There is basically the big guys that everyone's heard of (google, meta, microsoft/openAI, and anthropic) and then a handful of smaller players who are training foundation models mostly so that they can prove to VCs that they are capable of doing so -- to acquire more funding/access to compute so that they may eventually dethrone openAI and take a piece of the multi-billion dollar "enterprise AI" market for themselves.
Below that, there is a frothing ocean of mostly 7B finetunes created mostly by individuals who want to jailbreak base models for... reasons, plus the occasional research group.
The most oddball one I have seen is the databricks LLM which seems to have been an exercise of pure marketing. Those I suspect will disappear when the bubble deflates a bit.
Snowflake and the like are training and releasing new models because they intend to integrate the AI into their existing product down the line. Why not use and fine-tune an existing model? Their in-grown model maybe better suited for their product. This can also fail like Bloomberg's financial model being inferior to GPT-4, but these companies have to try.
Not all of them have permissive licenses for whatever the companies may want (or their clients want). Kind of a funny situation where everyone would benefit, but no one wants to burn their money for the greater good.
Interesting you'd say that in a discussion on Snowflake's LLM, no less. As someone who has a good opinion of Databricks, genuinely curious what made you arrive at such a damning conclusion.
Yes. Great choice of words. A lot of non-frontier models look like "an exercise of pure marketing" to me.
Still, I fail to see the rationale for telling the world, "Look at us! We can do it too!"
There are DOZENS of orgs releasing foundational models, not "a handful."
Salesforce, EleuthierAI, NVIDIA, Amazon, Stanford, RedPajama, Cohere, Mistral, MosaicML, Yandex, Huawei StabilityLM, ...
https://docs.google.com/spreadsheets/d/1kT4or6b0Fedd-W_jMwYp...
It's completely bonkers and a huge waste of resources. Most of them will see barely any use at all.
Furthermore, I'm happy to be in a golden age with lots of orgs trying things and many options. It's going to suck once the market eventually consolidates us and we have to take whatever enshittified thing the ologopolists feed us.
I want someone to spend a million on a Chess LLM so we can get a sense of how sophisticated they can get at non-linguistic pattern matching.
I want someone to spend a million on an LLM trained on Python program traces so we can try to teach it cause and effect and "debugging". Maybe it will emulate a Python interpreter and get highly reliable at predicting the outcome of Python code.
etc.
In the grand scheme of things, though, most of these are quite small -- 7b range. A 7b model is nothing to sneeze at but it's not megacorp resources either. It's in the range of "VC check" size.
The "big boys" who are training 70b plus are FAANG or government-scale entities. Microsoft, Google, and Meta have multiple entries on that "big" LLM foundation list -- it's because the GPUs are already bought, have to train something to keep utilization up. Also bear in mind that training of these things is still something closer to an art than a science; you put terabytes of data into the cauldron, let it brew, and only after it's done can you taste what you've made. Makes sense that some of these models will be junk.
Did you have a script to generate it or was it manually done?
A group at the CRFM maintains a bigger list of models (their goal is stated for cataloguing foundation models, but it looks like they have some tunes mixed in these days): https://crfm.stanford.edu/ecosystem-graphs/
This site also seems to keep track of models, with more closed/announced models that I don't bother to track: https://lifearchitect.ai/models-table/
A big one is that it shows investors, partners, and future recruits that you are both willing and capable to work on frontier technology. Hard to put a price on this, but it is important.
For the rest of us, it turns out you can use this bestiary of public models, mixing pieces of models with their own secret sauce together to create something superior than any of them [1].
I am speculating here, they would use their own OSS models to create a proprietary version which does one thing well. Answering questions for customers based on their own data. It's not an easy problem to solve as it seemed initially given enterprises need high reliability. Need models which are good at tool use, and can be grounded well. They could have done it on an oss model, but only now we have Llama-3 which is trained to make tool use easy. (Tool use as in function calling and use of stuff like OpenAI's code interpreter)
[1]: https://www.databricks.com/product/data-intelligence-platfor...
And streamlit, if you're as old as me, looks an awful lot like a MS-Access application for today. Again, it lives in the database, runs on a Snowflake warehouse and consumes credits, which is their revenue engine.
> offers over 0.6 million different pretrained open models.
One estimate I saw was that training GPT3 released 500 tons of CO2 back in 2020. Out of those 600k models, at least hundreds are of a comparable complexity. I can only hope building large models does not become analogous to cryptocoin speculation, where resources are forever burned only in a quest to attract the greater fool.
