Rice’s theorem is about the worst case. If you are thinking about the worst case for example because the model explicitly optimizes to not be understood then yes that can be a problem. Rice’s theorem tells you that you should not pick the goal of being able to interpret any model even if someone explicitly aimed for training a model such that it will be hard for you to understand. If you have control over what data you train on, and you try to understand the model at different stages during training, and you stop training once you really don’t understand what is going on, then you should be in a better shape. Python is also Turing complete. But lots of python snippets people write are quite easy to understand. Programmers reason about the semantics of some code all the time. The more code and the less optimized for understanding the more difficult it gets though. In some ways LLM’s are pretty close to a random binary, so yeah mechanistic interpretability is in trouble.
Morpheus
Congrats on making it! Good to link to your substack I would otherwise not have found your other posts! I was quite tempted to join your attempt once I realized it wasn’t an April fools joke. But I do not in fact currently have the time or energy to pull this off.
You are telling me this unbelievable result with suspiciously high success rate with the ants hasn’t been reproduced in 10 years? This experiment shouldn’t even be this hard? If you are a bright high schooler or undergrad with a microscope, I am happy to pay you $100 if you end up reproducing this.
My best guess would be that the sun breaking down folate (Vitamin B9) would be more of a reproductive fitness disadvantage especially during pregnancy compared to the very few skin cancer cases late in life. I don’t have strong evidence either way though. Probably both effects are roughly within one order of magnitude? Also light skin could be faster selected for than dark skin since vitamin D is so important. As Carl Feynman notes, there could be other biochemistry stuff.
Labor is only one input to algorithmic progress (compute for experiments is another), and algorithmic progress itself is only one component (though probably the majority, perhaps around 60% or 80%) of overall AI progress (scaling up training compute and spending more on data also contribute).
I notice I am confused about this algorithmic progress. Is this compatible with what Steven Byrnes wrote on LLM algorithmic progress? Does anyone here know? Or is there already something written up on how this conflicts with Steven’s picture. I was quite confused why there was so little discussion from people I would expect to know about that topic, to figure out what the consensus on this is. I find this somewhat cruxy for how confident to be that we are going to get to AGI soon, which is relevant for me, since currently I work on enabling genetic human enhancement. I guess if people believe discussing this in too much detail is exfohazardry, that would be good to know, so I can just look into it on my own.
that’s before we even consider IQ as only a reliable measure of a narrow construct—where is the reliable measure of creativity, practical judgment, personality, motivation and all the other social skills?
Let me know once you find a reliable measure of those constructs with enough DNA data attached to make a predictor :)
Importantly, there is limited evidence supporting almost all the claims, however well intentioned. That PGT-P improves real-world child health outcomes in a way that justifies its routine use is unproven, as is that it reliably predicts or increases intelligence in any practical or meaningful sense
What type of evidence would change your mind?
While I’m normally using a trackball as a mouse, two years ago I went to go co-working and used a normal mouse. I made a bad movement while using the mouse and afterwards my right hand hurt a bit. A few days later my hand was relatively okay, but my hand and arm were still more tense than before.
I asked multiple bodywork people to fix it, but while the arm got more relaxed the issue didn’t fully resolve. This week I decided to investigate how my right hand and left hand differ to find out what’s going on. I noticed that if I extend my right arm my right hand goes in the direction of the ulna side unless I add tension to keep it in place.
When palpating the ulna head from the dorsal side of my left hand I was touching the ulna head directly. When doing the same thing on the right side, there was something above the ulna head. I formed the hypothesis: “Maybe, the thing I’m palpating is out of place. How about I move it laterally?” I used my fingers to slowly push it laterally.
Afterwards, my right arm started relaxing. I fixed the problem that I produced two years ago in 10-15 seconds of action. I looked up the anatomy and deduced that I moved the tendon of the muscle extensor carpi ulnaris. The tendon is supposed to be lateral of the ulna head and not dorsal. This explains why my hand moved before when extending my arm. Part of extending the arm involves turning the ulna and as the ulna turns, the ulna head presses a bit in the dorsal direction and pushed on the tendon. As a result of pushing on the tendon the extensor carpi ulnaris contract resulting in the movement I observed.
Untreated, this issue might have resulted down the line in carpal tunnel syndrome or back pain down the line. Plausibly, it would have even produced those effects in the two years if I wouldn’t regularly do effective interventions to remove tension.
I noticed anatomy was another one of those areas where it wouldn’t even occur to me that I should probably learn the very basics of the subject, and perhaps I should just use some common sense and observation to make sense of things. For example, my doctor recommended that I get a tooth grinding guard (I used to have one as a teenager, that I stopped using after a while, and it didn’t fit anymore), but the only way they assessed whether I needed one was not based on my teeth, but the strength and tightness of my jaw muscle. I am pretty sure I don’t grind my teeth during the day (I can’t fully rule out yet that I do so during the night). Before I got an expensive guard, I decided to just set an hourly reminder to check what I am currently doing with my jaw. I noticed I do in fact tense up my jaw when I concentrate, but I seem to be doing so without having my teeth touch each other. Sadly the issue is symmetric, and looking at some jaw images online also didn’t give me obvious ideas what to do about it, but at least I won’t waste my time with a guard that I don’t actually need.
