Science

This is one of the objections I have when people talk about how useful LLMs can be for literature searches or for any kind of search, really. Even beyond the biases that inevitably (and, in some cases, intentionally) show up, part of a goo…

A good rule of thumb: If the press release (or stenographic news reporting masquerading as journalism) says that an "AI" did something, that's deliberate obfuscation. Software doesn't do anything of its own accord, but folks sure like to…

I have colleagues who use LLMs to find papers to study and cite and (while there are studies showing that those searches can build biases in) I’m sure they can often find work otherwise overlooked. But their architecture also inherently in…

I was talking with a colleague the other day who said that while he’s not sure if AI will change physics research so much as to unemploy the physicists, he’s confident that teaching will always need humans.

None of that is "empirical evidence" unless and until the public (and independent, non-AI-enthralled scientists) can review all inputs (including all training data, prompts, etc) and all actual outputs. >>

It is strange how much LLMs turned out to be able to solve such a wide range of hard problems that would not, instinctively, seem to be problems that a model of human language would be able to solve This is from a Stanford project that le…

Scientist here: It is widely known inside science that this—top-down imposition of a research agenda—leads to mostly rote, boring, non-innovative work. Building an institute to do discovery science is hard: you hire young people, put them…

A point I frequently make when talking about chatbots is that their output only seems to make sense because we are making sense of it. And worse, we can only do that sense-making by projecting a mind behind the text, something we do reflex…

🤖🧠NEW PAPER🧠🤖 (The result of an 8-year project!) LLMs seem very different from symbolic systems. Yet LLMs excel in symbolic domains (e.g., language/code/math). How do they do it? Our finding: LLM representations have implicit symboli…

Unfortunately, any more detailed debunking of each claim takes time (and isn't necessarily possible, without said access, though the mathematicians whose work was stolen have made some headway) >>

Another good rule of thumb: Extraordinary claims require extraordinary evidence. We need not just e.g. the counterexample to an outstanding mathematical conjecture, but also all of the inputs that went into finding it. >>

And as for fantasies of "AI civilizations" independently "hacking" into various companies' networks: it's worth asking -- who is describing it that way an why? In what other field would the companies perpetrating cybercrimes brag about it?…