Anders Sandberg

This paper is a goldmine on scientific self-experimentation. -14 Nobel Prizes have gone to self-experimenters. - Of 465 scientific self-experiments documented over a 203-year period, there have been 8 deaths. - The most recent recorded death from a self-experiment was in 1928, when Alexander Bogdanov injected himself with an incompatible blood type. - Many universities say that self-experimentation would require IRB approval because it violates "ethical norms for medical research," which is not true; the Nuremberg Principles make an explicit exception for people experimenting on themselves, and the Declaration of Helsinki just says the subject must consent. Also, "there is no law nor regulation identified that requires investigators experimenting on themselves to consult an ethics committee." - There are lots of recent self-experiments; "In 2014, Philip Kennedy had electrodes implanted into his speech center to further his research on direct brain interfaces. In 2016, Alex Zhavoronkov self tested drugs which his software algorithms identified as likely candidates."

I'm one of the 'additional expert contributors' to this report! It's a big deal to see the Allen Institute share this, I'm glad whole-brain emulation (WBE) is gaining recognition as a legitimate goal by the scientific community. alleninstitute.org/news/how-far-a…

@HopeExistential @johanknorberg It was great fun! Hopefully a start of a lot of progress.

Every year the economic gap between the US and Europe grows wider. Progress studies is the field that asks why: why civilizations flourish or stall, and what we can do about it. It has gained real traction in the US, but the conversation is just starting in Europe. Last Monday in Stockholm, we hosted the first in a new salon series to help change that. @johanknorberg talked about why golden ages end and how we can keep Europe’s alive; @StefanFSchubert about why we underestimate progress; and @beatrice_erk about what progress studies is and why Europe needs it. Big thanks to everyone who joined us.

@WallieF20103 @perrymetzger So, use named theorems or problems instead.

@anderssandberg @perrymetzger Using publication rate as a measure for productivity is bad.

Which is the stochastic parrot, the thing that can solve an Erdős problem, or the thing that just mindlessly repeats the same statements over and over again no matter what evidence is presented to it?

@DrPhiltill @andmar74 @is_OwenLewis Yes. I think bottlenecks in life are a bigger cause of Fermi than rare earth's. You need earth to be extremely unusual for it to work, and "extremely unusual" may mean something like one in 10^20 rather than one in a million.

I never read that book, but we now know solar systems like ours are the oddballs. Most have gas giants too close to the stars so terrestrial planets couldn’t have stable orbits over sufficient time. I am not sure what the statistics are for “good” star systems and planets, though. My view is shaped more by the origin of life having many unexplained links. I’m not saying they won’t be filled in. I’m just saying that I don’t think we can rule out rarity of life and hence advanced civilizations yet. There was a great study by @anderssandberg that showed the bottlenecks in our history were spaced roughly evenly in geological time, which is the expected spacing if planets with observers like us are rare, so there is at least some potential evidence for our rarity. I’m not arguing positively for this view. I’m just saying it isn’t BS, and my guess places it fairly high for at least moderate rarity.

This is a great summary of what I think the answer to the Fermi Paradox probably looks like. Also, AI *probably* isn't the Great Filter, but I guess we'll find out.

@kareem_carr I think this applies to humans too. We don't get as many stories about afternoons when Euler or Chauchy failed. Some anecdotes about struggling (e.g Wiles), but we are biased towards the results and less the process.

There's a kind of bias in AI results where they only tell you about the one time a prompt succeeded, and not the hundreds of times the prompt failed.

@Amrak999999999 @perrymetzger The prompt is in the paper.

@WallieF20103 @perrymetzger Actually, his amphetamine use boosted productivity for the last 25 years of his life. He was so productive that one can actually check it statistically by publication rate and key paper rate. If there is any poster boy for successful cognitive enhancement it was Erdös.

@perrymetzger Yes, problems posed by a meth head who tried to make himself into a computer. Once his meth stopped, all his math stopped. Really tells you something about how he did math, right?

@andrew_v10209 Yes, maybe a definition of a moment is when it makes people in denial to change their mind.

@anderssandberg For the type of people in denial, I don't think this quite changes anything yet. You could still just ignore this if you really wanted to. Maybe some of the people who were dismissive yet were still paying attention had their minds changed.

