I hope we find a path to at least fine-tuning medium sized models for prices that aren't outrageous. Even the tiny corp's tinybox [1] is $15k and I don't know how much actual work one could get done on it.
If the majority of startups are just "wrappers around OpenAI (et al.)" the reason is pretty obvious.
When I was at Rad AI we managed just fine. We took a big chunk of our seed round and used it to purchase our own cluster, which we setup at Colovore in Santa Clara. We had dozens, not hundreds, of GPUs and it set us back about half a million.
The one thing I can't stress enough- do not rent these machines. For the cost of renting a machine from AWS for 8 months you can own one of these machines and cover all of the datacenter costs- this basically makes it "free" from the eight month to three year mark. Once we decoupled our training from cloud prices we were able to do a lot more training and research. Maintenance of the machines is surprisingly easy, and they keep their value too since there's such a high demand for them.
I'd also argue that you don't need the H100s to get started. Most of our initial work was on much cheaper GPUs, with the A100s we purchased being reserved for training production models rapidly. What you need, and is far harder to get, is researchers who actually understand the models so they can improve the models themselves (rather than just compensating with more data and training). That was what really made the difference for Rad AI.
That said, a lot of other businesses don't want to take on the capex, but they do need to train some models... and those models can't run on just a half a million worth of hardware. In that case, someone else is going to have to do it for you.
It works both ways and there are no absolutes here.
My response was more for these folks the OP mentioned-
> There is almost no universe where a couple of guys in their garage are getting access to 1000+ H100s with a capital cost in the multiple millions.
I'm pointing out that this isn't true. I was the founding engineer at Rad AI- we had four people when we started. We managed to build LLMs that are in production today. If you've had a CT, MRI, or XRay in the last year there's a real chance your results were reviewed by the Rad AI models.
My point is simply that people are really overestimating the amount of hardware actually needed, as well as the costs to use that hardware. There absolutely is a space for people to jump in and build out LLM companies right now, and the don't need to build a datacenter or raise nine figures of funds to do it.
The guy could wake up tomorrow and decide he didn't feel like developing this stuff any more and you're going to be stuck with a dead project. In fact, he already did that once when he found a bug in the driver.
People RIP on Google for killing projects all the time and now you want to bet your business on a guy who livestreams in front of a pirate flag? Come on.
Never mind that even in my own personal dealings with him, he's been a total dick and I'm far from the only person who says that.
Even if I validate my idea on a RTX 4090, the path to scaling any idea gets expensive fast. 15k to move up to something like a tinybox (probably capable of running 65B model but is it realistic to train or fine-tune 65B model?). Then maybe $100k in cloud costs. Then maybe $500k in research sized cluster. Then $10m+ for enterprise grade. I don't see that kind of ramp happening outside well-financed VC startups.
That said I'm mostly responding to the "two guys in a garage" comment with this. Larger companies are going to have different needs altogether.
Serious question: Where does an aspiring AI/ML dev get that expertise. From looking at OMCS I'm not convinced even a doctorate from Georgia Tech would get me the background I need...
To put it another way, the $10m+ for enterprise grade just seems wrong to me. It's more like $10m+ for mediocre responses to a lot of things. Rad AI didn't spend $10m on their models, but they absolutely are professional grade and are in use today.
I also think it's important to consider capital costs that are a one time thing, versus long term costs. Once you purchase that $10m cluster you have that forever, not just for a single model, and because of the GPU scarcity right now that cluster isn't losing value nearly as rapidly as most hardware does. If you purchase a $500k cluster, use it for three years, and then sell it for $400k you're really not doing all that bad.
I can't tell you if one program is better than another, as it's a bit out of my area of expertise.
Another absolute. I try to not be so focused on single points of input like that.
From what I can tell, sitting on the other side of the wall (GPU provider), there is metric tons of demand from all sides.
But my (totally amateur and outsider informed) intuition is that the innovative work will still happen at the edge of model size for the next few years. We literally just got the breakthroughs in LLM capabilities around the 30b parameter mark. These capabilities seemed to accelerate with larger models. There appears to be a gulf in the capabilities from 7B to 70B parameter LLMs that makes me not want to bother with LLMs at all unless I can get that higher level performance of the massive models. But even if I did want to play around at 30B or whatever I have to pay 15k-100k.
I think we are just in a weird spot right now where the useful model sizes for a large class of potential applications is at a price point that many engineers will find prohibitively expensive to experiment with on their own.
