The Apple paper does not look at its own data — the model outputs become short past some thresholds because the models reflectively realize they do not have the context to respond in the steps as requested, and suggest a Python program instead, just as a human would. One of the penalized environments is proven impossible to solve in the literature for n>6, seemingly unaware to the authors. I consider this and more the definitive rebuttal of the sloppiness of the paper: https://www.alignmentforum.org/posts/5uw26uDdFbFQgKzih/bewar...
If I read a comment that has any probability of changing my mind about a fact or opinion, I always go to the user page to check their registration date. No hard cut-off date but I usually discount or ignore any account >= 2020.
an American corporation does in fact have some recognized legal personhood, and so a 'bill of attainder' could technically be found to exist within a legislative act which violates the legal rights of one.
Very cool result but the title is overselling the "AI" contribution. It seems like they trained a few standard binary classifiers (Naive Bayes, decision trees, kNN). The novelty is the independent variable coming from an attribute precomputed for many known elliptic curves in the LMFDB database, namely the Dirichlet coefficients of the associated L-function; and the dependent variable being whether or not the elliptic curve has complex multiplication (CM), an important theoretical property for which lots of flashy theorems begin with assuming whether or not the curve has CM. They go on to train another binary classifier (and a separate size k classifier) to determine a curve's Sato-Tate identity component using the Euler coefficients and group-theoretic information about the Sato-Tate group (constructed by randomly sampling elements and representing the two non-trivial coefficients of their characteristic polynomials as independent variables in the classifier). They also run a PCA: https://arxiv.org/pdf/2010.01213.pdf
The cool part is that they then stepped back and scratched their heads wondering why the classifier was so good at achieving separation for these dependent variables in the first place, and plotting the points showed them to be (non-linearly) separable due to a visually clear pattern! The punchline and the reason it's so important to understand these data points, the Euler coefficients for elliptic curves, is because they contain all the relevant number-theoretic information about the curve. With some major handwaving, understanding them perfectly would lead to things like the Langlands program (and some analogues of the Riemann hypothesis) getting resolved. These wide reaching conjectures are ultimately structural assertions about L-functions, and L-functions are uniquely specified by their Euler coefficients (the a_p term in their Euler factors). Will murmurations help with that? Who knows, but the more patterns the better for forming precise conjectures.
Relevant intersectional credentials: I have lead ML engineering teams in industry and also did my doctorate work in this area of math, including using the LMFDB database referenced in the article for my research (which was much smaller back then and has grown a lot, so very neat to see it's still a force for empirical findings!).
This is something I've been thinking about a lot lately. Especially in combinatorics and number theory, there are databases like oeis, LMFDB, etc that contain tons of data with the ability to generate more algorithmically (sometimes easier said than done). Using ML to get heuristics and really good guesses on where the next opportunities lie and then formalizing it once you have a good guess would be SO cool.
Is there a name for that? Or groups working on that stuff that I could follow?
My own little pet project was I scraped OEIS and built a graph of sequences where 2 were connected if one mentioned the other in its related sequences section. You got these huge clusters around prime powers and other important sequences. Then I thought maybe you could use a GNN to do link prediction providing an estimation of a relationship that should exist but hasn't been discovered yet.
The Lean 4 Focused Research Organization has ML interoperability in its roadmap. Since Lean 4 is shaping up to be a capable general purpose language as well, I can imagine a Lean project that retrieves and formats LMFDB data, uses it to train and test a NN, gets Lean 4 proof code from it, verifies or rejects it (possibly with more detailed feedback) and loops this like a "conversation".
However, Lean 4 still has a long way to go in terms of speed and library features, and I at least have given up on writing optimized code until we get the new compiler (whose timeline seems optimistic to me, but Leo de Moura knows much better).
We would be interested in actual automation of theorem production, but this pipeline would automate approximately 0% of (interesting) mathematics research. It does have the potential to automate some boring parts and enable mathematicians to make better conjectures faster.
I think I may be missing something. Why would you be interested in the automation of theorem production? Wouldn’t this make mathematicians obsolete? How far away do you think we are from that?
