Basically - "any model trained with ~28M H100 hours, which is around $50M USD or - any cluster with 10^20 FLOPs, which is around 50,000 H100s, which only two companies currently have " - hat-tip to nearcyan on Twitter for this calculation.
Specific language below.
" (i) any model that was trained using a quantity of computing power greater than 1026 integer or floating-point operations, or using primarily biological sequence data and using a quantity of computing power greater than 1023 integer or floating-point operations; and
(ii) any computing cluster that has a set of machines physically co-located in a single datacenter, transitively connected by data center networking of over 100 Gbit/s, and having a theoretical maximum computing capacity of 1020 integer or floating-point operations per second for training AI."
I don’t know how big 10^20 floating points is, and if 70b was made with something bigger or smaller. But I think that figure is the more important one as I think Meta uses a single datacentre.
These figures in context:
(b) The Secretary of Commerce, in consultation with the Secretary of State, the Secretary of Defense, the Secretary of Energy, and the Director of National Intelligence, shall define, and thereafter update as needed on a regular basis, the set of technical conditions for models and computing clusters that would be subject to the reporting requirements of subsection 4.2(a) of this section. Until such technical conditions are defined, the Secretary shall require compliance with these reporting requirements for:
(i) any model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, or using primarily biological sequence data and using a quantity of computing power greater than 10^(23) integer or floating-point operations; an
(ii) any computing cluster that has a set of machines physically co-located in a single datacenter, transitively connected by data center networking of over 100 Gbit/s, and having a theoretical maximum computing capacity of 10^(20) integer or floating-point operations per second for training AI.
Assuming number of FLOPs in compute is 6ND (N = number of parameters, D = dataset size in tokens) you could take the full RedPajama dataset (30T tokens) and a 500B parameter model and it’d come out to:
6*(30*10^12)*(500*10^9) = 9*10^25
In order to qualify, you would need a cluster that could train this beast in about:
10^26 / 10^20 = 1000000 seconds = 11.57 days
Ok, as a baseline for everyone who, like me, doesn’t understand all the big words and numbers on why this is great news:
So, if I’m understanding correctly, one of our most powerful open source models is so far from this benchmark that it can’t even been seen.
Someone please correct me if I’m wrong.
Someone please correct me if I’m wrong.
Think of it like regulating all use of 50Mhz+ computers, back in the early 80s when most people had 5Mhz or less. At the time, you might have thought “OK, I’ll never be able to afford that anyway – that’s like Space Shuttle computing power.” Yet, with such a restriction, this timeline, where everyone has smartphones and smartwatches and smart TVs, self-driving cars, robots, and millions of servers combine to create the internet, would not exist.
I imagine creating an app, putting it on everyone’s cell phone, and using a fraction of the power, you can build an llm easily that would surpass any single data center.
You have the connection speed between phones to worry about, as well as a different architecture. There’s a big difference running the kernel over a new layer and its inputs locally within a GPU chip, vs. copying that data to into packets, filling in all of the rest of the information associated with the packets, sending it to the phone’s radio, having it turned into radio waves, transmitting that to a cell tower, routing it through the network to the cell co, routing it on to the receiving phone’s cell tower (maybe via a satellite or two), transmitting it to the destination phone, decoding the radio waves, etc. I’m deliberately leaving out some details (like the bsd socket layers and encryption and decryption), and I’m sure I’m missing many other complications.
BUT, it’s conceivable, in future, as tech improves and the gap between consumer hardware and what’s needed to run AGI narrows , and so on.
The numbers appear to have OpenAI’s finger-prints on them. I don’t know if they’re from an AI-risk mitigations perspective or for laying foundations for competitive barriers. Probably a mix of both.
At 30 trillion tokens, 10^26 float ops caps you at ~550 billion parameters (using float ops = 6 * N * D). Does this indirectly leak anything about OpenAI’s current scaling? At 10 trillion tokens, it’s 1.7 Trillion parameters. Bigger vocabularies can stretch this limit a bit.
They must be prepping the field for tomorrow rather than trying to introduce immediate trust market conditions.
https://www.youtube.com/watch?v=8K6-cEAJZlE&t=6m39s
Where did it start? It started right here. And this is where it could’ve been stopped! If those people had stood together. If they had protected each other, they could’ve resisted the Nazi threat. Together they would’ve been strong. But once they allowed themselves to be split apart, they were helpless. When that first minority lost out, everybody lost out.
“Give me a big number in units that will be very hard to understand by anybody.”
“28M pigeon feet”
“It’s too on the nose.”
“28M H100 hours”
If I am not mistaken “28M H100 hours” roughly equals to “87M tetryliodo hexamine” or “32M hydrocoptic block rounds” given by the equation P = 2.5 times C times n to the 6th power, minus 7.
All of this is a red herring. The bigger issue is going to be checking of the data for biological sequences and such.
Sequences? Why?
Because they’re very concerned about using LLMs for help in creating bioweapons, and a small portion of the data will go a long way. I believe this will lead to scrutinizing datasets.
OHHHH that’s what that’s about. Makes sense.
Recombining elements of existing pathogens or chemicals using non-AI modelling is what current biolabs already do - and they still need to make them, test them, because all modelling gets you is good guesses. If anything my guess is that LLM’s will be worse at that task than human expert plus non-AI modelling. Still I guess I get the caution.
Haha great, open source models will now have a chance, and China will be able to catch up with their hypercensorship models
People say we should use the government to crack down on monopolies. But it’s the government that owns the monopolies.
It’s the other way around.
Its not a mutually exclusive thing. Its very obviously true that unregulated markets devolve into extreme anti competitive practices quickly. It’s also very obviously true large corporations can reach their hand into the government and “encourage” extreme anti competitive practices for their favor.
The legislation is going to get old, fast.
The legislation is going to get old, fast.
It’s an Executive Order, so they can just, like, make another one. It’s not legislation that has to pass the House and Senate.
1026
1023
1020
That would include models trained on a calculator.
So if I make a 10 Trillion param mixture of experts model out of fine-tuned variations of the same 300b model I am safe right?
Or how about I train a 4 Trillion param model on a new architecture that can utilize distributed compute? If contribution to GPU pools for training is encrypted and decentralized then good luck.
Fuck OpenAI. We will take this fight to the streets and win.
huh what’s that about biological sequence data , a MUCH lower number , huh , um do they know something specific about how that’s dangerous :o
just train/finetune outside of the us
Some exponent marks missing here.
The most interesting part is the focus on biological sequence data. This means that generative AI for synthetic biology is on the policy makers/risk assessors radar, and probably rightly so.