Is the unit distance conjecture disproof a "moment" like the Sputnik, ChatGPT, DeepSeek or Mythos moments? I think it is a small one, but a pretty clearly heard bell to academics. To the wider society is just more "AI does stuff".

@Traansistor That will likely cause waves. But it also requires more than a proof (for many of the problems, maybe not the Riemann hypothesis): they are more about getting an explanation or understanding of something. Not there yet, but I expect it eventually.

@anderssandberg I don’t think you can beat one of the AIs solving a Millenium Problem

@BishopBlougram @shallit43 The spikiness of current AI is interesting. I agree that LLMs are lousy at even well described games like chess while now mastering mathematical problem solving - something many of us mentally treat like a chess game. But clearly the AI did not.

@shallit43 @anderssandberg Happy to spread mirth. The complexity conundrum is very non-intuitive. Open math problems? Check. Figuring out chess rules after being fed 10,000+ games? No. Math professors fail at the former; five-year-olds ace the latter (after watching a couple games) @Blougram/grookie-state-of-the-art-ai-cosplaying-at-chess-3481040ed396" target="_blank" rel="nofollow noopener">medium.com/@Blougram/groo…

This is impressive: it is a problem I had actually heard of. It looks like the solution approach is surprising to mathematicians. It was a general reasoning model rather than a specialized one: bitter lesson time. I think the stochastic parrot is now nuked from orbit.

OK, looks like we need to ensure that our office toilets are now unisex. Which is fairly reasonable in my opinion. bbc.co.uk/news/articles/…

@robinhanson I agree. The epistemic argument for markets is one of the strongest.

I know many of you will roll your eyes at this, but I actually think his point is quite deep.

@nycthinker Have you checked the Mozilla vulnerabilities? If this is a pre-IPO gambit, it is darn effective one. arstechnica.com/information-te…

@anderssandberg Both this and the Mythos moment are pre-IPO marketing moments with no significance outside of narrow niches where LLMs are exceptionally strong due to the semantic stability and constraints of the systems for which outputs are generated.

@olorininir @grayhamandeggs I think this will be settled if they keep on going through the Erdös conjectures at a brisk pace. You predict it was a one-off, I predict it will just keep going (likely with less fanfare).

@grayhamandeggs @anderssandberg My point is that the team behind this *probably* 1. Tried lots of prompts 2. Did alot of pre-training and talking with the model 3. Fed in lots of synthetic data 4. Probably spat out very many solutions that needed to be checked This required lots of work and money, by humans

The cool, hard, and current question is what a PhD will mean in math and soon physics. We likely have to do some reinvention fast. Ah, academia, the institution well-known for pivoting fast and well...

Real capital-M "Moments" are signals that reshape the discourse, become common knowledge, and focus attention in new ways in the large. But society is made of subsets, and there are smaller moments when they react.

@pharmst See the graph, there is a curve for finding stuff in the literature. There are earlier genuinely original proofs as far as we know, but this is much more of a "real" problem than many of the earlier.

@anderssandberg Didn’t even the minor Erdős conjectures turn out to already be in the literature? I think this might possibly be the first genuinely original proof of a real problem by AI.

@DanielleFong @lightcellenergy Really neat! Textbooks always give a static picture, and even animations rarely allow the arbitrary disturbances from poking that tell you things about the dynamics.

Claude and I got a live sodium atom simulation! @lightcellenergy I've long wanted to be able to visualize what is really happening when a neutral sodium atom is excited in a flame or a bulb. Classical theory teaches that the single outer electron orbits the nucleus and full shells, and forms a dipole when this gets "wanged" but, physicists have long wondered, what does it look like when the electron gets wanged? is it like ringing a bell? i thought that this was the perfect time to really push what claude.ai is capable of, in terms of building artifacts. grabbed zig to compile wasm and webgpu to make it super performant, and work with it for a bit and ran the REAL TIME-DEPENDENT Schrödinger equation on a 3D grid in a potential in the browser. and you can kick it! what!! it's soo beautiful... this is going on the lightcellenergy.com website overhaul for sure here is the artifact. i don't endorse that it's exactly right, but i feel it is probably spiritually a good sodium simulation. it certainly rings my dipole! claude.ai/public/artifac…