YC is wrong about LLMs for chip design
zach.be
zach.be
A few months ago I saw a post on LinkedIn where someone fed the leading LLMs a counter-intuitively drawn circuit with 3 capacitors in parallel and asked what the total capacitance was. Not a single one got it correct - not only did they say the caps were in series (they were not) it even got the series capacitance calculations wrong. I couldn’t believe they whiffed it and had to check myself and sure enough I got the same results as the author and tried all types of prompt magic to get the right answer… no dice.
I also saw an ad for an AI tool that’s designed to help you understand schematics. In its pitch to you, it’s showing what looks like a fairly generic guitar distortion pedal circuit and does manage to correctly identify a capacitor as blocking DC but failed to mention it also functions as a component in an RC high-pass filter. I chuckled when the voice over proudly claims “they didn’t even teach me this in 4 years of Electrical Engineering!” (Really? They don’t teach how capacitors block DC and how RC filters work????)
If you’re in this space you probably need to compile your own carefully curated codex and train something more specialized. The general purpose ones struggle too much.
RC circuits man.
…Then takes a class on anything with 3d graphics… “oh shit matrix algebra again!”
…then takes a class on machine learning “urg more matrix math!”
Linear algebra, optimization, gradients, etc are 1st and 2nd year topics for STEM.
The software industry is dominated by self taught hackers and EE/math/physics drop outs. CS prestige is less than 10 years old.
Intro EE is kinda brutal in that there’s a lot of theory to cover, and you need to build the intuition on how it applies to real world circuit design on the fly.
I had a bit of an epiphany when I was in a set theory/number theory class and some classmates were breezing through proofs that I struggled with. I was having to do algebraic manipulations in a way that was novel to me, but was intuitive to math nerds. I felt like that guy who didn’t “get” the intuition in an intro programming or circuits class.
But yeah, students often get some context for math or programming in high school, but rarely for circuit design. E&M in physics at best. EE programs have solved this by weeding out anyone who can’t bash their way through the foundational theory… which isn’t great.
If you’re still interested, I would recommend the Student Manual to the Art of Electronics. It’s a very practical, lab-based book that throws out a lot of the math in favor of rules of thumb and gaining intuition for circuit design.
I graduated in chemistry, and Chemistry 1 in engineering had tests much more difficult than any other Chemistry 1 in any other faculty. After noticing that the same pattern applied to Physics 1 or Calculus I started realizing it was an engineering thing, which was later confirmed to me by an associate professor that was the design.
I asked him why, and he told me that it's a long established thing that you don't want people that struggle with science fundamentals to build bridges, ships or electrical circuits so the first semesters are very focused on this weeding.
And this at a top ten school for CS.
There are healthy ways to exploit an urge to procrastinate but this is just feeding the monster, and I hope the prof was ashamed of himself.
Why should we expect a general-purpose instruction-tuned LLM to get this right in the first place? I am not at all surprised it didn't work, and I would be more than a little surprised if it did.
The argument goes: Language encodes knowledge, so from the vast reams of training data, the model will have encoded the fundamentals of electromagnetism. This is based in the belief that LLMs being adept at manipulating language, are therefore inchoate general intelligences, and indeed, attaining AGI is a matter of scaling parameters and/or training data on the existing LLM foundations.
In reality, and with my biases as self-taught person, experience is crucial. Learning on the field. 10,000 hours of practice. Something LLMs are not very good at. You train them a priori, then it's a relatively static product compared to how human brains operate and self-adjust.
https://www.scientificamerican.com/article/you-dont-need-wor...
And I'm baffled that HN is not picking up on that and ACTUALLY BELIEVES that you can achieve AGI with a simple language model scaled to billions of parameters.
It's as futile as trying to explain vision to a blind man using "only" a few billion words. There's simply no string of words that can create a meaningful representation in the mind of the blind man.
> “they didn’t even teach me this in 4 years of Electrical Engineering!” (Really? They don’t teach how capacitors block DC and how RC filters work????)
My experience with being an adult, in general, is that many people who went to university don't believe that any given course taught them anything meaningful.
I can absolutely believe that such people didn't learn and remember anything meaningful from those courses. Whether the course is to blame, is far more questionable.
It's the same as all the people who say "Why didn't high school teach me how to balance a check book or calculate a mortgage or blah blah?"
In nearly every case, they literally did, but you weren't paying attention.
You also had to cheat off me to pass biology, so I'm going to go ahead and press X to doubt that you "understand the immune system"
We are surrounded by people who failed to invest in their own education, and instead of facing that awful reality, they INSIST that WE are the dumb ones.
It's infuriating.
“Where are you from?”
“What’s the chemistry of your required sustenance?”
“How long is your sleep cycle as measured with physical time constants?”
And similar basic questions could not be answered by 99.9% of the human population.
Fundamentally, almost none of us can give an accurate answer to what were made of, where we’re from, or what we need to survive.
It made me aware of how unlikely I'd be to ever get it right.
Where are you from is difficult given that we don't know how the alien cop's map is drawn. Third planet from a star shining at 5700K that is 8 kiloparsecs from the galactic center is only slightly more useful than lost kid and saying that their mom's name is mommy.
Chemistry of sustenance. We're carbon based and everything comes from that, but constructing a description of edible food from raw elements is going to take a lot more than drawing some hexagons with C H N and O, along with other required elements. Before we get to food and H₂O though, we'd need an atmosphere to breathe, I wouldn't presume the alien cops know to have an oxygen/nitrogen mix for humans, and not something that's poisonous for humans, like CO.
Time is something that's possible to express though. SI defined the second as a number of vibrations of a Cesium-133 atom, 8 hours of sleep is just multiplication.
Don't think anybody could describe what/where/what to an alien cop that doesn't even speak English to get themselves home or even to not die in an alien atmosphere.
Edit: I believe that LLM's are eminently useful to replace experts (of all people) 90% of the time.
What do you mean by "expert"?
Do you mean the pundit who goes on TV and says "this policy will be bad for the economy"?
Or do you mean the seasoned developer who you hire to fix your memory leaks? To make your service fast? Or cut your cloud bill from 10M a year to 1M a year?
Experts capable of critical thinking and reflecting on evidence that contradicts their world model (and thereby retraining it on the fly)? Most likely not, at least not in their current architecture with all its limitations.
I’d disagree, though: humans are still easier to predict and understand (and trust) than AI, typically.
In this example, GPT-4o cannot tell that GitHub is spelled correctly:
https://app.gitsense.com/?doc=6c9bada92&model=GPT-4o&samples...
In this example, Claude cannot tell that GitHub is spelled correctly:
https://app.gitsense.com/?doc=905f4a9af74c25f&model=Claude+3...
I still believe LLM is a game changer and I'm currently working on what I call a "Yes/No" tool which I believe will make trusting LLMs a lot easier (for certain things of course). The basic idea is the "Yes/No" tool will let you combine models, samples and prompts to come to a Yes or No answer.
Based on what I've seen so far, a model can easily screw up, but it is unlikely that all will screw up at the same time.
