ftacv.bsky.social @ftacv.bsky.social · 01/10/2026Personally the most likely outcome (i.e. the one that only requires inertia) is stratification into different publishing streams, with assessment of quality in the high tier stream primarily on "prestige", and the low tier stream a deep and wide ocean of slop 030
ftacv.bsky.social @ftacv.bsky.social · 24/09/2026Rereading the paper, I think they have 1) found an RNAseq database and said there's overpression of the locus and 2) expressed the gene, done RNA seq and said there's overpression of the repeats so therefore RT activity. Don't like the elision between Claude and the researchers at all tho 020
ftacv.bsky.social @ftacv.bsky.social · 24/09/2026[sorry for the spam, last one!] A grad student could do it, but they wouldn't because it would be insane, miserable drudgery. 001
ftacv.bsky.social @ftacv.bsky.social · 24/09/2026As a internal mental model of AI scientist "being able to try lots and lots of different analyses and data mining operations in parallel -> handoff to a properly defined experiment so you're not p-hacking" I think is the most likely approach to bear fruit at the minute. 110
ftacv.bsky.social @ftacv.bsky.social · 24/09/2026From my very cursory reading I don't think they did any wetlab stuff which is the acid test of this sort of thing - you can hang any neat explanation you want around sequence data, but you have to show it does things! 100
ftacv.bsky.social @ftacv.bsky.social · 24/09/2026[Not a bioinformatician] central claim in the preprint imo is about being able to pursue unexpected structure in a highly distributed way - not the recognition itself but the correlation to co-expression and finding a study that shows higher relative prevalence over the time course of an infection 100
ftacv.bsky.social @ftacv.bsky.social · 18/09/2026I did like lots of Gaiman's work, but thinking about it since the allegations came out - there's an incipient (sexual) nastiness to lots of his non YA stuff. Now it just feels like he's laughing at you for leaving it in plain sight. 030
ftacv.bsky.social @ftacv.bsky.social · 08/09/2026Banging your idea against the anvil of "why should anyone care about this" *is* productive but goodness does it hurt 001
ftacv.bsky.social @ftacv.bsky.social · 02/08/2026It makes you dumb and weak! It's frustrating, Claude was oneshotting some really hard numerical analysis problems that would have taken me a couple of days at least, but every problem I'm now like eh.... Let's see what the AI thinks 010
ftacv.bsky.social @ftacv.bsky.social · 26/07/2026No one ever brings up cottaging or dogging in these contexts, but I think we should have more pride in our heritage 0110
ftacv.bsky.social @ftacv.bsky.social · 29/06/2026(sorry this is a bit imprecise but you get my meaning) 010
ftacv.bsky.social @ftacv.bsky.social · 29/06/2026Obviously the above is getting close to the famed AGI, but for me the continually surprising finding of the LLM era is how far you can get (i.e. possibly remaking the economy, society etc) by remixing already existing concepts on an industrial scale. 110
ftacv.bsky.social @ftacv.bsky.social · 29/06/2026I think that's still explicable if you consider it as a translator of concepts. What I'm talking about is something like "we've sequenced this archea from a hydrothermal vent and the LLM has successfully predicted the role of the proteins with no known sequence alignment" 210
ftacv.bsky.social @ftacv.bsky.social · 26/06/2026I think the accurate world model argument won't be convincing to me personally until it makes OOD predictions, which is hard to assess because people are bad at that as well! I personally like the set of disclaimers here and find it a bit moving github.com/teorth/erdos... 110
ftacv.bsky.social @ftacv.bsky.social · 26/06/2026I think the "just text" school is unfortunately associated with the stochastic parrot people, but I think it's possible that this level of ability is encoded into the structure of human language/the corpus. 310
ftacv.bsky.social @ftacv.bsky.social · 26/06/2026Yeah I think the thing I'm most agnostic on is how much of that represents an internal world model (i.e. capable of extrapolation) and how much is extant in the corpus with an extremely generalised ability to translate (less so?) 110
ftacv.bsky.social @ftacv.bsky.social · 26/06/2026I'm sure you've seen this, but I think this is where I fall on the agentic side - www.argmin.net/p/secrets-of... but I think it's an arguement about why they are effective, rather than why they'll hit a limitargmin.netSecrets of Intelligence ServicesI asked Claude code to build Claude code. You won’t believe what happened next. 010
ftacv.bsky.social @ftacv.bsky.social · 26/06/2026And sorry to ramble - if we extend this to scientific hypothesis generation, we would expect them to be good at applying existing concepts between different fields, but much poorer at suggesting the genuinely new. But much of science is the former, which is exciting! 110
ftacv.bsky.social @ftacv.bsky.social · 26/06/2026Goes to your other point about human memory - we have to consciously apply the rules, the LLM follows the rules as a consequence of how it works, so what feels harder for us is easier for them (shades of moravec's paradox) 120
ftacv.bsky.social @ftacv.bsky.social · 26/06/2026I think also they're more likely to be a priori good at coding? Personally I think of them as interpolators over text, consequently the more structured and "rule-bound" the text or task (coding, poetry, translation) the better they will perform 200
ftacv.bsky.social @ftacv.bsky.social · 20/06/2026As an aside, I'd really like someone to make a serious case as to why an AI is going to be scarier than actually existing natural selection. Pointing to the capabilities of alphafold isn't enough! 010
