AI Agents can now automate scientific analyses that took years in mere hours.
Excited to announce our new pre-print on automating the data analysis process using Artificial Intelligence in High Energy Physics and Particle Physics.
With @softcollinear, Andrzej Novak, Dolores Garcia, and Phil Harris we show that just furnishing context (#JFC) with a thorough methodology is enough for AI agents to autonomously perform a HEP analysis end-to-end, complete with a planning, exploration, execution, statistical analysis, all performed by AI agents. Each step reviewed by more AI agents and looped as long as necessary.
We discuss prospects on how the field should evolve given the increasing power of AI models. Attached in the appendices/linked on GitHub are not only the review-ready analysis notes, but we also publish the full, reproducible codebase for each analysis with tracked history, setting a benchmark for openness and replicability in HEP.
Check out the pre-print: https://t.co/A1U8pv8wUL itβs a fun read and an exciting time to be in #AIforScience!
@MIT@MIT_Physics@AnthropicAI@claudeai@iaifi_news@CERN
#AIRevolution #ClaudeCode #AIAgents
AI Agents Can Already Autonomously Perform Experimental High Energy Physics
Eric A. Moreno, Samuel Bright-Thonney, Andrzej Novak, Dolores Garcia, Philip Harris
https://t.co/TmuhTb3wAA [πππ-ππ‘ ππ.π°πΈ ππ.π»πΆ]
@kchonyc Haha!
That's what I told Max Tegmark: the probability that one of his MIT students is a genocidal psychopath is non-zero, hence he should stop teaching.
Brings me back to a conversation I had with @vlimant and @EricMetodiev at @NeurIPSConf 2019 - We need an AI scientist. One that is curious about the world, able to form conjectures, and test them.
Attended a talk by @demishassabis from @DeepMind today where he discussed AI as a tool for scientific discovery (#AlphaFold etc.). After, posed two questions to him:
1) Humans are innately curious, creative, and explorative beings. If you give a baby a stack of blocks, they will
- Highly parallelizable & extremely fast
- Energy efficient
- Temporally based
- Analog compute engines
That is my take - does anybody have any others? @demishassabis seemed interested but pointed out that these solutions are still very far away. Maybe so! I hope not!
(1/3) Super excited! Our new paper in @MLSTjournal on detecting #gravitationalwaves without ever telling the ML model what a GW looks like. This builds on work by @ElenaCuoco et. al by introducing recurrent autoencoders.
(2/3) We are currently exploring more complex versions of this architecture (with larger dimension anomaly/loss spaces) to possibly detect exotic GWs that aren't currently identified by LIGO/Virgo! #Semisupervised looks like the way to go!