But we have had extensive experience with humans, it is normal to have better defined trust, LLMs will be better understood as well. There is no central understander or truth, that is the interesting part, it's a "Blind men and the elephant" situation.
Sure, EDA tools are deterministic, but the humans who apply them are not. Introducing LLMs to these processes is not some radical and scary departure, it’s an iterative evolution.
Its really just that the "in principle" part of the overall implication with your comment and so many others just doesn't make sense. Its very much cutting off your nose to spite your face. How could science itself be possible, much less engineering, if this is how we decided things? If we regarded ourselves always from the outside? How could even be motivated to debate whether we get the computers to design their own chips? When would something actually happen? At some point, people do have ideas, in a full, if false, transparency to themselves, that they can write down and share and explain. This is not only the thing that has gotten us this far, it is the very essence of why these models are so impressive in the certain ways that they are. It doesn't make sense to argue for the fundamental cheapness of the very thing you are ultimately trying to defend. And it imposes this strange perspective where we are not even living inside our own (phenomenal) minds anymore, that it fundamentally never matters what we think, no matter our justification. Its weird!
I'm sure you have a lot of good points and stuff, I just am simply pointing out that this particular argument is maybe not the strongest.
I accept that I’m fallible, both in my areas of expertise and in all the meta stuff around it. I code bugs. I omit requirements. Not often, and there are mental and technical means to minimize, but my work, my org’s structure, my company’s processes are all designed to mitigate human fallibility.
I’m not interested in “defending” AI models. I’m just saying that their weaknesses are qualitatively similar to human weaknesses, and as such, we are already prepared to deal with those weaknesses as long as we are aware of them, and as long as we don’t make the mistake of thinking that because they use transistors they should be treated like a mostly deterministic piece of software where one unit test pass means it is good.
I think you’re reading some kind of value judgement on consciousness into what is really just a pragmatic approach to slotting powerful but imperfect agents into complex systems. It seems obvious to me, and without any implications as to human agency.
By and large, the processes people are scrambling to place LLMs in are ones that typical machines struggle or fail and humans excel or do decently (and that LLMs are making some headway in).
There's no point comparing LLM performance to some hypothetical perfect understanding machine that doesn't exist. It's nonsensical actually. You compare it to the performance of the beings it's meant to replace or augment - humans.
Replacing non-deterministic black boxes with potentially better performing non-deterministic black boxes is not some crazy idea.
If you could that would be nice wouldn't it? And if you couldn't?
If people were saying, "let's replace Casio Calculators with interfaces to GPT" then that would be crazy and I would wholly agree with you but by and large, the processes people are scrambling to place LLMs in are ones that typical machines struggle or fail and humans excel or do decently (and that LLMs are making some headway in).
You're making the wrong distinction here. It's not Dave vs your nifty script. It's Dave or nothing at all.
There's no point comparing LLM performance to some hypothetical perfect understanding machine that doesn't exist.
You compare to the things its meant to replace - humans. How well can the LLM do this compared to Dave ?
I'm pretty sure they are scrambling to put them absolutely anywhere it might save or make a buck (or convince an investor that it could)
For example, using a LLM to transform structured data into JSON, and doing it with two LLMs in parallel to try to catch the inevitable failures, instead of just writing code that outputs JSON.
I guess I just don't see your point. So a few purported applications are not very sensible. So what ? This is every breakthrough ever.
LLM’s are not good at actually doing the processing, they are not good at math or even text processing at a character level. They often get logic wrong. But they are pretty good at looking at patterns and finding creative solutions to new inputs (or at least what can appear creative, even if philosophically it’s more pattern matching than creativity). So an LLM would potentially be good at writing a first draft of that script, which Dave could then proofread/edit, and which a standard deterministic computer could just run verbatim to actually do the processing. Eventually maybe even Dave’s proofreading would be superfluous.
Tying this back to the original article, I don’t think anyone is proposing having an LLM inside a chip that processes incoming data in a non-deterministic way. The article is about using AI to design the chips in the first place. But the chips would still be deterministic, the equivalent of the script in this analogy. There are plenty of arguments to make about LLM‘s not being good enough for that, not being able to follow the logic or optimize it, or come up with novel architectures. But the shape of chip design/Verilog feels like something that with enough effort, an AI could likely be built that would be pretty good at it. All of the knowledge that those smart knowledgeable engineers which are good at writing Verilog have built up can almost certainly be represented in some AI form, and I wouldn’t bet against AI getting to a point where it can be helpful similarly to how Copilot currently is with code completion. Maybe not perfect anytime soon, but good enough that we could eventually see a path to 100%. It doesn’t feel like there’s a fundamental reason this is impossible on a long enough time scale.
Right, and there’s nothing fundamentally wrong with this, nor is it a novel method. We’ve been joking about copying code from stack overflow for ages, but at least we didn’t pretend that it’s the peak of human achievement. Ask a teacher the difference between writing an essay and proofreading it.
Look, my entire claim from the beginning is that understanding is important (epistemologically, it may be what separates engineering from alchemy, but I digress). Practically speaking, if we see larger and larger pieces of LLM written code, it will be similar to Dave and his incomprehensible VBA script. It works, but nobody knows why. Don’t get me wrong, this isn’t new at all. It’s an ever-present wet blanket that slowly suffocates engineering ventures who don’t pay attention and actively resist. In that context, uncritically inviting a second wave of monkeys to the nuclear control panels, that’s what baffles me.
Tangent for a slight pet peeve of mine:
"We" did joke about this, but probably because most of our jobs are not in chip design. "We" also know the limits of this approach.
The fact that Stack Overflow is the most SEO optimised result for "how to center div" (which we always forget how to do) doesn't have any bearing on the times when we have an actual problem requiring our attention and intellect. Say diagnosing a performance issue, negotiating requirements and how they subtly differ in an edge case from the current system behaviour, discovering a shared abstraction in 4 pieces of code that are nearly but not quite the same.
I agree with your posts here, the Stack Overflow thing in general is just a small hobby horse I have.
I look up "how do I sort a list in language X" because I know from school that there IS a defined good way to do it, probably built into the language, and it will be extremely idiomatic, but I haven't used language X in five years and the specifics might have changed and I don't remember the specific punctuation.
Or Dave could write a first draft of that script, saving him the time needed to translate what the LLM composed.
No, I'm just disappointed in the decision of Black Box A and am bound to be even more disappointed by Black Box B. If we continue removing thoughtful design from our systems because thoughtlessness is the default, nobody's life will improve.
I like my job.
My job also involves cooperating with other non-deterministic black boxes (colleagues).
I can totally see how artificial non-deterministic black boxes (artificial colleagues) may be useful to replace/augment the biological ones.
For one, artificial colleagues don't get tired and I don't accidentally hurt their feelings or whatnot.
In any case, I'm not looking forward to replacing my deterministic tools with the fuzzy AI stuff.
Intuitively at least it seems to me that these non-deterministic black boxes could really benefit from using the deterministic tools for pretty much the same reasons we do as well.
