November 5, 2025
Lab Coats vs Chatbots
Kosmos: An AI Scientist for Autonomous Discovery
AI Scientist Works 12 Hours Straight; Hype vs 'Just Rehash'
TLDR: Kosmos says it can run 12 hours, read 1,500 papers, and produce cited discoveries with 79.4% accuracy. Commenters split between “four real new findings” hype and skepticism that it merely rehashes pre‑baked datasets, with extra side‑eye at its rule‑like “world model” and jokes about the robot intern doing all‑nighters.
Kosmos is pitched as an AI “scientist” that grinds for up to 12 hours, reading 1,500 papers, running 42,000 lines of code, and spitting out fully cited reports. Fans cheered claims of “six months of research” packed into a 20‑cycle run and seven cross‑field discoveries. Then the comments erupted: skeptics like isuguitar121 said most “discoveries” are just Kosmos confirming what datasets were built to prove; boosters like adt waved receipts, pointing to four novel findings. The big spark: its “world model,” a shared map between the code bot and the literature bot.
Andy99 warned this sounds like rulebook AI in a new outfit, predicting we’ll hit the same old wall of hand‑crafted rules. Others joked about the bot pulling all‑nighters: “My PhD advisor could never.” Leptons dropped the meme‑y mic: “How much slop will it discover?” Still, Kosmos citing everything and scoring 79.4% accuracy earned cautious respect. The real drama: discovery engine or super‑fast copy editor? If it can find new patterns without cherry‑picked data, labs might trade coffee for code; if not, it’s just a very diligent robot intern.
Key Points
- •Kosmos is an AI system that automates data-driven scientific discovery through iterative cycles of analysis, literature search, and hypothesis generation.
- •A structured world model enables coherent information sharing between a data analysis agent and a literature search agent over long horizons.
- •In typical runs, Kosmos executes about 42,000 lines of code, reads around 1,500 papers, and operates for up to 12 hours across 200 agent rollouts.
- •Independent scientists found 79.4% of statements in Kosmos-generated reports to be accurate, with all statements cited to code or primary literature.
- •Seven discoveries were highlighted; three reproduce findings from preprinted or unpublished work not accessed at runtime, and four are novel contributions.