default I'm sceptical for the same reasons given by sailingparrot: https://news.ycombinator.com/item?id=49982068
─ 5 more ─
You are 100% correct an atomic weapon is not capable of destroying a city without humans doing a bunch of things. Hell putting a normal bomb on a plane and exploding it takes several people with specific technical skills and knowledge doing things.
>Questions of agency are for lawyers, questions of personhood for philosophers, we're engineers and our question is capability.
But it's objectively not capable without a human specifying things through prompts and training. Same as an oven can't cook a three course meal without a chef. We get around that with training data but there will always be things with no/less data or outdated knowledge.
─ 3 more ─
That limitation feels very much artificially imposed today, much like LLMs (mostly) not being able to learn from prompts or alter their weights on the fly.
One can stand up an agent like openclaw and prompt it with something very general. That would kick off a recurring loop that could indeed see the agent decide for itself it needs to design new chip hardware. Technically a human kick started that loop, but when does that stop being important? I don't attribute all of my actions and decisions to the fact that my parents brought me into this world, for example, and I'd hope they aren't legally on the hook if I screw up.
>But it's objectively not capable without a human specifying things through prompts and training. Same as an oven can't cook a three course meal without a chef.
Not in the same way. The only thing an oven can do by itself is thermostatic.
All machine learning (LLMs included, but way broader than that) is bad at learning compared to any organic brain, to the extent that any organism this bad would starve before learning to eat. However, even a human can't become a chef unless trained, we learn a lot more than we innovate, and what we happen to want without prompting is not generally well aligned with what is desired by people who pay us, which is why we need all those boring workplace things like "a boss".
But even then, this diversion is like saying "an oven can't cook a meal" in response to someone saying they built an oven and "it cooked a meal". Like, it's obvious they didn't mean it did every step by itself without anyone ever even bothering to ask it to: if they meant that version, they'd be a lot louder about it.
>We get around that with training data but there will always be things with no/less data or outdated knowledge.
And? The linked git page (implicitly) claims that there is sufficient training data to do this task.
It may, of course, be wrong. I won't be surprised if it turns out this simulation is too far from reality. But the claim is "it does ${thing} now", not "it's generally intelligent and can do everything now".
>objectively not capable without a human specifying things through prompts and trainin
huh.
Neither can you. If I drop your ass off in the woods at a few days old, you're back to 10,000 BC, hell more like 200,000 BC. So I don't get why you have these weird pendetic responses that are completely out of scope.
Also, a huge portion of AI training these days has nothing to do with humans, AI trains AI.
>there will always be things with no/less data or outdated knowledge.
And guess what, you're not doing them either! HN posters keep acting like humans are an island, but nothing in the modern world works without a society. Once you put an AI in a harness that can ask questions it isn't really much different than you.
And then in response you get ridiculous point-missing strawman non-analogies like ovens not cooking without a chef. Indeed, ovens and AI are quite different, in very relevant ways.
This conflates different issues and is akin to "guns don't kill people, people kill people" -- a denial that "has real and harmful consequences".
─ 1 more ─
Is your argument that the "guns don't kill people, people kill people" line is inaccurate or simply unhelpful towards certain end goals?
>It's meant to assign agency and accountability where it actually lies instead of mystifying it with anthropomorphic language.
Yesterday, my boss told me to fix a bug in our product. Then, I fixed it. Today my boss is taking credit, saying that he fixed it. I guess he's right, since he told me to do it.
>It's meant to assign agency and accountability where it actually lies instead of mystifying it with anthropomorphic language.
> Failing to do so has real and harmful consequences, such as enabling OpenAI to escape accountability for clearly criminal behavior.
Lol, I guess we cannot even say that AI is capable of doing something (obviously kicked off with a prompt, that goes without saying), because some people immediately get OpenAI hacking derangement syndrome.
OpenAI should be held accountable if actual damages happened, but I am not going to change completely normal speech figures in order to maybe bring it 0.01% closer.
How many "prompts" from our parents, teachers, bosses, etc. it took for all of us to come here and discuss this? Or for that human to think of that prompt for that LLM?
These are clearly rhetorical questions, but think of the metaphysical implication of your contestation. Ex nihilo nihil fit.
What if that initial prompt never asked for this hardware to be developed, and it was just one piece of the puzzle to answer to that prompt? That it took a chain of thousands of agents to prompt each others to come up with that?
well, let me know when they have proper context management, etc.
oh. they do. I built that. It's pretty fancy.
But there's still human direction behind most projects in the cybersphere.
─ 1 more ─
Text files?
Bulldozers, excavators and rollers are now capable of building roads.
Opened the RTL, looked at the floating point math, learned that apparently you don't need correct floating point operations for LLMs, closed the page.
Indeed there are some bugs/"non standard behavior" regarding very small or very big floating point values. All of those where proven harmless for LLM inference. Thanks for pointing out the bug. If you caught something outside of that, please point that out so I can fix it ;)
Well, you don't. That's why 1.58-bit (ternary) quants are often used. But if that's what they're going for, no need to dress it up in a floating-point facade.
I've been vibecoding an open source hardware AV1 decoder: https://github.com/Jonahss/openav1
After using AI to develop risc-v CPU cores, the same technique was used for developing openTPU. An open source AI inference engine. It's able to run most of the modern models like Qwen 3.5, Gemma 4, and many others. The TPU started able to produce only a few tokens per second and trough a recursive self improvement loop got to 80+ tok/sec on the smallers models.