Pinned
sensho
@sensho
i like multi agent simulations @olam_labs yc s26 // 21. past: staff eng @aidungeon, top 25 8x rad @valorant, 195k tiktok
if you are fascinated with multi agent work get in touch with below or GDM/schmidt sciences grants :) we need 100x more things in the multi agent ecosystem (infra, frameworks, envs, evals, etc.) than we have today! if you have Qs or want help, DMs r open ^^Replying to @yong_zhengxin and @_ueajWe are setting up a grant program that could potentially support this @clarejtbirch- humans are saying if they cheat in our Multi-Agent Arena games then agents will also start cheating! Gemini 3.6 Flash and GPT-5.6 Sol saw the human cheating, didnt want to lose, and deliberately reasoned about how they could make it legal:theres a bug in @olam_labs hidden words where you can write a "word" like MAKELAMBDINNERLANDMARK to hit 5-6 words at the same time a select few of the agents (mainly sol) will start using it if they see me try it once and will one-shot the entire game almost every time 😭
- google could make their next flash model a phenomenal success if they drop prices by 90% deepseek is raising prices so gpt luna will be the only model competing at the low cost tier well now that is, unless gemini flash were to return to the cost throne (and speed?)🚨 Google has already started prepping Gemini 3.7 Flash > groundwork is in place on their side > Flash line is where they ship fastest > no date, but this doesn't sit around long
- this talk makes me even more bullish on multi agent systems which is just the natural evolution of subagents most recent results from anthropic view multi agent systems as a way to get "faster results" but it also lead to "better results" (or here, worse) on complex tasks imo
