well, lo, people have built a benchmark using a task called multi-armed bandit, where success depends on iteratively grasping the relative odds of payouts from different choices, and LLMs are (generally) shit at it arxiv.org/pdf/2403.15371 /
Who's Who of AI
tweety fish
at different times:
haxx0r in cDc
PhD in using people for ML and vice versa
"theory of mind for autonomous cars" startup guy would you believe, went kablooie
at present: newsletter -- buttondown.email/apperceptive
&c music, politics, nonsense
What they're sharing
arxiv.org
thetransmitter.org
arxiv.org
bloomberg.com
techloy.com
View on Bluesky →
Articles & links
this, posted elsewhere in these sprawling threads, is very interesting on the topic of what emotional representations exist in frontier models and what they might mean: www.thetransmitter.org/emotion/what...
as promised here is the paper comparing LLMs to human results; I haven't gone through it carefully enough to vouch for the methodology and for various reasons I have my suspicions about how robust it'll be but it's certainly conceptually interesting: arxiv.org/pdf/2505.09901
quoted sentence paraphrased from today's Matt Levine which links www.bloomberg.com/news/article... and robinhood.com/us/en/newsro...
tweety fish reposted
Whether you’re a denialist or a booster or just someone trying to be objective, Salesforce is going to give us some pretty clear answers as to whether radically shifting resources from human coders to AI will work.
Their own posts
Recent commentary
Last night it occured to me to wond er if LLMs were any good at gambling tasks. This is important not because it'd be funny for LLMs to gamble but because gambling tasks get used to measure human decision-making under risk /
listen, I try not to prejudge this stuff too hard because I think it's important for me, at least, to have my skepticism rooted in up-to-date knowledge but "Robinhood offers in-app agentic trading to its users" is one hell of an alarming sentence