We need more people who "think different" and push back against the status quo instead of carrying out ad hominem attacks on public forums.
Should I complain that to drill oil I need hundreds of millions of dollars to even start?
Your VPS example was doing barely any computation. You're conflating web 1.0 and web 2.0 with neural networks and they are nothing alike in terms of FLOPS.
I also think that you're way off on the second point. I'm not saying that to be rude, because it does seem to be a popular opinion. It's just that if you read papers most people publishing aren't using giant clusters. There's a whole field of people who are finding ways to shrink models down. Once we understand the models we can also optimize them. You see this happen in all sorts of fields beyond "general intelligence"- tasks that used to take entire clusters to run can work on your cell phone now. Optimization is important not just because it opens up more people to work on things, but also because it drops down the costs that these big companies are paying.
Lets think about this in another direction. ML models are based off of how the brain is thought to work. The human brain is capable of quite a bit, but it uses very little power: about 10 watts. It is clearly better optimized than ML models are. That means there's a huge gap we still have to fit on efficiency.
It's not that bad; there are lots of things you can do with a hobbyist budget. For example, a consumer GPU with 12 or 24 GB VRAM costs $1000-2000 and can let you run many models and do fine-tuning on them. The next step up, for fine-tuning larger models, is to rent an instance on vast.ai or something similar for a few hours with a 4-8 GPU instance, which will set you back maybe $200—still within the range of a hobbyist budget. Many academic fine-tuning efforts, like Stanford Alpaca, cost a few hundred dollars to fine-tune. It's only when you want to pretrain a large language model from scratch that you need thousands of GPUs and millions in funding.
2) Speaking of VPSes and web 1.0 in the same breath is a little anachronistic. Servers had much lower capacity in 1999, and cost much more. Sun was a billion dollar company during the bubble because it was selling tens of thousands of unix servers to startups in order to handle the traffic load. Google got a lot of press because they were the oddballs who ran on commodity x86 hardware.
There is a massive difference between what is necessary to prove a scientific thesis and what is necessary to run a profitable business. And what do you mean "giant clusters" in this context. What is the average size of the clusters used in ground breaking papers and what is their cost? Is that cost a reasonable amount for a boot-strapped startup to experiment with or are we getting into the territory where only VC backed ventures can even experiment?
> There's a whole field of people who are finding ways to shrink models down
Of course the cost of running models is going to come down. The literal article we are responding to is a major part of that equation. You seem to be making arguments about how the future will be as support for an argument against how the present is.
Presently, hardware costs are insanely high and not coming down soon (as per the article). Presently, useful models for a large set of potential applications require significant cluster sizes. That makes it presently difficult for many engineers to jump in and play around.
My opinion is that the cost has to come down to the point that hobbiest engineers can play with the high-quality LLMs at the model sizes that are most useful. That doesn't imply that there are no model sizes for other use-cases that can't be developed today. It doesn't imply that the price of the hardware and size of the models will not fall. It just implies that dreaming of a business based around a capable LLM means your realistic present day costs are in the 10's of thousands at a minimum.
Of course, it is the businesses that find a way to make this work that will succeed. It isn't an impossible problem, it is just a seemingly difficult one for now. That is why I mentioned VC funding as appearing to have more leverage over this market than previous ones. If you can find someone to foot the 250k+ cost (e.g. AI Grant [1] where they offer 250k cash and 350k cloud compute) then you might have a chance.
Maybe he'll succeed, but this definitely doesn't scream stability to me. I'd be wary of investing money into his ventures (but then I'm not a VC, so what do I know).
[1] https://www.youtube.com/watch?v=Mr0rWJhv9jU
[2] https://github.com/RadeonOpenCompute/ROCm/issues/2198#issuec...
[3] https://twitter.com/realGeorgeHotz/status/166980346408248934...
If you want to compete on the actual model, then yes, this is not the time for garage shops.
If your business plan is so good, then it will work without H100 "cards" too, or if it's even better and you know it'll print money with H100 cards then great, just wait.
By the way, I also got a bug in the AMD drivers fixed too [0]. That bug fix enabled me to fully automate the performance tuning of 150,000 AMD gpus that I was managing. This is something nobody else had done before, it was impossible to do without this bug fix. We were doing this by hand before! The only bummer was that I had to upgrade the kernel on 12k+ systems... that took a while.
I went through the proper channels and they fixed it in a week, no need for a public meltdown or email to Lisa crying for help.
[0] https://patchwork.freedesktop.org/patch/470297/?series=99134...
These skills are developed when they study phd