Does an LLM know math? Not like we do. There’s no deductive logic in there; it’s all statistical inferences from language. An LLM doesn’t “work through” a circuit diagram systematically the way a physics student would. It observes the entire diagram at once, and then guesses the most likely next token.
Hello, fellow tech enthusiasts, just stopping by to announce I performatively can't tell the difference between "Latest big tech product (TM)" and Homo Sapiens Sapiens!!!
I'll be seeing you in the next LLM related message thread with the same exact comment!!! As you were!!!
But I certainly agree in general. It’s been years and there are still no independent novel discoveries afaik.
ESM3: https://www.evolutionaryscale.ai/blog/esm3-release
AlphaProof/AlphaGeometry2: https://deepmind.google/discover/blog/ai-solves-imo-problems...
MatPilot discovering new materials: https://arxiv.org/abs/2411.08063
Then of course NVidia Omniverse with their digital-twin learning.
https://blog.google/technology/ai/google-ai-big-scientific-b...
YC is technically incompetent and isn't about making the world better. Every single one of their words is a lie and hides the real intent: make money.
Second, want to give any examples of "shitty, hype-based compan[ies]" (I assume you mean companies with no real revenue traction) getting bought out for "a few billion".
Third, investment banks facilitate sales of assets, they don't buy them themselves.
Maybe sit out the conversation if you don't even know the basics of how VC, startups, or banking work?
https://www.reuters.com/article/business/peloton-raises-12-b...
The author here is missing a few important things about chip design. Most of the time spent and work done is not writing high performance Verilog. Designers spent a huge amount of time answering questions, writing documentation, copying around boiler plate, reading obscure manuals and diagrams, etc. LLMs can already help with all of those things.
I believe that LLMs in their current state could help design teams move at least twice as fast, and better tools could probably change that number to 4x or 10x even with no improvement in the intelligence of models. Most of the benefit would come from allowing designers to run more experiments and try more things, to get feedback on design choices faster, to spend less time documenting and communicating, and spend less time reading poorly written documentation.
> Well, it turns out that LLMs are also pretty valuable when it comes to chips for lucrative markets -- but they won’t be doing most of the design work. LLM copilots for Verilog are, at best, mediocre. But leveraging an LLM to write small snippets of simple code can still save engineers time, and ultimately save their employers money.
I think designers getting 2x faster is probably optimistic, but I also could be wrong about that! Most of my chip design experience has been at smaller companies, with good documentation, where I've been focused on datapath architecture & design, so maybe I'm underestimating how much boilerplate the average engineer deals with.
Regardless, I don't think LLMs will be designing high-performance datapath or networking Verilog anytime soon.
At large companies with many designers, a lot of time is spent coordinating and planning. LLMs can already help with that.
As far as design/copilot goes, I think there are reasons to be much more optimistic. Existing models haven't seen much Verilog. With better training data it's reasonable to expect that they will improve to perform at least as well on Verilog as they do on python. But even if there is a 10% chance it's reasonable for VCs to invest in these companies.
Anyway that’s largely anecdata/sample size of 1, and it could very well be a case of me holding the tool wrong, but that’s what I observed.
There simply isn't enough of that code in existence.
Writing Verilog code is about mapping the constructs onto your theory of mind about the underlying hardware. If that were easy, so many engineers wouldn't have so much trouble writing Verilog code that doesn't have faults. You can't write Verilog code just by pasting together Stack Overflow snippets.
Look at the confusion that happens when programmers take their "for-loop" understanding into the world of GPU shaders or HDLs (hardware description languages) where "for-loops" map to hardware and suddenly are both finite and fixed. LLMs exhibit the exact same confusion--only worse.
But there may still be value in YC calling for innovation in that space. The article is correctly showing that there is no easy win in applying LLMs to chip design. Either the market for a given application is too small, then LLMs can help but who cares, or the chip is too important, in which case you'd rather use the best engineers. Unlike software, we're not getting much of a long tail effect in chip design. Taping out a chip is just not something a hacker can do, and even playing with an FPGA has a high cost of entry compared to hacking on your PC.
But if there was an obvious path forward, YC wouldn't need to ask for an innovative approach.
seen here as well when george-hotz attempts to overthow the chip companies with his plan for an ai chip https://geohot.github.io/blog/jekyll/update/2021/06/13/a-bre... little realizing the complexity involved. to his credit, he quickly pivoted into a software and tiny-box maker.
How many experts do YC have on chip design?
Even obvious can be risky. First it's nice to share the risk, second more investments come with more connections.
As for LLMs boom. I think finally we'll realize that LLM with algorithms can do much more than just LLM. 'algorithms' is probably a bad word here, I mean assisting tools like databases, algorithms, other models. Then only access API can be trained into LLM instead of the whole dataset for example.
A bit level (non von Neumann) general purpose systolic array could greatly speed up AI computations, along with almost everything else. It's a chip to do general purpose computation.
The chip design is almost trivial. I'd expect someone with a few years of experience could knock it out in a few days. I hope to field a design in the next TinyTapeout (I'm on a fixed income, so I've had to wait a while)
The real problem is programming. We're talking vast greenfields that go on forever. There's no good way to target the architecture, you certainly wouldn't want to use Verilog or any other HDL.
VCs are not investing in the current LLM-based systems to improve X, they're investing in a future where LLM based systems will be 100x more performant.
Writing is complex, LLMs once had subhuman performance, and yet. Digital art. Music (see suno.AI) There is a pattern here.
If that’s the biggest gap, then YC is correct that it’s a good area for a startup to tackle.
The reason is that those experts do not own the code that they have written.
The code is owned by big companies like NVIDIA, AMD, Intel, Samsung and so on.
It is unlikely that these companies would be willing to provide the code for training, except for some custom LLM to be used internally by them, in which case the amount of code that they could provide for training might not be very impressive.
Even a designer who works in those companies may have great difficulties to see significant quantities of archived Verilog/VHDL code, though it can be hoped that it still exists somewhere.
Not my field of expertise but there seem to be experts founding startups etc in the ASIC space, and Bitcoin miners were designed and built without any of the big companies participating. So I’m not following why we need Intel to be involved.
An obvious way to set up the flywheel here is to hire experts to do professional services or consulting on customer-submitted designs while you build up your corpus. While I said “fine-tuning”, there is probably a lot of agent scaffolding to be built too, which disproportionately helps bigger companies with more work throughput. (You can also acquire a company with the expertise and tooling, as Apple did with PA Semi in ~2008, though obviously $100m order of magnitude is out of reach for a startup. https://www.forbes.com/2008/04/23/apple-buys-pasemi-tech-ebi...)
One could author some projects that can be implemented in FPGAs, but those do not provide good training material for generating code that could be used to implement a project in an ASIC, because the constraints of the design are very different.
Designing an ASIC is a year-long process and it is never completed before testing some prototypes, whose manufacture may cost millions. Authoring some Verilog or VHDL code for an imaginary product that cannot be tested on real hardware prototypes could result only in garbage training material, like the code of a program that has never been tested to see if it actually works as intended.
