nmca

You only need one, plus a couple recovery spares, in any sane implementation.

hardware tokens are the way! Everyone has had a house key their whole lives, and understands how to keep a spare to prevent lock-outs.

This would be much better with an example

Only tangentially related really but almost all of the variance I have ever observed in ability to remember things like this is just trying at all to use a mnemonic or routine.

The above routine is a specific practice to remember something — I basically think it would be effective for anyone that actually did it, and almost all non-adherence would be at step zero.

The original dream of crypto was a good one.

The linked USAMO math results are in an exam that requires proofs. The same authors, on the same website, ran AIME 2025 shortly after it happened and found it was totally consistent with the o1 announcement numbers; the difference being that the AIME requires only short answers and no proof.

If you are a skilled mathematician, it is quite easy to verify both that (as of 7th April) models excel at novel calculations on held out problems and mostly shit the bed when asked for proofs.

Gary cites these USAMO as evidence of contamination influencing benchmark results, but that view is not consistent with strong performance of the models on clearly held out tasks (arc test, AIME 25, HMMT 25, etc etc).

If you really care, you can test this by inventing problems! It is a very very verifiable claim about the world.

In any case, this is not the pundit you want. There are many ways to make a bear case that are much saner than this.

This whole thread is true consistent statements that differ only in emphasis.

in 2019 these things could hardly speak, in 2024 we trained one that outperforms all but the very best in the world at competition programming.

the bull case very obviously speaks for itself!

A truly wonderful novel that I also opened the comments to shill.

Been using elevenlab reader, but these are much better!

very big if true!

I downloaded their data from the metr public evals GitHub repo and independently re-implemented the very basic maximum likelihood analysis, which gave very similar results to what they shared.

The authors own summary from the position paper is:

In particular, we examine the one-step iterative process of DDIM [39, 19, 21] and show that it has limited capacity with respect to the target timestep under the current denoising network design. This can be addressed by adding the target timestep to the inputs of the denoising network [15].

Interestingly, this one fix, plus a proper moment matching objective [5] leads to a stable, single-stage algorithm that surpasses diffusion models in sample quality while being over an order of magnitude more efficient at inference [50]. Notably, these ideas do not rely on denoising score matching [46] or the score-based stochastic differential equations [41] on which the foundations of diffusion models are built.

As an aside, an FM index can be used to efficiently turn an LLM into an actual stochastic parrot (one that emits only substrings of some dataset). This is more useful than it sounds because you can use it for quoting from large corpora.

Whilst brexit was a bad idea —- look forward not backward! Planning reform is a viable opportunity and making it as good as possible is a much better use of energy than moaning about mistakes past.

I’ve had three show-stopping bugs across core apple software in the last month — I wish the competition was better!

This is grok 3, so not a debut

I’ve listened to a few audiobooks on long drives, and have been surprised how hard it is to find good voices on audible. Often a book that might otherwise be good has a prohibitively annoying tone. So honestly the exciting thing here is the customisation.

That said, even in their cherries the emphasis still isn’t quite right in the Tolkien example.

The discourse around synthetic data is like the discourse around trading strategies — almost anyone who really understands the current state of the art is massively incentivised not to explain it to you. This makes for piss-poor public epistemics.

(while I work at OAI, the opinion below is strictly my own)

I feel like the current version is fairly hazardous to students and might leave them worse off.

If I offer help to nontechnical friends, I focus on:

- look at rate of change, not current point

- reliability substantially lags possibility, by maybe two years.

- adversarial settings remain largely unsolved if you get enough shots, trends there are unclear

- ignore the parrot people, they have an appalling track record prediction-wise

- autocorrect argument is typically (massively) overstated because RL exists

- doomers are probably wrong but those who belittle their claims typically understand less than the doomers do

It’s about 1oom off. In some contexts, one oom is pretty close.

I for one am hugely supportive of compute that is red white and blue.

Can you define what you mean by novel here?

The real turker studies, resulting in the ~70% number, are scored correctly I believe. Higher numbers are just speculated human performance as far as I’m aware.

75% of 400 is 300 :)

It is allowed exactly two guesses, per the ARC rules.

I can’t provide more information than is currently public, but from the ARC post you’ll note that we trained on about 75% of the train set (which contains 400 examples total); which is within the ARC rules, and evaluated on the semiprivate set.

We worked with ARC to run inference on the semi-private tasks last week, after o3 was trained, using an inference only API that was sent the prompts but not the answers & did no durable logging.

You’re trying to to characterise personal risk and they were tying to characterise societal risk; both are valid.

The problems with python types are:

A) The type system is fairly bad, as type systems go (forgivable) B) the type checkers are, for large codebases, excruciatingly slow, to the point that the tests are faster!!

The second is not forgivable.

You get what you incentivise, not what you pay for.