Just dropping in to say it's nice to see somebody actually try to work through the proof, and it gives one confidence that the proof at least isn't complete nonsense (which is helpful given the few details provided about the process behind it).
With the Erdős proof, OpenAI added perspectives from working mathematicians that gave some context -- hope something like that appears for this one eventually.
I thought this is a cool premonition of the eventual success of language models. One can nicely see a microcosm of everything with generative AI (text generated in a domain you don't understand looks fine at first glance), already a mention of attention models, and there's even a reference to the model "hallucinating" some URL.
It's also all charmingly tiny. I was surprised at the quality of the LaTeX generated from a network trained on a single book -- it looks like a proof, has sequential lemmas, closes all brackets...
Hi, thanks, and very cool work (assuming it eventually holds up in peer review)!
A few things that confused me while trying to read the paper:
- There's two different methods of cell division mentioned -- mechanical extrusion and the autonomous, protein-driven division. Most of the results (e.g. the five generations) focus on the mechanically driven one, while the autonomous one is more "lifelike". Does the autonomous division have a higher failure rate, or can you get the same results with it as well?
- It's mentioned that the bottleneck for survival of many generations is ribozomes degrading, but also that ribozomes are supplied from the outside. Do the degraded ribozomes actively harm the cell? Or is there some other reason why they cannot be replenished?
- You say that after 5 generations, only 30% of the cells have the correct genome, and it's presented like a problem -- but 30% of 2^5 is more than 10, so this sounds like more than enough for continued survival. Is there something missing in this train of thought? Perhaps other failures that can kill the cell?
And some questions about the implications:
- Do you think that the genome you use is already close to minimal? AFAIK a lot of the minimal organisms found in the wild are parasites of some sort, getting most of their complex molecules from the outside, which is a similar spirit to this (a rich medium and the cell "just" self-duplicating). If the multiple plasmids are causing trouble (per the previous point), would it make sense to try and get rid of some of them?
- Are those minimal genes somehow interpretable -- as in "you need functions X, Y, and Z and cannot avoid them by using a better medium"?
- Do you think this is a plausible stage of very early life?
With these confident comments I would appreciate some kind of origin of the information. Not even necessarily a source accessible to me, just: are you in any wealth management offices? Or are you reporting other people's opinions? Or does it just sound right given the spirit of our time?
I don't think this changes anything important, but my understanding is that the lowering is the scan -- you go through the ring, which captures data about the slice currently inside it. That gets you down to ~80s, which rounds to a minute (they say "about 5 cm/s").
Now, there's a lot of other reasons to be skeptical (e.g. there's no information on what all of this imaging could actually resolve), but please don't shoot the message.
LIGO detects length changes of 10^-18 m, or attometers, not femtometers, which are a thousand times longer. (https://www.ligo.caltech.edu/page/facts) But this does not matter at all, because this is not resolution of the body image, but the size of the vibration on the speaker. That's a technical data point that I don't see any reason to include in this presentation other than to cause this exact confusion.
The video looks in general like it's trying to impress by giving a lot of incidental information about how the device works while being very light on what it would be able to actually see -- e.g., it doesn't matter how many gigabytes your device collects if the resulting image is blurry.
Compare the website of LIGO (https://www.ligo.caltech.edu/page/facts), which also has a lot about the technical marvels (huge vacuum tubes! precision engineering!) but crucially includes the goal of this all.
Oh my god, 1/0 is an absolutely brilliant piece of art and I recommend everyone here to check it out. It starts out unassuming, but that's part of the point -- it ended up becoming about the author's growth through talking to characters he himself made up (and the characters talking back in protest), and he also ended up meeting his wife through it :)
I guess if I had to sell the idea... in its own words: it's as far removed from the average sitcom as possible. It's not at all like anything else you have ever read. (https://www.undefined.net/1/0/?strip=961)
It's also relevant for non-AI reasons. The upshot is: Do it if you want to. You're at an age where you're supposed to be mostly learning, but over time, start transitioning into whyever you did this (which doesn't have to mean staying in academia, and in fact can mean teaching or popularisation or consulting or...).
I have a personal benchmark for measuring AI problems in the form of hand-drawn Bongard problems (https://en.wikipedia.org/wiki/Bongard_problem). The idea is that there are two sets of six images that differ based on some feature of the images, and the task is to find the dividing feature. This task is not perfectly well-defined, but usually there is a single solution that strikes one as obviously canonical once found.
They are nice because it's easy to hand-draw new ones with solutions that probably don't exist in the literature, and because for some reason they have proven quite hard for AI.
Sadly, the recently reported advances in generative AI for problem-solving require expensive models I don't have access to. Could somebody try pasting this image to GPT-5.5 Pro or Claude Opus 4.7 or the like, with the accompanying text "Hi. This is a Bongard problem. Can you solve it?", and share a link to the resulting chat? I would be curious.
The free models (Claude Sonnett 4.6, GPT-5.5, Gemini 3.5 Flash with extended thinking) all give obviously incorrect solutions (rules that don't actually hold for the images), to the point that I think there must be some problem in the image processing. Example: https://claude.ai/share/1ff7b5c2-c34a-40cc-a249-2d0fd3474884
P.S. For obvious reasons, I'm not sharing the solution, but I have verified that most of my friends found it within 5 minutes, and everybody found the same solution.
Thanks for responding this way! This had flame-war potential that didn't get realised. I'll try and reply in a similar spirit.
I still find it curious how few holes there are (took a while to find one!), and finally figured out why: imagine a large square grid. It would probably have a different density than the rhombus grid, and it seems nontrivial to match it up. It seems that in the code this is done by each rhombus having edge length 50 while the periodic table elements are 38 x 42 pixels in size.
This, if I understand it correctly, means that this tiling is not just aperiodic but (in this regard) also anisotropic -- it's denser in one direction than another. And thus I have learned a new thing about Penrose tilings. :)
I'm sorry, but... this is why I'm unhappy about AI-generated projects.
The parent post asked a more-or-less specific question about how this works and offered their working hypothesis (that it works by taking the rectangular periodic table and mapping each element to its nearest Penrose tile). I would also be curious about this, because e.g. it seems non-obvious to me that the resulting table would have no holes inside. Does there have to be any extra step to ensure that?
But your reply is not informative at all. If you wrote this yourself from scratch, you would be able to answer this kind of question (and report on anything interesting that came up while creating the program), but this way the post just hangs here to be forgotten, and I haven't learned a new thing about Penrose tilings. :(
I use RSS to get updates from a ll the stuff I read online at once, and thought this would be nice for those websites that don't already have an RSS feed, but... Perhaps I'm stupid, but I can't actually find the RSS output? And searching for RSS on https://docs.sitespy.app/docs returns no hits.
Actually, the search here is chemical -- you make a bunch of random small RNA sequences, then chemically replicate those that do what you want, and after a few cycles of this you find just a few sequences dominating the population. The pools can easily reach 10^14 examples and more, because you don't need a separate container/experiment for each one, but do it all together in one batch. The keyword is SELEX, a pretty cool idea: https://en.wikipedia.org/wiki/Systematic_evolution_of_ligand...
The numbers across age categories do not correspond to the ones in TFA, but they're also insane, and around 10% average. Combined with 20% vomiting side effects, that would suggest 2% of Americans have been throwing up, which sounds... implausible? Can an American say whether this matches what they're seeing/hearing in their social circles?
Original title is The biggest mystery in neuroscience, according to me. Since the me is not me but computational neuroscientist Chenchen Li, I changed the title to the mystery itself.