I ask as a newbie in math; math is a passion of mine. I am genuinely reconsidering going into math research as I fear just being automated away.
I am not a mathematician but have some interest on a pop-sci level.
I believe this presentation at G-Research by Alex Davies would be of interest.
https://www.youtube.com/watch?v=Mp_skPK-X9M
In these area of physics informed machine learning this is refered to as "discovering new physics". Probably there are analogs in computational mathematics, biology, chemistry, etc.
> Very cool result but the title is overselling the "AI" contribution. It seems like they trained a few standard binary classifiers (Naive Bayes, decision trees, kNN).
But it seems they would never have even suspected there were such patterns if the "AI" had not provided evidence for them?
By the way: the tools mentioned, like decision trees, Bayes and kNN were all taught in the AI course I attended one and a half decade ago... AI was basically ML at the time, but nowadays it seems that ML has become "just statistics", and AI only includes LLMs.
There are plenty of companies using ML methods (DT, Bayes, kNN), normal NN etc now that the AI money spigot is wide open, if only as part of the "shit in, shit out" process.
They are excellent, and not requiring more than high school maths knowledge to really get quite deep into the mysterious connections between prime numbers, Riemann hypothesis, elliptic curves and L-Functions.
As someone who understands about 2% of the GP but maybe 85% of TFA, I'd suggest diving into the various topics explored there. Galois Fields, for instance, are a rich topic for Wikipedia research and have intuitive and surprising properties that make them fun to learn about.
This will lead you deeper into study of abstract algebra concepts like groups and rings. If you haven't done much set theory you will probably go deep on that and develop an opinion on the Axiom of Choice.
Then you'll probably surface a bit to look at elliptic curves and consider their many applications in abstract and concrete topics like cryptography and the elusive proof of Fermat's Last Theorem.
By then you'll have caught up to me. In the meantime I'll be reading up on module forms and L-functions.
Sounds like it's far more about "big" data analysis, and recognising that elyptic curves encryption has a statistically apparent signature. AI/ML was just the analysis that exposed it.
Not really. Only a tiny slice of the historical person’s memories and persona is recorded. There is a lot more entropy to their representation that died when their brain did. Ergo, whatever “perfect” simulacrum is presented will need to infer the gaps and ultimately be fictional.
By the pigeonhole principle, there is a sentence that writes out its entire SHA256 representation this way. Alternatively, the map from these kinds of sentences with 256 terms to 2^256 given by SHA256 admits a fixed point.
The pigeonhole principle does not say that. It can be used to show that there are two different sentences with the same hash as each other (among any collection of 2^256 + 1 sentences), but it tells you nothing about hashes that agree with the content of the sentence. The probability that a random hash function on a collection of 2^256 sentences has a fixed point is about 1 - 1/e, and it approaches 1 as you add more variations to grow the collection infinitely. But SHA-256 isn’t actually random, so the only way to know this for sure would be to find an example.
I don't believe this is necessarily true. Unless I'm misunderstand you, each of the possible variants of spelling out 32 hexadecimal characters could theoretically SHA-256 into the spelled-out hash + 1 (looping around at ff…ff).
I don't see how pigeonhole principle applies to that situation. It could well be that "zero" hashes to 1, "one" hashes to 2... and "f" hashes to 0, extended out to the hash's length.
Just a note that this is by Geoff Anders from Leverage Research, an organization historically plagued with some controversy in the level of psychological experimentation it is willing to perform on its members:
Hi, this is the official Leverage Hacker News account. Some clarification on the research that took place:
During the relevant period, researchers were permitted substantial freedom in determining what experiments to run and hypotheses to explore. Researchers also participated in experiments as they saw fit. This was voluntary, not all members participated in psychological research, and promotions and salaries were not tied to participation.
One should imagine a purposefully unstructured environment with 30+ people trying to figure out how the mind worked and which self-improvement modalities worked best, rather than subjects at a clinic being experimented on. Our researchers explored tons of hypotheses and we think that was great.
There are difficult questions about balancing people’s freedom to experiment with their own minds with safety in experimentation, including in an institutional context, and that’s something we think there should be more public discussion about.