Hedgehog thing got resolved. Downstream of the hedgehog gene being mentioned somewhere in the references of one paper I pasted into the chat and after gpt-4o searched for it once it must have influenced other chats.
Yeah, the low extrinsic mortality is definitely something NMR have going for them that ant workers don’t (queens do since they don’t leave the nest usually). Indian jumping ants also compete for being queen like NMRs, but they forage and leave the nest. There are definitely academics that think in detail what differences in extrinsic mortality risk predict for life-span. When I just asked Claude on that it apparently is actually more complicated than lower extrinsic mortality=slower rate of senescence? ¯\_(ツ)_/¯ Similarly, I am not sure if the low oxygen environment is actually that great for NMR longevity. As ChristianKi mentions your theory is going to be more interesting if you can look across species what it would predict or what your evolutionary theory predicts regarding mechanistic/biochemistry and genetic stuff in NMRs. As ChristianKi, I would not call myself an expert, but I might be somewhere on the Pareto frontier of having looked into aging, eusocial animals and naked mole rats. Not sure about relevant experts. Maybe check the people who wrote the textbook on naked mole rats? You would need some generalist that would know enough about these senescence models and about particular biology of naked mole rats and this expert might just not exist in 1 person.
I am curious: What is the theory? I’d be surprised if your theory works, applies to naked mole rats, but not to ants and other eusocial animals. I always thought naked mole rats live long because they are eusocial, so you having a theory specifically for naked mole rats sounds ominous.
I like the idea of leveraging the GWAS data and just throwing some omics data at it to see what you get, but your particular experiment isn’t the most interesting from my perspective? Why look at protein interactions? Bipolar disorder isn’t a disorder where something goes wrong with all cells, it is something where something goes wrong with the brain in particular. If your analysis had told you that you get some genes clustering around mitochondria regulation in particular, would you have believed that and what would it have told you? After skimming the Wikipedia article on bipolar, I guess maybe you could have found some evidence for why on earth lithium seems to help with bipolar? Also one obvious pitfall is linkage disequilibrium which means not all your snp’s are going to be causal (because closeby SNPs are inherited together). Probably you can do some sophisticated statistics to filter for only those snp’s where it’s relatively obvious what gene is influenced because the gene is large or the region of possibly causal snp’s is small. Let’s assume your sample size is large enough and you accounted for that.
To me the most obvious thing to look at would be what specific type of brain cells are associated with bipolar. It also seems like the type of thing that has already been done. Some diggin with claude finds these two papers with the graphs below.
In the above I find it surprising that the pancreas is showing up for character traits I would think of as more mental traits? What is up with that? Is bad energy household making people more impulsive or something like that? Seems worth learning more about.
Overall nothing really interesting with the above graph for bipolar here: (Medium spiny neoron->striatium-> something something reward/feedback? Duh! Or perhaps this is just pointing at the intelligence confounder) To me it was surprising to see Endothelial-mural cells this high up for cognitive performance, but maybe just a fluke.
I would possibly find it interesting to see what happens if you take the large datasets we have now on mouse whole brain cell atlases and combine this with GWAS data on mental disorders, personality etc. Mouse are obviously very different, so not sure how informative that would be for the smaller cell clusters.
Writing up my thoughts here, since I did some more research a month ago, but unfortunately didn’t write it up at the time. I will probably not get to looking further into this any time soon. If someone else wants to pick up the torch on this feel free to do so. I find this is a good exercise to broaden my biology knowledge.
The number of peptides measured in that study seemed kinda small after looking at the table again. The long lifetime for Histone 3.1 etc. is not super robust. I looked at this other study that measured lifetime of proteins in mice and IIRC the lifetimes were just generally shorter. Also I am not sure how easily the lifetimes from mice translate to human etc., so the protein thing in particular I find less plausible now.
