Rex Kerr

Neuroscientist specializing in behavior and microscopy; aging researcher; and expert in Scala and high performance programming. How are these things related???

[Rex Kerr @_ichoran_]

@FoilingMothra @AmandaAskell @elonmusk @WSJ That's not adaptive in the evolutionary sense. Evolution doesn't care about "cost plus", only offspring. Quality matters! Pooling resources into fewer children can be adaptive; cooperative societies can do this naturally. But pro-adult child-hostile actions = fail at evolution.

[Rex Kerr @_ichoran_]

@AmandaAskell @elonmusk @WSJ Childless members of highly social species, if behaving adaptively (this is not guaranteed!), _widen_ the scope of their stake in the future because there is less (genetic) reason to engage in selfish non-cooperative behavior to aid their offspring at the expense of others'.

[Rex Kerr @_ichoran_]

@davidad What wins depends on the exploration/exploitation landscape. People tend to make the Donald Hoffman mistake of fixing the landscape--usually favoring exploitation--and then "discovering" that highly optimized non-veridical solutions win. Veridical models win when reality varies.

[Rex Kerr @_ichoran_]

@TBen_Yassine @mbeisen It would be strikingly odd if there were no way to use money to help prepare students to excel in college, so the first step would be causal modeling to determine which correlations reflect likely-potential-at-time-of-entry and which reflect money-makes-me-look-better-than-I-am.

[Rex Kerr @_ichoran_]

@chubbicat4641 @davidad You don't need any sentences at all. You can use shapes, symbols, numbers, whatever. That we happen to like patterning correspondences of states of affairs by using language does not entail that it has anything intrinsically to do with sentence structure.

[Rex Kerr @_ichoran_]

@gkossakowski In particular, even if nominally the dimensionality of text could be adequate for the task, that it is adequate when adhering to ordinary language use is far from clear. Comprehensive summaries are hard; it's part of why people use other methods (shorthand, mind-maps, etc.).

[Rex Kerr @_ichoran_]

@gkossakowski I like the direction, but in order to move on from embeddings as you suggest, one needs an inverse mapping from the embedding layer(s) to tokens, and it's not clear to me that even if trained, this can be done in a sufficiently lossless way for large texts e.g. a research paper.

[Rex Kerr @_ichoran_]

@davidad In a situation with changing context and/or extent of deployment, a pastiche can have vastly more serious consequences than the original ever had in its context and scale of deployment. It wouldn't be brilliantly Machiavellian, but it could still be existentially dangerous.

[Bertrand Meyer @Bertrand_Meyer]

We lost a titan of programming languages, programming methodology, software engineering and hardware design. Niklaus Wirth passed away on the first of January. We mourn a pioneer, colleague, mentor and friend.

[Rex Kerr @_ichoran_]

@cshperspectives @steinly0 I side with Sydney Brenner on this one. You'll find the quote if you use this search string: brenner probably in that order

[Rex Kerr @_ichoran_]

@lab_oren @Nancy_Kanwisher Which straw is considered the last one, and which were not, sends a powerful signal about what is considered important.

[Rex Kerr @_ichoran_]

@klaehnr @jdegoes If people do (1), (2), and (3), then effective, benevolent leaders should emerge from the process of voting. (Might need voting system reform too--IRV or Condorcet? Something not too easily gamed.) There is still the question of what to do when no such leader is on the ballot.

[Rex Kerr @_ichoran_]

@davidad Here's an example (used standard web-interface ChatGPT-3.5 this time), where it gets a binary question correct by making an even number of errors (2n>0):

[Rex Kerr @_ichoran_]

@davidad Try asking LLMs about the relative position of the stars and stripes in the U.S. flag. Most of the answers they give me are badly wrong. I'm not surprised about tic-tac-toe! You have to talk a *lot* about a spatial problem before linguistic computations on it become useful.

[Rex Kerr @_ichoran_]

@mbeisen Do you need a characteristic spatial scale of the lumpiness, or just to detect nonstationarity at a particular window size? Autocorrelation would be my first try (e.g. Moran's I). If that fails, then spectral density analysis of standard deviation.

[Rex Kerr @_ichoran_]

@norabelrose @robinhanson @anderssandberg From a first-person perspective, it's obvious: a pain state has qualia unlike a non-pain state. From a third-person perspective, pain states tend to be associated with evocation of homeostatis-restoration mechanisms (like inflammation), if we can hold Descartes' demon at bay.

[Rex Kerr @_ichoran_]

@mbeisen Maybe it's good if people in science are confused? I've written and received multiple reviews saying, "You claim P, but that seems impossible given Q, which you don't talk about." Validation is usually impractical, so review is not primarily about that. But let's try when we can!

[Rex Kerr @_ichoran_]

@mbeisen Oh, well, that reminds me of 6, 7, 9, 8, 6, ..., but I don't recognize the double 8 in yours.

[Rex Kerr @_ichoran_]

@mbeisen Yes, I also see the stylized C. elegans with a slightly-too-large egg. Decent, but I don't think they'll win the International C. Elegans Conference best artwork prize.

[Rex Kerr @_ichoran_]

@jdegoes @soronpo Inlining plus type unions render this a largely solvable problem for Scala 3 (within JVM limits). Rust just makes it easy. But Scala makes easy some abstractions that are tricky in Rust. Both are high-productivity and fast; Scala just trends to the former and Rust the latter.