ftacv.bsky.social @ftacv.bsky.social · 16/06/2026giftarticle.ft.com/giftarticle/... (3 clicks only unfortunately)giftarticle.ft.comOpenAI spending hit $34bn last year ahead of planned IPOAudited figures reveal mounting costs from model development, infrastructure and rapid expansion 000
ftacv.bsky.social @ftacv.bsky.social · 12/06/2026Coupled with the fact that the majority of "genetic inheritance of intelligence" research is just cover for a more genteel class of racist, you can see why most people dismiss such discourse out of hand 000
ftacv.bsky.social @ftacv.bsky.social · 12/06/2026The request usually ends with the demands to "acknowledge" this, not the plainly noxious policy implications. Genetic testing (which genes?) Abandoning the educational outcomes of those judged to be morons? 100
ftacv.bsky.social @ftacv.bsky.social · 06/06/2026Yeah, this is what kills it I think. The LLM is recovering the conditional distribution over its training set, and so you can assign a probability of a piece of text being drawn from that distribution, but you can always train to change the shape of it 001
ftacv.bsky.social @ftacv.bsky.social · 03/05/2026Yeah, as ever with Google, it could fall either side of genuine money -grubbing malfeasance or that every measure they use for indexing gets Goodhart'd within three days 100
ftacv.bsky.social @ftacv.bsky.social · 03/05/2026The correction is much more irritating in the UK ("Apple crumble" in there as a control) 110
ftacv.bsky.social @ftacv.bsky.social · 01/05/2026Don't love that the solution to the challenges of 21st century democratic governance appears to be the iron fist, but at least the winners are the ones making solar panels 010
ftacv.bsky.social @ftacv.bsky.social · 01/05/2026Gets better then worse twice imo, great ranking by @leftoutside.bsky.social medium.com/@leftoutside...medium.comEvery Discworld novel ranked definitively by meUnlike the coward Tom Chivers I rank all the Discworld novels 120
ftacv.bsky.social @ftacv.bsky.social · 28/04/2026Ok fun, are there cases of "seat shocks" where you get a change in SP^1/6 and you see a transition to different coalition dynamics? 000
ftacv.bsky.social @ftacv.bsky.social · 21/02/2026The mental model that has been personally most useful to me is them as text interpolators; the more structured the rules of the text (code, poetry, translation) the easier it is for the interpolator to find the best correspondence for the prompt within the latent space of its training dataset 000
ftacv.bsky.social @ftacv.bsky.social · 22/01/2026Sooooo much of this around train travel generally. Split-ticketing, advances, Railcard rules, operators, routing, demand management... It makes people feel people feel special and knowledgeable to be able to game the system but surely it's not a good think for public transport to be so labyrinthine 000
ftacv.bsky.social @ftacv.bsky.social · 26/12/2025As always the problems begin when *other people* use python, a group that includes your past self from two weeks ago 0212
ftacv.bsky.social @ftacv.bsky.social · 26/12/2025👍 Neat way of providing user-alterable defaults, convenient way to report config errors 👎👎👎 Spelunking a massive kwargs object through a series of nested function calls 1120
ftacv.bsky.social @ftacv.bsky.social · 02/11/2025Yes!! Piketty talks about the "machinery of justification" which allows us to permit a particular inequality regime. Ours is the fetishisation of intelligence, and almost every argument about intelligence, artificial or otherwise, buys into this machinery 010
ftacv.bsky.social @ftacv.bsky.social · 13/10/2025This does make me wonder though, if you took the connectome of everyone in Tokyo, would you be able to do something similar? Like a MDL for the set of connections that refer to living in tokyo 030
ftacv.bsky.social @ftacv.bsky.social · 13/10/2025Right, so this is coming in as how many bits to specify a person so log2(Hpop) (gwern.net/death-note-a... ), and I guess people are going "but quantum!" because they're thinking of it as a "how much information to reproduce a person"gwern.net<em>Death Note</em>: L, Anonymity & Eluding EntropyApplied Computer Science: On Murder Considered As STEM Field—using information theory to quantify the magnitude of Light Yagami’s mistakes in <em>Death Note</em> and considering fixes 110
ftacv.bsky.social @ftacv.bsky.social · 13/10/2025Is that because it'd be in the form of a posterior over the parameters and then each instantiation is just a draw from that distribution? Otherwise I don't see how you guarantee how the path from one node to another can be specified by a few bits 110
ftacv.bsky.social @ftacv.bsky.social · 13/10/2025I suppose you'd store this as a graph or a series of embeddings, so then who's the node with the highest degree? Or the person that's the closest to the primary eigenvector. Presumably a Chinese guy that likes beer 130
ftacv.bsky.social @ftacv.bsky.social · 08/10/2025An average of under 65?!?! I thought the SMBC guy was better than this tbh 000
ftacv.bsky.social @ftacv.bsky.social · 03/10/2025If we were productive members of society we wouldn't be here! But thanks, if you do formalise it, my priority for any biological modelling attempt is "what is the actual measurement going to be" 010
ftacv.bsky.social @ftacv.bsky.social · 03/10/2025Would you mind writing downa and sharing the actual maths at some point? 110
ftacv.bsky.social @ftacv.bsky.social · 03/10/2025Incidentally, although the "intuition through natural language -> repeated observation" approach pisses everyone off, biology has eaten many modellers who believed that the problem was a lack of mathematical sophistication amongst biologists. 110
ftacv.bsky.social @ftacv.bsky.social · 03/10/2025Ok, thanks! So, as I understand it your point is that unless you can observe a direct mechanistic improvement, you can't map a property to a causal mechanism because there are too many competing paths? 210
ftacv.bsky.social @ftacv.bsky.social · 03/10/2025Apologies, I can't follow this, would you mind restating your point here? 110