Learning to design an ASIC is not very difficult for a human, because a human does not need a huge number of examples, like ML/AI. Humans learn the rules and a few examples are enough for them. I have worked in a few companies at designing ASICs. While those companies had some internal training courses for their designers, those courses only taught their design methodologies, but with practically no code examples from older projects, so very unlikely to how a LLM would have to be trained.
IE I think you are wildly underestimating both the scale of training data needing, and wildly overestimating the amount of verilog code possessed by nvidia.
GPU's work by having moderate complexity cores (in the scheme of things) that are replicated 8000 times or whatever. That does not require having 8000 times as much useful verilog, of course.
The folks who have 8000 different chips, or 100 chips that each do 1000 things, would probably have orders of magnitude more verilog to use for training
This validation costs millions, which is why it is hard to enter this field, even as a fabless designer.
Many design errors are not caught even during hardware testing, but only after mass production, like the ugly MONITOR/MWAIT bug of Intel Lunar Lake.
Randomly-generated HDL code, even if it does not have syntax errors, and even if some testbench for it does not identify deviations from its specification, is not more likely to be valid when implemented in hardware, than the proverbial output of a typewriting monkey.
Ever try to get ChatGPT to play scrabble? Ever try to describe the board to it and then all the letters available to you? Even its fancy pants o1 preview performs absolutely horrible. Either my prompting completely sucks or an LLM is just the wrong tool for the job.
It’s great for asking you to score something you just created provided you tell it what bonuses apply to which words and letters. But it has absolutely no concept of the board at all. You cannot use to optimize your next move based on the board and the letters.
… I mean you might if you were extremely verbose about every letter on the board and every available place to put your tiles, perhaps avoiding coordinates and instead describing each word, its neighbors and relationships to bonus squares. But that just highlights how bad a tool an LLM is for scrabble.
Anyway, I’m sure schematics are very similar. Maybe somebody we will invent good machine learning models for such things but an LLM isn’t it.
To a nail, every hammer has a purpose.
The main problem is an optimal decomposition of the big project into a collection of interconnected modules and in defining adequate interfaces between modules.
This is not difficult when the purpose of the project is to just take an older project and make some improvements to it, when a suitable structure is already known, but it is always the main difficulty when a really new problem must be solved.
I have yet to see any example when a LLM can be used to help even in the slightest way to solve such an example of "divide et impera" for something novel, where novel by definition means that the training set has not contained the solution for an identical project.
There is pretty much no relationship between the 2-dimensional or multi-dimensional structural graph of the interconnected modules, together with the descriptions of their matching interfaces, and the proximity or frequency of tokens in the description of the circuit by a hardware design language. So there is little that a LLM could use to generate any HDL program for an unknown circuit.
What a LLM could do is only after a good designer has done the difficult job to decompose the project into modules and define the interfaces. When given a small module with its defined interfaces, a LLM might be able to find some boilerplate code to speed up the implementation of the module.
However, any good designer would already have templates for the boilerplate code and I can not really imagine how a LLM could do this faster than a designer who just selects the appropriate templates and pastes them into the module.
And now they can easily replace mediocre human performance, and since they are tuned to provide answers that appeal to humans that is especially true for these subjective value use cases. Chip design doesn't seem very similar. Seems like a case where specifically trained tools would be of assistance. For some things, as much as generalist LLMs have surprised at skill in specific tasks, it is very hard to see how training on a broad corpus of text could outperform specific tools — for first paragraph do you really think it is not dubious to think a model trained on text would outperform Stockfish at chess?
Also, they might not be able to do it. eg most models can't generate "horse riding an astronaut" or "upside-down car".
I just entered your prompt into an AI image generator and in under a second it gave me an image[0] of what looks to me like an anthropomorphic dolphin sitting at a desk writing a letter in a little study. I then had to google what the difference between a porpoise and a dolphin was because I genuinely thought porpoises looked much more like manatees. While I could nitpick the AI's work for making the porpoise's snout a little too long, had I drawn it the porpoise would have been a vaguely marine looking blob with no anatomy detailed enough to recognize let alone criticize. I am quite confident that if you asked for a large number of images based on that prompt from humans, it would easily rank among the best, and it's unlikely you'd get any which were markedly better. The fact it can generate this image nearly instantaneously though is astounding. If your goal was to get one masterpiece hanging in the Louvre, this particular tool would not suffice, but if your goal was to illustrate children's books, this tool could do in hours what would have taken a team of humans months. That is superhuman performance.
[0] https://api.deepai.org/job-view-file/e0b80ca6-d934-42e4-9a7e...
(Sorry if the link doesn't remain good for long)
Whether humans or AI are better at the task overall is probably too vague a question to answer, depending a lot on how you weight different desirables.
But if you're using AI to create art, you're typically not trying to move someone's soul. You're trying to create a work that depicts something in a particular style with a particular fidelity with a certain amount of resource consumption. That is the only metric by which it makes any sense to evaluate the machine designed to do that specific task.
If LLMs will do well in the space for some use case it's the established chip designers that will benefit from it, not a small startup.
I do not think they mean to say that an AI would be 100 times better at designing chips than a human, I assume this is the engineering tradeoff they refer to. Though I wouldn't fault anyone for being confused, as the wording is painfully awkward and salesy.
I also think OP is missing the point saying the target applications are too small of a market to be worth pursuing.
They’re too small to pursue any single one as the market cap for a company, but presumably the fictional AI chip startup could pursue many of these smaller markets at once. It would be a long tail play, wouldn’t it?
I agree they're probably wrong but this article doesn't actually explain why they're wrong to bet on exponential progress in AI capabilities.
We have a bunch of AI initiatives in my company but most of them are about using Copilot to help write scripts to automate the design flow. Our physical design flow are thousands of lines of Tcl and Python code.
The article mentions High Level Synthesis. I've been reading about this since my first job in the 1990's. I've worked on at least 80 chips and I've never seen any chip use one of these tools except for some tiny section that was written by some academics who didn't want to learn Verilog for reasons.
It's fundamentally important when doing hardware design to work in a language that _expresses_ itself like you're designing hardware. Verilog (for all its faults) shines there because it feels like you're writing a slightly higher level netlist. That's not the case with SC and friends, which doesn't allow you to think in hardware. Languages like BSV and SV are functionally similar but they force you to think in similar ways to Verilog, meaning you can write much tighter high-level code.
I'd be interested in your experience, but I feel that using normal programming languages to build hardware is an abstraction failure. Which is why it performs so poorly.
Anyways, their IP very clearly violated the standards of a very well known interface, which could have spelled disaster at tape-out. I had to fight tooth-and-nail, and spent lots of my company's time trying to convince this third-party vendor that this was an actual issue. Only months later were they convinced. The revised code kept coming back and failing interface checks, which shows that they weren't doing these checks on their end. All I could think is, "this can't go well..."
Software engineers get hyped when they see the progress in AI coding and immediately begin to extrapolate to other fields—if Copilot can reduce the burden of coding so much, think of all the money we can make selling a similar product to XYZ industries!