Those startups and researchers would better invest in smarter algorithms and approaches instead of trying to outpolute OpenAI, Meta and Microsoft.
500t to train a model at least seems like a more productive use of carbon than spending a few days on the beach. So I don’t think the carbon use of training models is that extreme.
If it's just a handfull of companies doing it, fine, it's negligible versus benefits. If it starts to chase the marginal cost of the resources in requires, so that every mid to large company feels that a few million $ or so spent training their a model on their own dataset makes them more in competitive advantages, then it quickly spirals out of control hence the cryptocoin analogy. That's exactly what many AI startups are proposing.
Just do them when renewable supply is high and demand is low; that energy can't be stored and would have been wasted anyway.
Again, look at bitcoin mining, the miners will happily pay any carbon tax to work 24/7, it's better to run the farm to cover electricity prices and then make some pennies then to keep it off and incur depreciation costs.
In addition, even 100% renewable would be attainable, that would still require massive infrastructure investment, resource use and associated emissions, since most of the corresponding industries, such as concrete and steel production, aluminum and copper ore mining and refining etc. are very far from net zero and will stay that way for decades.
To throw into this planet-sized bonfire a large uninterruptible consumer, whose standby capital depreciation on things like state of the art datacenters far exceeds the cost most industries are willing to pay for energy, all predicated on the idea that "demand spurs renewable investments", is frankly idiotic.
There aren't that many base models, and finetunes take very little energy to perform.
It's likely not the model size that's bigger, but the training corpus (see 15T for llama3). I doubt anyone has a model with “trillions” of parameters right now, one trillion maybe as rumored for GPT-4, but even for GPT-4 I'm skeptical about the rumors given the inference cost for super large models and the fact that the biggest lesson we got since llama is that training corpus size alone is enough for performance increase, at a reduce inference cost.
Edit: that doesn't change your underlying argument though: no matter if it's the parameter count that increases while staying at “Chinchilla optimal” level of training, or the training time that increases, there's still a massive increase in training power spent.
https://www.newsweek.com/taylor-swift-coming-under-fire-co2-...
I’m not sure if any of these things would be the first on the chopping block if a carbon tax were implemented, but it is worth a shot.
So absolute nothing in the grand scheme of things?
Done once, I agree, that's very little.
But if each of those 600,000 other models used that much (or even a tenth that much), then that now becomes impactful.
Releasing 500 tons of CO2 600,000 times over would amount to about 1% of all human global annual emissions.
And those 600k models are mostly fine-tunes. If running your 4090 at home is too much then we're going to have to get rid of the gamers.
This CO2 objection is such an innumerate objection. Just making 100 cars already is more than making one of these LLMs from scratch. A finetune is so cheap in comparison.
In fact, I bet if you asked most LLM companies they'd gladly support a universal carbon tax with even dividend based on emissions and then you'd see who's actually emitting.
One sees the high impact of the large model, and the growth of model training, and is concerned with how much that could increase in coming years.
The other group assumes the first group is complaining about right now, and thinks they're being ridiculous.
This whole thing reminds me of ten years ago when people were pointing out energy waste as a downside of bitcoin. "It's so little! Electricity prices will prevent it from ever becoming significant!" was the response that it was met with, just like people are saying in this thread.
In 2023, crypto mining accounted for about 0.5% of humanity's electricity consumption. If AI model training follows a similar curve, then it's reasonable to be concerned.
Except this is obviously not the case, as "the other group" is aware that many of these large training companies, such as Microsoft, have committed to being net negative on carbon by 2030, and are actively making progress with this whereas the other group seems to be motivated by flailing for anything they can use to point at AI and call it bad.
How many carbon-equivalent tons does training an AI in a net negative datacenter produce? Once the datacenters run on sunlight what is the new objection which will be found?
The rest of the world does not remain static with only the AI investments increasing.
Are you claiming that by 2030, the majority of AI will be trained in a carbon-neutral-or-better environment?
If not, then my point stands.
If so, I think that's an unrealistic claim. I'm willing to put my money where my mouth is. I'll bet you $1000 that by the year 2030, fewer than half of (major, trailed-from-scratch) models are trained in a carbon-neutral-or-better environment. Money goes to charity of the winner's choice.
