Views my own, not my employers.
cdt
Karma: 284
drift-diffusion threshold
I don’t think this is a typical phrase, can you cite this? Perhaps you mean the drift barrier hypothesis instead.
I have no idea what innovation looks like in PRS. ML is all good but the bottleneck is (presumably) phenotyping which none of these companies want to do.
Google, PacBio and ONT all produce software for e.g. base/variant-calling, but it’s not obvious to me why they would want to go beyond that. What’s the natural next step? Variant effect prediction? Really tough problem with no natural commercialisation link (in their product, at least).
Strongly upvoted for a series of concrete, object-level arguments with reference to literature. Unclear why people perceive this comment as emotional. The use of “lizardman constant” here seems to conflate many sources of measurement error in surveys in a way that it didn’t upon the original definition.
Strong upvote for your courage.
These points do not seem like the same points as the other plot? (Look at the x-y of the orange dots).
This confusion comes about because natural selection has no mechanism to maintain variation. Equivalently, gradient descent can only work with the data provided or in other words it has no “proposal” step like Gibbs sampling or MCMC. So the idea that gradient descent and natural selection are the same feels intuitive to me. (Caveat being frequency-dependent selection.)
It is also known that some models of evolutionary game theory recover Fisher’s theorem of natural selection as a consequence of the replicator equation (a model of natural selection) as a gradient flow, see this arxiv paper. (Might have bungled the explanation on this one, so take with some salt.)
This may go beyond aphantasia and to something called Severely Deficient Autobiographical Memory (SDAM). It looks like our understanding of SDAM is still in its infancy though.
Sometimes I think about the arguments about LLMs being conscious/not conscious because they have X or don’t have Y capability, and then I think about this. I wonder sometimes if they knew what it was like, people would consider this “less than human”. Quite like your last paragraph here.
It is amazing that a paper that is essentially just a vaguer form of Hamilton’s Rule only cites him once.
As it stands, I think the table is incorrect but “right” in the sense that it really depends on which random constants you assign to these calculations, and I can’t see find any evidence of a careful selection in the paper or in his code.
Paper here: https://eigenism.org/paper.pdf
However, traditional ML approaches are usually preferred over the deep learning models associated with modern AI. AlphaFold aside, there aren’t too many situations where you’d pick a stack of transformers over a traditional method
Is this really true? Consider:
DeepConsensus and other deep variant callers
ONT base-calling and related research
Whatever AlphaGenome and Evo is/are doing
More traditional-style annotation toolsDefinitely feels like there’s a push towards this, although many examples I am more familiar with are model-based so synthetic data approaches.
What would be your bar from banning someone from Lighthaven fully? Feel free to ignore if you would rather not be pinned to it, or if it you feel it may encourage bad behaviour.
Ah I see I have been a little sloppy with my language—mea culpa
The extent to which traits and phylogenies are correlated is an open research question, see this wikipedia article. But aspects of biology that are unique to phylogenies such as diversification interact with traits in complex ways. The SSE models are a good introduction to the methodology of this (background here).
I don’t know how to attach these ideas to linguistics because I can’t think of a good concrete example.
As an aside, you also say that
For example, a notion of difficulty of the variants can be used (with care) to infer which variant is more likely to be original
There are things of this kind in historical linguistics, though less subtle and complex, such as analogy. But nothing that I know of in phylogenetics.But this is very well understood in phylogenetics. This is the basis of codon models and the “maximum likelihood school” of phylogenetic modelling. You can see this by looking at IQ-TREE (a modern phylogenetic inference tool): https://iqtree.github.io/doc/Substitution-Models
Just throwing this out there as an idle thought when reading this post* - I wonder if the reason linguistic phylogenetics and biological phylogenetics differ is that the phylogeny and the biological traits of the underlying species are often correlated with each other. I don’t know if that’s true for linguistics, and indeed I’m not sure what traits you are capable of identifying in a text or language that would affect the phylogenetic reconstruction itself.
Thanks for this info—very useful datapoint, I have strong-upvoted.
I think you are right that “humans that want things” will always have more things than “humans that don’t want things” (or even replace human with X, see this talk by Joe Carlsmith). You must be much more explicit drawing a line between this and “post-scarcity”—which frame of post-scarcity is it that you are disagreeing with here? You seem half to be making an argument about the inherent nature of competition, and half an argument about the desire of humans for status, which leads to an argument about neither.
One thing I’m curious about is given SpaceX (containing xAI), Anthropic and OpenAI all going public at a similar time, can the public market financially support all these companies at once? Does this make the competitive effect on culture stronger or weaker? I have no financial expertise so I don’t know what to think about it.
I agree that adjusting self-reports by the estimated discrepancy in effect sizes informed by the METR study is the way to go here—so maybe somewhere up to 1.5x is more likely. 4x speedup would require some really convincing evidence.
Was the date of publication of the RSPv3 or the Risk Report fixed? It’s entirely possible that the internal deployment of Mythos was pushed back to elide speculation.
Okay, I should probably caveat this by saying that Google does have a phenotyping service in the sense of Google Health and related products, but I’m not aware of any work in this area and I’m not sure how much things like heart rate/BMI etc. are usable in genomic research. Or that they can even be released to be used in this way.