The problem with this extrapolation is that the software industry is pretty much unique in the amount of information about its inner workings that is publicly available for training on. We've spent the last 20+ years writing millions and millions of lines of code that we published on the internet, not to mention answering questions on Stack Overflow (which still has 3x as many answers as all other Stack Exchanges combined [0]), writing technical blogs, hundreds of thousands of emails in public mailing lists, and so on.
Nearly every other industry (with the possible exception of Law) produces publicly-visible output at a tiny fraction of the rate that we do. Ethics of the mass harvesting aside, it's simply not possible for an LLM to have the same skill level in ${insert industry here} as they do with software, so you can't extrapolate from Copilot to other domains.
> Nearly every other industry (with the possible exception of Law) produces publicly-visible output at a tiny fraction of the rate that we do.
You are correct! There's lots of information available publicly about certain things like code, and writing SQL queries. But other specialized domains don't have the same kind of information trained into the heart of the model.
But importantly, this doesn't mean the LLM can't provide significant value in these other more niche domains. They still can, and I provide this every day in my day job. But it's a lot of work. We (as AI engineers) have to deeply understand the special domain knowledge. The basic process is this:
1. Learn how the subject matter experts do the work.
2. Teach the LLM to do this, using examples, giving it procedures, walking it through the various steps and giving it the guidance and time and space to think. (Multiple prompts, recipes if you will, loops, external memory...)
3. Evaluation, iteration, improvement
4. Scale up to production
In many domains I work in, it can be very challenging to get past step 1. If I don't know how to do it effectively, I can't guide the LLM through the steps. Consider an example question like "what are the top 5 ways to improve my business" -- the subject matter experts often have difficulty teaching me how to do that. If they don't know how to do it, they can't teach it to me, and I can't teach it to the agent. Another example that will resonate with nerds here is being an effective Dungeons and Dragons DM. But if I actually learn how to do it, and boil it down into repeatable steps, and use GraphRAG, then it becomes another thing entirely. I know this is possible, and expect to see great things in that space, but I estimate it'll take another year or so of development to get it done.
But in many domains, I get access to subject matter experts that can tell me pretty specifically how to succeed in an area. These are the top 5 situations you will see, how you can identify which situation type it is, and what you should do when you see that you are in that kind of situation. In domains like this I can in fact make the agent do awesome work and provide value, even when the information is not in the publicly available training data for the LLM.
There's this thing about knowing a domain area well enough to do the job, but not having enough mastery to teach others how to do the job. You need domain experts that understand the job well enough to teach you how to do it, and you as the AI engineer need enough mastery over the agent to teach it how to do the job as well. Then the magic happens.
When we get AGI we can proceed past this limitation of needing to know how to do the job ourselves. Until we get AGI, then this is how we provide impact using agents.
This is why I say that even if LLM technology does not improve any more beyond where it was a year ago, we still have many years worth of untapped potential for AI. It just takes a lot of work, and most engineers today don't understand how to do that work-- principally because they're too busy saying today's technology can't do that work rather than trying to learn how to do it.
This will get harder I think over time as low hanging fruit domains are picked - the barrier will be people not technology. Especially if the moat for that domain/company is the knowledge you are trying to acquire (NOTE: Some industries that's not their moat and using AI to shed more jobs is a win). Most industries that don't have public workings on the internet have a couple of characteristics that will make it extremely difficult to perform Task 1 on your list. The biggest is now every person on the street, through the mainstream news, etc knows that it's not great to be a software engineer right now and most media outlets point straight to "AI". "It's sucks to be them" I've heard people say - what was once a profession of respect is now "how long do you think you have? 5 years? What will you do instead?".
This creates a massive resistance/outright potential lies in providing AI developers information - there is a precedent of what happens if you do and it isn't good for the person/company with the knowledge. Doctors associations, apprenticeship schemes, industry bodies I've worked with are all now starting to care about information security a lot more due to "AI", and proprietary methods of working lest AI accidentally "train on them". Definitely boosted the demand for cyber people again as an example around here.
> You are correct! There's lots of information available publicly about certain things like code, and writing SQL queries. But other specialized domains don't have the same kind of information trained into the heart of the model.
The nightmare of anyone that studied and invested into a skill set according to most people you would meet. I think most practitioners will conscious to ensure that the lack of data to train on stays that way for as long as possible - even if it eventually gets there the slower it happens and the more out of date it is the more useful the human skill/economic value of that person. How many people would of contributed to open source if they knew LLM's were coming for example? Some may have, but I think there would of been less all else being equal. Maybe quite a bit less code to the point that AI would of been delayed further - tbh if Google knew that LLM's could scale to be what they are they wouldn't of let that "attention" paper be released either IMO. Anecdotally even the blue collar workers I know are now hesitant to let anyone near their methods of working and their craft - survival, family, etc come first. In the end after all, work is a means to an end for most people.
Unlike us techies which I find at times to not be "rational economic actors" many non-tech professionals don't see AI as an opportunity - they see it as a threat they they need to counter. At best they think they need to adopt AI, before others have it and make sure no one else has it. People I've chatted to say "no one wants this, but if you don't do it others will and you will be left behind" is a common statement. One person likened it to a nuclear weapons arms race - not a good thing, but if you don't do it you will be under threat later.
Also consider that there exist quite a lot of subject matter experts who simply are not AI fanboys - not because they are afraid of their job because of AI, but because they consider the whole AI hype to be insanely annoying and infuriating. To get them to work with an AI startup, you will thus have to pay them quite a lot of money.
After all in a capitalist economy the last to be disrupted generally gets "all the spoils" as purchasing power (and hence prices/wages) move from least scarce/disrupted skills to more scarce skills which allows the last to be disrupted to have more time to accumulate wealth/assets to shield themselves from AI even more.
In software, we've all self taught, improved, posted Q&A all over the web. Plus all the open source code out there. Just mountains and mountains of free training data.
However software is unique in being both well paying and something with freely available, complete information online.
A lot of the rest of the world remains far more closed and almost an apprenticeship system. In my domain thinks like company fundamental analysis, algo/quant trading, etc. Lots of books you can buy from the likes of Dalio, but no real (good) step by step research and investment process information online.
Likewise I'd imagine heavily patented/regulated/IP industries like chip design, drug design, etc are substantially as closed. Maybe companies using an LLM on their own data internally could make something of their data, but its also quite likely there is no 'data' so much as tacit knowledge handed down over time.
1. More because fine-tuning with enough good Verilog as data should let the LLMs do better at avoiding mediocre Verilog (existing chip companies have more of this data already though). Plus non-LLM tools will remain, so you can chain those tools to test that the LLM hasn't produced Verilog that synthesizes to a large area, etc
2. Less because when creating more chips for more markets (if that's the interpretation of YC's RFS), the limiting factor will become the cost of using a fab (mask sets cost millions), and then integrating onto a board/system the customer will actually use. A half-solution would be if FPGAs embedded in CPUs/GPUs/SiPs on our existing devices took off
> If Gary Tan and YC believe that LLMs will be able to design chips 100x better than humans currently can, they’re significantly underestimating the difficulty of chip design, and the expertise of chip designers.
