Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Can AI Trained on Historical Data Rediscover Relativity?

AI summary

A Nature feature by Philip Ball examines experiments using historical-data language models to test whether AI can recreate scientific breakthroughs. At the India AI Summit, Google DeepMind co-founder Demis Hassabis proposed training a model on knowledge available through 1911 and asking whether it could independently rediscover general relativity, calling this a useful test of AGI. Preliminary experiments produced mixed results. Michael Hla's Machina Mirabilis model showed occasional flashes of intuition, such as describing light as discrete pulses, but often failed to understand the physics it invoked. Other researchers reported difficulty separating correct from incorrect theories, judging which ideas were worth testing and preventing information from beyond the historical cutoff from entering the training data.

In-depth insight

The experiments provide a concrete test of whether AI can move beyond statistical prediction toward historically constrained scientific discovery, while their mixed results show that plausible ideas alone do not demonstrate sound physical reasoning or an ability to identify meaningful theories.

本篇《自然》长文共5937字,干货满满,预计阅读时间18分钟,时间不够建议可以先“浮窗”或者收藏哦。原文作者:Philip Ball科学家正在探究以历史数据训练的语言模型能否重现创造性突破。1915年,阿尔伯特·爱因斯坦(Albert Einstein)提出了广义相对论,改变了我们对物理世界基本结构的认识。这......

本文来自微信公众号: 自然系列 ,作者:Nature Portfolio

本篇《自然》长文共5937字,干货满满,预计阅读时间18分钟,时间不够建议可以先“浮窗”或者收藏哦。

原文作者:Philip Ball

科学家正在探究以历史数据训练的语言模型能否重现创造性突破。

#Google DeepMind#Truthful AI#Massachusetts Institute of Technology#Machina Mirabilis#Ranke-4B#AlphaFold#OpenAI#ETH Zurich#University of Zurich#Technion

Original · 12 min readRead the original
Group chat

No comments yet. Start the conversation.