It then also occured to me that maybe Lysosomes filling up with gunk or whatever could be the problem. Currently I don’t have a great understanding of Lysosomes, and Claude Opus 4.6 even tells me there might even be this mechanism how post-mitotic cells get rid of stuff in Lysosomes, also generally they don’t really take up that much mass, so I don’t have great intuitions if this is generally plausible. I was hoping there is maybe just something that you can feed to mice that lands in their lysosomes and can definitely not be digested and then you could check if those mice look like they are suddenly aged by a few years, but I am not sure if something like that exists or could be designed. If that’s a thing, that seems like a great legible experiment. Also after thinking about it more if it was mostly some specific long-lived cell types that make the difference, I’d expect if something messes with those cell types in particular, it would lead to some type of progerias, so this probably makes the idea of it being the somatic cell types less plausible. I am also not sure, if it is something from the post-mitotic cells that feeds into a self-destructive feedback loop in the mitotic cells, then adding a 70 year old heart or 70 year old brain would not suddently make you 70, it might just give your aging a 2x speedup boost which is harder to distinguish from things that are not the root cause like bad diet. Not sure but that could be tricky. Also if it is just those specific long-lived cell types that predicts you’d expect animals that are selected for different lifetimes or that rely on those cell types more or less to have different adaptions, so that’s a thing to check.
I just quickly tried to actually look at some graphs for some mouse papers that look at organ transplants, but I really don’t have the capacity right now to understand the details of those experiments. But yeah if it was the post-mitotic cells that don’t turn over, you’d expect a mouse to age faster after a heart or brain transplant, than a liver transplant.
High Modernist
If execution in a lot of areas becomes cheap and having good ideas to execute becomes the bottleneck, generating and sharing your ideas online might be worth it more often on the margin. If the project easier to do in 12 months and it is not mission critical right now, why not delay execution for now.
Mitochondria and Lysosomes could also run into issues in post-mitotic cells, as well as other things I haven’t thought of.
I was especially thinking of your recent post “on goal models”. I tried to look up other posts, then I just saw your post on distributed agents and then this clicked and I feel like I better understand now where you are going with this. I find distributed agents and how to think about them confusing.
With most of your posts I already agree with the conclusion, I disagree and am still not convinced after reading the post, or the topic seems really confusing and after reading your post the topic seems more confusing than before. With this post I thought I had a good explanation, but now I see it wasn’t adequate. When I saw the title of this post I was thinking this was going to be interesting. This post genuinely changed my mind.
Some more completionism for proteins that don’t turn over or only extremly slow in mammals:
Rec8 in Oocytes (it holds the chromosomes together)
Some histones (Histone H3.1 in this rat study)
Some proteins in the nuclear pore complex (Nup205, Nup93 etc.)
Nuclear pore complex proteins not turning over in post-mitotic cells suggests a new hypothesis for aging I haven’t heard before: Most cells turn over, but there are quite a few cell types that don’t or extremely slowly (brain, skeletal muscle, heart (only slowly), oocytes, podocytes etc.). Most things inside of those cells do turn over, but some structures only on the scale of decades (see above). I can see nuclear pore complex damage and histone damage feeding into DNA damage. If cells that don’t turn over spread their damage to cells that do, I could see how this might be the upstream ratchet on the “The DNA Damage <-> Mitochondrial ROS” loop, so it’s not the stem cells after all. A weakness of this model is that nuclear pore protein mutations aren’t really associated with progerias. The first thing I could think of to test this was looking at what happens if you transfer different organs from young to old mice and vice versa. Will report back what I find.
Artificial Yeast exist now! So someone might want to run the experiment!
I am less excited about doing this in yeast than when I wrote this comment, but overall would still find it interesting. I find it more likely now the bottlenecks in Saccharomyces cerevisiae are sufficiently different from multicellular eukaryotes that things might not translate (even if transposons were the root cause in mice and humans). Most importantly, in yeast since they are single celled, we are mostly talking about aging of the mother cells, that split asymetrically from their daughter cells. Evolution then found weird tradeoffs that lead to bugs in mother cells with improved results in daughter cells:
For example Extrachromosomal rDNA circles (ERC) accumulates in mother cells and shortens their lifespan, but somehow this mostly doesn’t end up in the daughter cells. From my shallow investigation it seems relatively well established that deleting Fob1 significantly improves the lifetime of mother cells. The story being Fob1 is involved in the repair of rDNA, which improves growth potential of the daughter cells, but creates these ERCs that accumulate in the mother sometimes accidentally. There might be many similar examples like this, but none of these problems really translates to multicellular organisms, where this gets solved the same way as in yeast colonies through selection.
If transposons were the root cause, for aging in mother cells, then just going through gametogenesis should not reset the life-span of yeast, but it does (note though that I couldn’t quickly find a replication of this paper). When yeast perform meiosis, they split their nuclei into 4 spores and a fifth dumpster compartment leaving behind ERCs and other things that would hurt the spores.
There is still no firealarm for ASI. How things will work out is still uncertain. The AI talking to you is more like smoke than a firealarm in the firealarm analogy. It’s still hard to coordinate a pause around. But yeah past ChatGPT your chances are still higher than post GPT-2 or post AlphaGo. I don’t know where I read it but yeah I think I read somewhere from some people involved in Miri explicitly that they are not trying to make the wider public pay too much attention to AI which makes complete sense given how people end up doing stupid things in response to that (the founding of OpenAI etc.). So not making too much of a fuss at that time just seems correct.