I may be confused, but isn’t the author fundamentally misunderstanding YC’s point? I read YC as simply pointing out the benefit of specialized compute, like GPUs, not making any point about the magnitude of improvement LLMs could achieve over humans.
From my reading of the RFS (not the video) it appears they are essentially asking for the next Groq or SambaNova.
Personally, this kind of communication issue would give me a long pause if I was considering YC for this segment, as this is a fairly basic thesis to communicate, and if a basic thesis can be muddled, can the advice provided be strong as well, especially compared to peer early stage funders in this space?
I'd want to know about the results of these experiments before casting judgement either way. Generative modeling has actual applications in the 3D printing/mechanical industry.
If the evaluation of the approach is "it works great if you train it on a few decades of the best designs from a successful fabless semiconductor company", I would say that if you plan to use that method as a startup, you're clearly going to fail. Nobody's going to give away their crown jewels to train an LLM that designs chips for other companies.
And much more important:
- LLMs can suddenly become more competent when you give them the right tools, just like humans. Ever try to drive a nail without a hammer?
- Models with spatial and physical awareness are coming and will dramatically broaden what’s possible
It’s easy to get stuck on what LLMs are bad at. The art is to apply an LLMs strengths to your specific problem, often by augmenting the LLM with the right custom tools written in regular code
I've driven a nail with a rock, a pair of pliers, a wrench, even with a concrete wall and who knows what else!
I didn't need to be told if these can be used to drive a nail, and I looked at things available, looked for a flat surface on them and good grip, considered their hardness, and then simply used them.
So if we only give them the "right" tools, they'll remain very limited by us not thinking about possible jobs they'll appear as if they know how to do and they don't.
The problem is exactly that: they "pretend" to know how to drive a nail but not really.
If you’re creative enough to figure out different tools for humans, you are creative enough to figure out different tools for LLMs
What is the added value of that combo and at what cost?
I don't know anything about chip design, but like any area in tech I'm certain there are cumbersome and largely repetitive tasks that can't easily be done by algorithms but can be done with human oversight by LLMs. There's efficiency to be gained here if the designer and operator of the LLM system know what they're doing.
We at Silogy [0] are directly targeting the problem of verification productivity using AI agents for test debugging. We analyze code (RTL, testbench, specs, etc.) along with logs and waveforms, and incorporate interactive feedback from the engineer as needed to refine the hypothesis.
A non-LLM monte carlo AI approach: "Pushing the Limits of Machine Design: Automated CPU Design with AI" (2023) https://arxiv.org/abs/2306.12456 .. https://news.ycombinator.com/item?id=36565671
A useful target for whichever approach is most efficient at IP-feasible design:
From https://news.ycombinator.com/item?id=41322134 :
> "Ask HN: How much would it cost to build a RISC CPU out of carbon?" (2024) https://news.ycombinator.com/item?id=41153490
Gary Tan's was right[1] in that there is a fundamental inefficiency inherent in the von Neumann architecture we're all using. This gross impedance mismatch[4] is a great opportunity for innovation.
Once ENIAC was "improved" from its original structure to a general purpose compute device in the von Neumann style, it suffered a 83% loss in performance[2] Everything since is 80 years of premature optimization that we need to unwind. It's the ultimate pile of technical debt.
Instead of throwing maximum effort into making specific workloads faster, why not build a chip that can make all workloads faster instead, and let economy of scale work for everyone?
I propose (and have for a while[3]) a general purpose solution.
A systolic array of simple 4 bits in, 4 bits out, Look Up Tables (LUTs) latched so that timing issues are eliminated, could greatly accelerate computation, in a far nearer timeframe.
The challenges are that it's a greenfield environment, with no compilers (though it's probable that LLVM could target it), and a bus number of 1.
[1] https://www.ycombinator.com/rfs-build#llms-for-chip-design
[2] https://en.wikipedia.org/wiki/ENIAC#Improvements
[3] https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu...
For example, how it would implement a 1-bit full adder? Like the nitty-gritty details: which input on which cell represents input A, which represents input B, and which represents carry-in? Which output is sum and which is carry-out? What are the functions programmed into each node that it uses?
Then show how to build a 2-bit adder from there.
I don't think he's arguing that. More that ASICs can be 100x better than CPUs for say crypto mining and that using LLM type stuff it may be possible to make them for other applications where there is less money available to hire engineers.
(the YC request https://www.ycombinator.com/rfs-build#llms-for-chip-design)
The key word here is "still".
We don't know what the limits of LLMs are.
It's possible that they will reach a dead end. But it is also possible that they will be able to do logic and math.
If (or when) they achieve that point, their performance will quickly become "superhuman" in these kinds of engineering tasks.
But the very next step will be the ability to do logic and math.
Even the serious idea that the article thinks could work is throwing the unreliable LLMs at verification! If there's any place you can use something that doesn't work most of the time, I guess it's there.
Replace all asserts with expected ==expected and most people won't notice.
Those tests were very common back when I used to work in Ruby on Rails and automatically generating test stubs was a popular practice. These stubs were often just converted into expected == expected tests so that they passed and then left like that.
Some development stacks are extremely underpowered for code verification, so they do patch the design issue. Just like some stacks are underpowered for abstraction and need patching by code generation. Both of those solve an immediate problem, in a haphazard and error-prone way, by adding burden on maintenance and code evolution linearly to how much you use it.
And worse, if you rely too much on them they will lead your software architecture and make that burden superlinear.
https://github.com/williamcotton/search-input-query/blob/mai...
It is a good test suite and it saved me quite a bit of typing!
In fact, Claude did most of the typing for the entire project:
https://github.com/williamcotton/search-input-query
BTW, I obviously didn't just type "make a lexer and multi-pass parser that returns multiple errors and then make a single-line instance of a Monaco editor with error reporting, type checking, syntax highlighting and tab completion".
I put it together piece-by-piece and with detailed architectural guidance.
It’s too resource intensive for all code, but mutation testing is pretty good at finding these sorts of tests that never fail. https://pitest.org/
It doesn't matter if it can't actually 'get there' as long as people still believe it can.
Come to think about it, a socioeconomic system dependent on population and economic growth is at a fundamental level driven by this balancing act: "We can solve every problem if we just forge ahead and keep enlarging the base of the pyramid - keep reproducing, keep investing, keep expanding the infrastructure".
It’s gonna be bad.
Should we expect money pumps to generate inflation quicker on this cycle than on the last ones? If so, why?
In these situations, I’ve been able to sufficiently program the agent that I haven’t seen too much of an issue as you described. Consistency is a feature.
Once it was spices. Then poppies. Modern art. The .com craze. Those blockchain ape images. Blockchain. Now LLM.
All of these had a bit of true value and a whole load of bullshit. Eventually the bullshit disappears and the core remains, and the world goes nuts about the next thing.
- LLMs are extremely competent at surface-level pattern matching and manipulation of the type we'd previously assumed that only AGI would be able to do.
- A large fraction of tasks (and by extension jobs) that we used to, and largely still do, consider to be "knowledge work", i.e. requiring a high level of skill and intelligence, are in fact surface-level pattern matching and manipulation.
