Hacker News Reader

Comment by mlinsey | Hacker News Reader

GPT4 is expensive enough to train that most organizations with that high of a compute budget also have compliance departments and will not flat-out ignore a law like that. State actors overseas and the indie hackers with open-source models will not be deterred, but they are probably six months behind anyway. I don't think the six-month moratorium (as opposed to say, a three year moratorium) would be so impractical to implement, for that reason. But I'm also not sure what it would get us, the past six months of alignment work doesn't suggest six more months would solve the problem.

View on HN

Nobody stopping anything ever.

I predict we will have distributed training systems soon on peer to peer like network for collaboratively building free models to use, if we can all pitch in some GPUs we can get 100s of thousand of individuals working on training giant LLM models together. It’s an unsolved problem currently but should have a solution soon.

Additionally today we can build sub-$10K systems to do self hosted inference on 65B weights models.

I’m seriously confused why anybody thinks anyone will even remotely care about any kind of pause is asked for.

For sure everyone with computers and a reasonable level of skills are now looking at this field. The progress will be exponential and uncontrollable in my opinion. I’m fine with that and I also think there is absolutely nothing that can stop it anyway.

> I’m seriously confused why anybody thinks anyone will even remotely care about any kind of pause is asked for.

I can say that Google and Microsoft would definitely care, just as two specific instances of "anybody" that I have worked for and seen what their legal review process gating new product launches is like.

You seem much more optimistic that the community efforts are <6 months behind those companies than I am; I would be happy to be wrong here and see an open-source self-hosted LLM with better than GPT4 performance by September.

I’ve read OpenAI only used around 4000 high end GPUs ($30K each) for around 4 months to train GPT4. Yes there are some reinforcement tricks that are still unclear but input data is basically known.

I think people underestimate how many entities have the ability to easily acquire that amount of resources. It’s less than $200M to purchase that amount of hardware. Militaries spend $2-4M per missile. This is nothing for a lot of organizations.

We already have self hosted LLM inference that reaches gpt 3.5 turbo level.

Google and Microsoft are notoriously slow moving and probably less relevant than smaller faster operations like OpenAI.

The fabs for gpu production are easily controllable. So are the data centers where many of them are currently concentrated.

There you go. I was waiting for this comment.

Yep so global totalitarianism is the clear answer. Makes sense.

I’m sure nobody will resist.

That's a rather strong assumption about other state actors, in terms of their current progress, future progress and their ability to attract researchers who get frustrated by the pause.

After the 6-month pause is over, then what? Another 6 months? What is to be hoped to be achieved in those initial 6 months other than non-participating entities getting a leg up? If there is no guarantee that the 6-month pause will not become one year or a few years, then maybe some researchers will move to non-participating organizations/states in anticipation of an extended pause.

I’d also point out that just the discussion about governments regulating LLM research is likely triggering many to start making contingency plans already. Make sure we have the datasets. make sure we have all the code and research papers.

Even on the individual level as for me the discussion of such ideas has me thinking individually if I should take any defensive steps or not.

Reply on news.ycombinator.com