Joe @twitskeptic@qoto.org

@tb Bull market for anodyne corporate PR statements!

Remarkable story by a student reporter at Stanford that reveals potential scientific fraud by the current Stanford president. https://stanforddaily.com/2023/02/17/internal-review-found-falsified-data-in-stanford-presidents-alzheimers-research-colleagues-allege/

@mcdanlj @JamesGleick That might be true if Elon was forcing everyone to abandon SMS based 2FA, but Twitter Blue users are exempt. Like everything Elon, it makes no sense.

@ct_bergstrom FWIW, since I'm in a contentious mood this afternoon, I disagree that hallucinations are necessarily pathological. A couple of counter examples are sleep onset/awakening hallucinations or children who are convinced that they have an imaginary friend.

That doesn't mean that I'm happy about applying "hallucinations" to LLMs since it incorrectly implies that there's a consciousness in there that's misperceiving the external state of the world. And "bullshitting" also implies intent, if not to deceive, then to enhance status or impress.

Another term is required. Blathering is the best I can do but I'm hoping that someone else can do better.

@ct_bergstrom
Perhaps Hanlon's razor should be applied here.

As you mentioned, "hallucinate" is an established AI term. Engineers inexperienced in public communications use the jargon when speaking to reporters and the reporters parrot it back in their headlines without understanding the context. This is not unusual - another example that comes to mind is "heritability" in biology.

So, at least for now, I'll give them the benefit of the doubt.

@ct_bergstrom Interesting - now it refuses to play. Maybe ChatGPT's Mom is monitoring Mastodon.

@ct_bergstrom But I digress! For those "theory of mind" proponents.

@ct_bergstrom Meanwhile, it looks like that a right wing organization can train a GPT-3 model to spew out right wing dogma for only $300. Truly scary stuff. https://twitter.com/davidrozado/status/1626151249795235840

@Riedl Panic at falling behind overcoming sensible consideration of reputational risk from bad LLM summaries. Right out of the Elon school of management.

@ct_bergstrom @moultano Ugh, not small LLM - small transformer based language model.

@ct_bergstrom @moultano I'm starting to have doubts about the idea tha LLMs are "stochastic parrots" that can't generalize after watching a short talk from Francois Charton of Meta at NeurIPS 2022.

TL;DR - he trained a small LLM to learn how to diagonalize matrices using only triplets of the similarity transform. No hallucinations were observed.

The talk was "Leveraging Maths to Understand Transformers" https://neurips.cc/virtual/2022/workshop/50015#wse-detail-63846

@ct_bergstrom Great shots but hope you have insulated waders.