Reconciling these facts raises some uncomfortable implications, and calling LLMs "actually intelligent" lets us avoid these.
Recently I came across some one advertising an LLM to generate fashion magazine shoot in Pakistan at 20-25% of the cost. It hit me then that they are undercutting the fashion shoot of country like Pakistan which is already cheaper by 90-95% from most western countries. This AI is replacing the work of 10-20 people.
There was a thread here about why ycombinator invests into several competing startups. The answer is success is often more about connections and politics than the product itself. And crypto, yes, is a good example of this. Musk will get his $1B in bitcoins back for sure.
> Most recent example was funneling money from Russia into Trump’s campaign.
Musk again?
It is a registration wall I think.
To be a bit acerbic, and inspired by Arthur C. Clarke, I might say: "Any sufficiently complex business could be indistinguishable from Theranos".
Oh yes.
I had a discussion with a manager at a client last week and was trying to run him through some (technical) issues relating to challenges an important project faces.
His immediate response was that maybe we should just let ChatGPT help us decide the best option. I had to bite my tongue.
OTOH, I'm more and more convinced that ChatGPT will replace managers long before it replaces technical staff.
That is one spicy article, it got a few laughs out of me. I must agree 100% that Langchain is an abomination, both their APIs as well as their marketing.
I have been working on FPGA's and, in general, programmable logic, for somewhere around thirty years (started with Intel programmable logic chips like the 5C090 [0] for real time video processing circuits.
I completely skipped over the whole High Level Synthesis (HLS) era that tried to use C, etc. for FPGA design. I stuck with Verilog and developed custom tools to speed-up my work. My logic was simple: If you try to pound a square peg into a round hole, you might get it done yet, the result will be a mess.
FPGA development is hardware development. Not software. If you cannot design digital circuits to begin with, no amount of help from a C-to-Verilog tool is going to get you the kind of performance (both in terms of time and resources) that a hardware designer can squeeze out of the chip.
This is not very different from using a language like Python vs. C or C++ to write software. Python "democratizes" software development at a cost of 70x slower performance and 70x greater energy consumption. Sure, there are places where Python makes sense. I'll admit that much.
Going back to FPGA circuit design, the issue likely has to do with the type, content and approach to training. Once again, the output isn't software; the end product isn't software.
I have been looking into applying my experience in FPGA's across the entire modern AI landscape. I have a number of ideas, none well-formed enough to even begin to consider launching a startup in the sector. Before I do that I need to run through lots of experiments to understand how to approach it.
[0] https://www.cpu-galaxy.at/cpu/ram%20rom%20eprom/other_intel_...
Returning to LLMs. I think the problem here may be that there is simply not enough learning material for LLM. Verilog comparing to C is a niche with little documentation and even less open source code. If open hw were more popular I think LLMs could learn to write better Verilog code. Maybe the key is to persuade hardware companies to share their closed source code to teach LLM for the industry benefit?
Current LLMs can’t do it, but the assumption that that’s what YC meant seems wildly premature.
The purpose of capital is to make progress from where we are now.
https://optics.ansys.com/hc/en-us/articles/360042305274-Inve...
https://optics.ansys.com/hc/en-us/articles/33690448941587-In...
YC did well because they were good at picking ideas, not generating them.
This doesn't line up with the perennial attitude (as discussed by pg) that YC picks people/teams and not ideas, because while ideas and approaches may change, the people are the same and having a good founder, co-founder and team matters the most.
Their M.O. is to avoid getting too attached to an idea because, in the process of actually building the company, pivots may be required. And so the focus is on a team moreso than a business plan, which again, is not something pg is particularly fond of seeing especially the ones that have lengthy (and therefore improbable/unrealistic) forecasts.
In my opinion, part of the problem i that training data is scarce (real world designs are literally called "IP" in the industry after all...), but more than that, circuit design is basically program synthesis, which means it's _hard_. Even if you try to be clever, dealing with graphs and designing discrete objects involves many APX-hard/APX-complete problems, which is _FUN_ on the one had, but also means it's tricky to just scale through, if the object you are trying to do is a design that can cost millions if there's a bug...
But selling shovels that are useful in many small markets can still be a viable play, and that’s how I understand YC’s position here.
Language is cool and immensely useful. LLMs, however, are fundamentally flawed from their basic assumptions about how language works. The distribution hypothesis is good for paraphrasing and summarization, but pretty atrocious for real reasoning. The concept of an idea living in a semantic "space" is incompatible with simple vector spaces, and we are starting to see this actually matter in minutia with scaling laws coming into play. Chip design is a great example of where we cannot rely on language alone to solve all our problems.
I hope to be proven wrong, but still not sold on AGI being within reach. We'll probably need some pretty significant advancements in large quantitative models, multi-modal models and smaller, composable models of all types before we see AGI
With respect to AGI in its broadest sense: indeed it is not in reach. I think that is for the better!
Other intelligent effects are coincidental.
Other intelligent effects are coincidental.
> If Gary Tan and YC believe that LLMs will be able to design chips 100x better than humans currently can, they’re significantly underestimating the difficulty of chip design, and the expertise of chip designers.
This is very obviously not the intent of the passage the author quotes. They are clearly talking about the speedup that can be gained from ASICs for a specific workload, eg dedicated mining chips.
> High-level synthesis, or HLS, was born in 1998, when Forte Design Systems was founded
This sort of historical argument is akin to arguing “AI was bad in the 90s, look at Eliza”. So what? LLMs are orders of magnitude more capable now.
> Ultimately, while HLS makes designers more productive, it reduces the performance of the designs they make. And if you’re designing high-value chips in a crowded market, like AI accelerators, performance is one of the major metrics you’re expected to compete on.
This is the crux of the author's misunderstanding.
Here is the basic economics explanation: creating an ASIC for a specific use is normally cost-prohibitive because the cost of the inputs (chip design) is much higher than the outputs (performance gains) are worth.
If you can make ASIC design cheaper on the margin, and even if the designs are inferior to what an expert human could create, then you can unlock a lot of value. Think of all the places an ASIC could add value if the design was 10x or 100x cheaper, even if the perf gains were reduced from 100x to 10x.
The analogous argument is “LLMs make it easier for non-programmers to author web apps. The code quality is clearly worse than what a software engineer would produce but the benefits massively outweigh, as many domain experts can now author their own web apps where it wouldn’t be cost-effective to hire a software engineer.”
What is the quality of Verilog code output by humans? Is it good enough so that a complex AI chip can be created? Or does the human need to use tools in order to generate this code?
I've got the feeling that LLMs will be capable of doing everything a human can do, in terms of thinking. There shouldn't be an expectation that an LLM is able to do everything, which in this context would be thinking about the chip and creating the final files in a single pass and without external help. And with external help I don't mean us humans, but tools which are specialized and also generate some additional data (like embeddings) which the LLM (or another LLM) can use in the next pass to evaluate the design. And if we humans have spent enough time in creating these additional tools, there will come a time when LLMs will also be able to create improved versions of them.
