Imagine if someone was lobbying for some reasonable regulation (we should regulate drugs, based around clinical trials) and then instead of a transparent system you get purely executive actions with little to no public justification (Trump declares all glp1s illegal no one knows why exactly)
Would you levy the same two quote criticism of the reasonable call for regulation?
There should be an anti leaderboard that highlight people under a threshold. Not trying to learn how to use ai while working at a company like Amazon is almost certainly a bad thing, and cause for looking into why.
My take: multi turn evals are hard because to do it really correctly you have to simulate a user. This is not yet modeled well enough for multi turn to work as well as it could.
The flip side of this is that for some tasks (especially in ml/ai), doing it manually at least a few times gives you a sense of what is correct and a better sense of detail.
For example, spending the time to label a few examples yourself instead of just blindly sending it out to labeling.
(Not always the case, but another thing to keep in mind besides total time saved and value of learning)
I think the methods here are highly questionable, and appear to be based on self report from a small amount of employees in Denmark 1 year ago.
The overall rate of participation in the labor work force is falling. I expect this trend to continue as AI makes the economy more and more dynamic and sets a higher and higher bar for participation.
Overall GDP is rising while labor participation rate is falling. This clearly points to more productivity with fewer people participating. At this point one of the main factors is clearly technological advancement, and within that I believe if you were to make a survey of CEOS and ask what technological change has allowed them to get more done with fewer people, the resounding consensus would definitely be AI
I’m talking about a general trend I see in use of this term, not that it’s always a bad thing to say “I’m not technical so someone else should write the script”
I agree with everything you said!
Both things are happening in the world: people using this terminology to throw work at others needlessly, and people doing good division of labor.
I think that “I’m not technical” is often an excuse for throwing work at other people and frankly can be a form of learned helplessness. Nowadays, there is less and less reason to ask other people to write one off scripts/queries, you can ask AI for help and learn how to do that.
Since this is HN some disclaimers
-no that’s not always what’s happening, when “not technical” is thrown around
-no it’s not always appropriate to use AI instead of asking an expert
It seems like if they in fact distilled then what we have found is that you can create a worse copy of the model for ~5m dollars in compute by training on its outputs.
“In my humble opinion, these companies would not allocate a second of compute to lightweight models if they thought there was a straightforward way to achieve the next leap in reasoning capabilities.”
The rumour/reasoning I’ve heard is that most advances are being made on synthetic data experiments happening after post-training. It’s a lot easier and faster to iterate on these with smaller models.
Eventually a lot of these learnings/setups/synthetic data generation pipelines will be applied to larger models but it’s very unwieldy to experiment with the best approach using the largest model you could possibly train. You just get way fewer experiments per day done.
The models bigger labs are playing with seem to be converging to about what is small enough for a researcher to run an experiment overnight.
The talking point has been that if we do this in cases where there isn’t shelter to offer, the people will come back. Let’s see how that plays out, will be informative.
I think it’s also quite possible for some people it’s a needed wake up call
I agree with you but it also makes me think: Google's TPUs are also fixed costs and these research experiments could have been run at times when production serving need isn't as high.
I think numpy closely maps to how I think so it’s not as hard to read these dense lines as it would be to read expanded versions. I think my point of view is shared by a lot of leading researchers and this is why it is used more heavily.
The kinds of type safety you want might be good for other use cases but for ML research they get in the way too much.