On this view, adversarial examples arise from gradient descent being "too smart", not "too dumb": the program is fine; if the test suite didn't imply the behavior we wanted, that's our problem.

Shouldn't we expect to see RL models trained purely on self play not to have these issues then?

My understanding is that even models trained primarily with self play, such as katago, are vulnurable to adversarial attacks. If RL models are vulnurable to the same type of adversarial attacks, isn't that evidence against this theory?

The amount of inference compute isn't baked-in at pretraining time, so there is no tradeoff.

This doesn't make sense to me.

In a subscription based model, for example, companies would want to provide users the strongest completions for the least amount of compute.

If they estimate customers in total will use 1 quadrillion tokens before the release of their next model, they have to decide how much of the compute they are going to be dedicating to training versus inference. As one changes the parameters (subscription price, anticipated users, fixed costs for a training run, etc.) you'd expect to find the optimal ratio to change.

Test-time compute on one trace comes with a recommendation to cap reasoning tokens at 25K, so there might be 1-2 orders of magnitude more there with better context lengths. They are still not offering repeated sampling filtered by consensus or a reward model. If o1 proves sufficiently popular given its price, they might offer even more expensive options.

Thanks, this is a really good find!

Thanks!! this is exactly what I was looking for

With the release of openAI o1, I want to ask a question I've been wondering about for a few months.

Like the chinchilla paper, which estimated the optimal ratio of data to compute, are there any similar estimates for the optimal ratio of compute to spend on inference vs training?

In the release they show this chart:

The chart somewhat gets at what I want to know, but doesn't answer it completely. How much additional inference compute would I need a 1e25 o1-like model to perform as well as a one shotted 1e26?

Additionally, for some x number of queries, what is the optimal ratio of compute to spend on training versus inference? How does that change for different values of x?

Are there any public attempts at estimating this stuff? If so, where can I read about it?

If someone wants to set up a figgy group to play, I'd love to join

I agree the conclusion isn't great!

Not so surprisingly, many people read the last section as an endorsement of some version of "RCTism", but it's not actually a view I endorse myself.

What I really wanted to get at in this post was just how pervasive priors are, and how difficult it is to see past them.

Just played through it tonight. This was my first D&D.Sci, found it quite difficult and learned a a few things while working on it.

Initially I tried to figure out the best counters and found a few patterns (flamethrowers were especially good against certain units). I then tried to look and adjust for any chronology, but after tinkering around for a while without getting anywhere I gave up on that. Eventually I just went with a pretty brainless ML approach.

I ended up sending squads for 5 and 6 which managed a 13.89% and 53.15% chance of surviving, I think it's good I'm not in charge of any soldiers in real life!

Overall I had good fun, and I'm looking forward to looking at the next one.

This wouldn't be the first time Deepmind pulled these shenanigans.

My impression of Deepmind is they like playing up the impressiveness of their achievements to give an impression of having 'solved' some issue, never saying anything technically false, while suspiciously leaving out relevant information and failing to do obvious tests of their models which would reveal a less impressive achievement.

For Alphastar they claimed 'grandmaster' level, but didn't show any easily available stats which would make it possible to verify.  As someone who was in Grandmaster league at the time of it playing (might even have run into it on ladder, some of my teammates did), its play at best felt like low grandmaster to me.

At their event showing an earlier prototype off, they had one player (TLO) play their off-race with which he certainly was not at a grandmaster level. The pro player (Mana) playing their main race beat it at the event, when they had it play with the same limited camera access humans have. I don't remember all the details anymore, but I remember being continuously annoyed by suspicious omission after suspicious omission.

What annoys me most is that this still was a wildly impressive achievement! Just state in the paper: "we managed to reach grandmaster with one out of three factions" - Nobody has ever managed to create AI that played remotely as well as this!

Similarly Deepminds no-search chess engine is surely the furthest anyone has gotten without search. Even if it didn't quite make grandmaster, just say so!

if it makes it easier, I can add the questions to manifold if you provide a list of questions and resolution criteria.

thanks for pointing that out, I've added a note in the description