Mistral appears to be using the Leonardo supercomputer, which doesn't seem to have direct numbers available, but I did find this quote upon its launch in 2022:
> One of the most powerful supercomputers in the world – and definitely Europe’s largest – was recently unveiled in Bologna, Italy. Powerful machine Leonardo (which aptly means “lion-hearted”, and is also the name of the famous Italian artist, engineer and scientist Leonardo da Vinci) is a €120 million system that promises to utilise artificial intelligence to undertake “unprecedented research”, according to the European Commission. Plus, the system is sustainably-focused, and equipped with tools to enable a dynamical adjustment of power consumption. It also uses a water-cooling system for increased energy efficiency.
You might have a greater chance to win the bet if we think about all models trained in 2030, not just flagship/cutting-edge models, as it's likely that all the GPUs which are frantically being purchased now will be depreciated and sold to hackers by the truckload here in 4-5 years, the same way some of us collect old servers from 2018ish now. But even that is a hard calculation to make--do we count old H100s running at home but on solar power as sustainable? Will the new hardware running in sustainable datacenters continue to vastly outpace the old depreciated?
For cutting-edge models which almost by definition require huge compute infrastructure, a majority of them will be carbon neutral by 2030.
A better way to frame this bet might be to consider it in percentages of total energy generation? It might be easier to actually get that number in 2030. Like Dirty AI takes 3% of total generation and clean AI 3.5%?
Something else to consider is the algorithmic improvements between now and 2030. From Yann LeCunn: Training LLaMA 13B emits 24 times less greenhouse gases than training GPT-3 175B yet performs better on benchmarks.
I haven't done longbets before, but I think that's what we're supposed to use for stuff like this? :) My email is in my profile.
One more thing to consider before we commit is that the current global share of renewable energy is something close to 29%. You should probably factor in overall renewable growth by 2030, if >50% of energy is renewable by then, I win by default but that doesn't exactly seem sporting.
Hmm. Yeah, we'll need to hammer out a solid definition. Further complicating things are models that may not be publicly available and are internally used by companies, though those may not be trained from scratch.
I would be fine with your suggestion to frame it in terms of percent power generation, though it might be hard to disentangle training costs from usage costs from that number. I would argue that including usage energy cleanliness is in the "spirit" of the bet but I'm happy to try to disentangle it as I originally proposed training-only.
> Something else to consider is the algorithmic improvements between now and 2030. From Yann LeCunn: Training LLaMA 13B emits 24 times less greenhouse gases than training GPT-3 175B yet performs better on benchmarks.
This is an excellent point, and definitely works in your favor. Really I'm on the good side of this bet. I win either way :) either my charity makes money, or I'm pleasantly surprised by climate impacts.
I've also not used longbets before. I would think we want to hammer out exact terms here before we set something up there?
Yes, but one can at least still imagine scenarios where AI training being 0.5% of electricity use could still be a net win.
(I hope we're more efficient than that; but if we're training models that end up helping a little with humanity's great problems, using 1/200th of our electricity for it could be worth it).
It does not seem well-poised to yield novel advancements in unrelated-to-AI fields, yet. Possibly genetics. But things like solving global warming, there is not any sort of path towards that for anything we're currently creating.
It's not clear to me that spending 0.5% of electricity generation to put a solid chunk of the lower-middle-class out of work is worth it.
Can LLMs help us save energy? It doesn't seem to be such a ridiculous idea to me.
And can they be an effort multiplier for others working on harder problems? Likely-- I am a high-skill worker and I routinely have lower-skill tasks that I can delegate to LLMs more easily than I could either do myself or delegate to other humans. (And, now and then, they're helpful for brainstorming in my chosen fields).
I had a big manual to write communicating how to use something I've built. Giving GPT-4 some bulleted lists and a sample of my writing got about 2/3rds of it done. (I had to throw a fraction away, and make some small correctness edits). It took much less of my time than working with a doc writer usually does and probably yielded a better result. In turn, I'm back to my high-value tasks sooner.
That is, LLMs may help attacking the great problems directly, or they may help us dedicate more effort to the great problems. (Or they may do nothing or may screw us all up in other ways).
The trouble I have is, what determines whether AI grow to 0.5% (or whatever %) of our electricity usage is not whether the AI is a net good for humanity even considering power use. It's going to be determined by whether the AI is a net benefit for the bank account of the people with the means to make AI.
