@CiItay@BenDror4@OriKatz3 ההבדל הכי גדול בין כל מה שהיה עד עכשיו למה שקורה עכשיו (אם לא קרה עדיין) זה שעכשיו אנחנו לא מסוגלים לפרמל שום מבחן שבו בני אדם מנצחים AI.
האם זה אומר שAI יכול להיות יותר טוב מבני אדם בכל עבודה? אנערף. אבל הטענה ההופכית היא היום unfalsifiable. ועד לפני שנתיים שלוש, היא לא הייתה.
Very excited to share the first paper from my postdoc, led by the talented @JieZhang_ETH . This was an extremely fun project with a great group of people 🥸
@iliaishacked In my cohort, to test the algorithm, I tried to simulate a group of movie reviewers. Starting from a few real reviewers I had access to, at $20 it converged to a set of really interesting personalities lol. Correlation with review-aggregator websites was about 0.92.
@iliaishacked It's an evolutionary algorithm that optimizes a population of agents for a task; the agents rank each other at every round and the rank determines degree of mutation of their "profiles" (lower->more mutation). Is this along the lines of what you meant?
@iliaishacked I think it's just still very hard to separate boilerplate from non boilerplate. Nothing so far could save me from writing the agent loop myself, by which I mean having claude code do it.
@iliaishacked Yeah I see.. it's 130k stars, larger than the other large ones (DSPy 30k stars, LlamaIndex 46k). But there are also way too many other smaller-but-not-small ones
Anyway I guess adoption is driven by hype and name recognition just as much as real advantage