I mean, when I once randomly checked the content of a file in The Pile, I found an Craigslist "ad" for an escort offering her services. No chip-generating AI does need to have this in its parameters in order to do its job. So there is a lot of room for improvement and this improvement will come over time. Such an LLM doesn't need to know that much about humans.
Whoever is recommending investing in better chip(ALU) design hasn't done even a basic analysis of the problem.
Tokens per second = memory bandwidth divided by model size.
First, I agree that the bar for HLS tools is relatively low, and they are not as good as they could be. Admittedly, there has been significant progress in the academic community to develop open-source HLS tools and integrations with existing tools like Vitis HLS to improve the HLS development workflow. Unfortunately, substantial changes are largely in the hands of companies like Xilinx, Intel, Siemens, Microchip, MathWorks (yes, even Matlab has an HLS tool), and others that produce the "big-name" HLS tools. That said, academia has not given up, and there is considerable ongoing HLS tooling research with collaborations between academia and industry. I hope that one day, some lab will say "enough is enough" and create a open-source, modular HLS compiler in Rust that is easy to extend and contribute to—but that is my personal pipe dream. However, projects like BambuHLS, Dynamatic, MLIR+CIRCT, and XLS (if Google would release more of their hardware design research and tooling) give me some hope.
When it comes to actually using HLS to build hardware designs, I usually suggest it as a first-pass solution to quickly prototype designs for accelerating domain-specific applications. It provides a prototype that is often much faster or more power-efficient than a CPU or GPU solution, which you can implement on an FPGA as proof that a new architectural change has an advantage in a given domain (genomics, high-energy physics, etc.). In this context, it is a great tool for academic researchers. I agree that companies producing cutting-edge chips are probably not using HLS for the majority of their designs. Still, HLS has its niche in FPGA and ASIC design (with Siemens's Catapult being a popular option for ASIC flows). However, the gap between an initial, naive HLS design implementation and one refined by someone with expert HLS knowledge is enormous. This gap is why many of us in academia view the claim that "HLS allows software developers to do hardware development" as somewhat moot (albeit still debatable—there is ongoing work on new DSLs and abstractions for HLS tooling which are quite slick and promising). Because of this gap, unless you have team members or grad students familiar with optimizing and rewriting designs to fully exploit HLS benefits while avoiding the tools' quirks and bugs, you won't see substantial performance gains. Al that to say, I don't think it is fair to comply write off HLS as a lost cause or not sucesfull.
Regarding LLMs for Verilog generation and verification, there's an important point missing from the article that I've been considering since around 2020 when the LLM-for-chip-design trend began. A significant divide exists between the capabilities of commercial companies and academia/individuals in leveraging LLMs for hardware design. For example, Nvidia released ChipNeMo, an LLM trained on their internal data, including HDL, tool scripts, and issue/project/QA tracking. This gives Nvidia a considerable advantage over smaller models trained in academia, which have much more limited data in terms of quantity, quality, and diversity. It's frustrating to see companies like Nvidia presenting their LLM research at academic conferences without contributing back meaningful technology or data to the community. While I understand they can't share customer data and must protect their business interests, these closed research efforts and closed collaborations they have with academic groups hinder broader progress and open research. This trend isn't unique to Nvidia; other companies follow similar practices.
On a more optimistic note, there are now strong efforts within the academic community to tackle these problems independently. These efforts include creating high-quality, diverse hardware design datasets for various LLM tasks and training models to perform better on a wider range of HLS-related tasks. As mentioned in the article, there is also exciting work connecting LLMs with the tools themselves, such as using tool feedback to correct design errors and moving towards even more complex and innovative workflows. These include in-the-loop verification, hierarchical generation, and ML-based performance estimation to enable rapid iteration on designs and debugging with a human in the loop. This is one area I'm actively working on, both at the HDL and HLS levels, so I admit my bias toward this direction.
For more references on the latest research in this area, check out the proceedings from the LLM-Aided Design Workshop (now evolving into a conference, ICLAD: https://iclad.ai/), as well as the MLCAD conference (https://mlcad.org/symposium/2024/). Established EDA conferences like DAC and ICCAD have also included sessions and tracks on these topics in recent years. All of this falls within the broader scope of generative AI, which remains a smaller subset of the larger ML4EDA and deep learning for chip design community. However, LLM-aided design research is beginning to break out into its own distinct field, covering a wider range of topics such as LLM-aided design for manufacturing, quantum computing, and biology—areas that the ICLAD conference aims to expand on in future years.
I mean I assume the best is heavily guarded.
the AI hype train is basically investors not understanding tech, don’t get me wrong AI in itself could be a huge thing if used right but the things getting the most attention in the current market aren’t it
Would appreciate the collective energy being spent instead towards adding to amor refining Garry’s request.
As we've seen in the recent past, it's difficult to predict what the possibilities are for LLMS and what limitations will hold. Currently it seems pure scaling won't be enough, but I don't think we've reached the limits with synthetic data and reasoning.
Do we know what LLMs will be able to do in the future? And even if we know, the startups have to work with what they have now, until that future comes. The article states that there's not much to work with.
For a counterexample I think I’d look to non-tech companies. OrangeTheory maybe?
This idea of “we’re a startup; we can’t actually make anything useful now, but once the tech we use becomes magic any day now we might be able to make something!” is basically a new phenomenon.
It used to be that startups would be created to do something different with existing tech or to commercialise a newly-discovered - but real - innovation.
1) all the domains there is no training data
Many professions are far less digital than software, protect IP more, and are much more akin to an apprenticeship system.
2) the adaptability of humans in learning vs any AI
Think about how many years we have been trying to train cars to drive, but humans do it with a 50 hours training course.
3) humans ability to innovate vs AIs ability to replicate
A lot of creative work is adaptation, but humans do far more than that in synthesizing different ideas to create completely new works. Could an LLM produce the 37th Marvel movie? Yes probably. Could an LLM create.. Inception? Probably not.
Yeah, blind hope and a bit of smoke and lighting.
> but I don't think we've reached the limits with synthetic data
Synthetic data, at least for visual stuff can, in some cases provide the majority of training data. For $work, we can have say 100k video sequences to train a model, they can then be fine tuned on say 2k real videos. That gets it to be slightly under the same quality as if it was train on pure real video.
So I'm not that hopeful that synthetic data will provide a breakthrough.
I think the current architecture of LLMs are the limitation. They are fundamentally a sequence machine and are not capable of short, or medium term learning. context windows kinda makes up for that, but it doesn't alter the starting state of the model.
I think that LLMs are plateauing, but I'm less confident that this necessarily means the capabilities we're using LLMs for right now will also plateau. That is to say it's distinctly possible that all the talent and money sloshing around right now will line up a new breakthrough architecture in time to keep capabilities marching forward at a good pace.
But if I had $100 million, and could bet $200 thousand that someone can make me billions on machine learning chip design or whatever, I'd probably entertain that bet. It's a numbers game.
Problem with this reasoning is twofold: start-ups will overfit to getting your money instead of creating real advances; competition amongst them will drive up the investment costs. Pretty much what has been happening.