We can just as easily have a situation where AI grows to 0.5% electricity usage, is economically viable for those in control of it, while having a net negative impact for the rest of society.
As a parent said, a carbon tax would address a lot of this and would be great for a lot of reasons.
Snowflake is a publicly traded company with a market cap of $50B and $4B of cash in hand. It has no need for venture capital money.
It looks like a case of "Look Ma! I can do it too!"
These fine tunes are not a huge amount of compute, most of them are doing these trainings on a single personal machine over a day or so of effort, NOT the six+ months across a massive cluster it takes to make a good base model.
That isn't wasted effort either. We need to know how to use these tools effectively, they're not going away. It's a very reductionist and inaccurate view of the world you're peddling in that comment.
However, in the long-term, as the hype dies down, so will the stock prices.
At the end of the day, I think it will be a transfer of wealth from shareholders to Nvidia and power companies.
*Depending on govt interventions
Honestly, Zuckerburg (seemingly the only CEO willing to actually invest in an open AI ecosystem for the obvious benefits it brings them) should just invest a few million into hiring a few real firmware hackers to port all the ML CUDA code into an agnostic layer that AMD can build to.
Which is to say, "everyone" knows that this stuff has a lot of potential. Everyone is also used to what often happens in tech, which is outrageous winner-take-all scale effects. Everyone ALSO knows that there's almost certainly little MARGINAL difference between what the big guys will be able to do and and what the little guys can do on their own ESPECIALLY if they essentially 'pool their knowledge.'
So, I suppose it's the whole industry collectively and subconsciously preventing e.g. OpenAI/ChatGPT becoming the Microsoft of AI.
This seems rather generous.
More like "little guys, or even literally any 'guy' that isn't dominant in this space -- which tends toward dominance -- have learned, perhaps counterintuitively, that free/open source is best for their own greedy interests."
:)
What a peculiar way to say: 600,000
The first droid armies will rapidly recoup the cost when the final wars for world domination begin…
2020's elections costed 15B USD in total, so we can't afford to lose (we are the good guys, right ?)
It’s mostly marketing for the company to appear to be modern. If you aren’t differentiated and if LLMs aren’t core to your business model, then there’s no loss from releasing weights. In other cases it is commoditizing something that would otherwise be valuable for competitors. But most of those 600K models aren’t high performers and don’t have large training budgets, and aren’t part of the “race”.
Hype and jumping on the bandwagon are perfectly good reasons for a business. There's no business without risk. This is the cost of doing business which you want to explore greenfield projects.
I heard this winner-takes-all spiel before - only last time, it was about Uber or Tesla[1] robo-taxis making car ownership obsolete. Uber has since exited the self-driving business, Cruise is on hold/unwinding and the whole self-driving bubble has mostly deflated, and most of the startups are long gone, despite the billions invested in the self-driving space. Waymo is the only company with robo-taxis, albeit in only 2 tiny markets and many years away from general availability.
1. Tesla is making robo-taxi noises once more, and again, to juice investor sentiment.
No investment in an emerging market is a sure thing, it's an educated guess. You have to take a lot of swings to occasionally hit a homerun, and investing in AI seems like the most plausible swing to make at this time.
I suspect the AI market will have a similar trajectory in the next decade: no actual AGI - maybe one company still plugging away at it, a couple of very successful companies whose core competencies don't include AI, but with billions in market cap, and a lot of failed startups littering the way there.
If the "best model" only stays the best for a few months and if, during those few months, the second best model is near indistinguishable, then it will be extremely hard to extract trillions of dollars.
It also means they build large scale rails necessary to use Snowflake for training and can market such at every release.
Having said that, I'm a big fan of Llama-3 at the moment.
https://www.snowflake.com/blog/arctic-open-efficient-foundat...
But the most interesting aspect about this, for me, is that every tech company seems to be coming out with a free open model claiming to be better than the others at this thing or that thing. The number of choices is overwhelming. As of right now, Huggingface is hosting over 600,000 different pretrained open models.
Lots of money has been forever burned training or finetuning all those open models. Even more money has been forever burned training or finetuning all the models that have not been publicly released. It's like a giant bonfire, with Nvidia supplying most of the (very expensive) chopped wood.
Who's going to recoup all that investment? When? How? What's the rationale for releasing all these models to the public? Do all these tech companies know something we don't? Why are they doing this?
---
EDIT: Changed "0.6 million" to "600,000," which seems clearer. Added "or finetuning".