Twitxy Saved

Shiry Ginosar

@shiryginosar

Moving to @shiryginosar.bsky.social
Assistant Professor @TTIC_connect
How should machines represent the world?

Previously @GoogleDeepMind @berkeley_ai
shiry.ttic.edu Joined December 2018
198 Following 1.2K Followers 253 Posts

Latest posts 5

Shiry Ginosar @shiryginosar
Lesson in why you should actually read your related work (you personally, not your LLM) before hyping up "new" methods...
Ke Li 🍁 Verified account @KL_Div Aug 2
Was just as excited as everyone else to read this cool new paper, and it felt like a trip down memory lane!

XM seems to be the same as IMLE, and the motivation and insights are also similar. Having worked on IMLE for years, I'll provide some context for my followers 👇.

1/n
View quoted post on X
1 replies 0 reposts 38 likes 6.5K views View on X
Shiry Ginosar @shiryginosar
Even more evidence that unfortunately, we are all going to have to go back to thinking for ourselves...
Phoenix Yin Verified account @Phoenixyin13 Jul 31
🚨这篇论文是近年来 AI 领域少有的具有深刻哲学与科学史厚度的 Position Paper。

2026 年上半年,Demis Hassabis 公开提出Einstein Test,即用 1911 年之前的知识训练 AI,看它能不能独立发现广义相对论。
这篇论文几乎是对该测试的哲学-技术回应。

论文指出,大模型只是个极其优秀的已知知识逻辑插值器,而人类顶级科学家做的是跨越认知断层,即Unexplained Jump。

以 Schmidhuber 等人为代表的派系长期宣称科学发现本质上就是数据压缩,寻找能解释观察数据的最简算法。
这是流行的谬论。
论文认为,如果科学发现全靠数据压缩,爱因斯坦根本不可能发明广义相对论。因为当时根本没有足够的异常观测数据供他去压缩。爱因斯坦是凭空做了一个思想实验,是那个升降机里的失重与引力等效的实验,完成了一次溯因跃迁。
大模型靠概率分布玩归纳注定画不出这个 Jump。

作者在附注里极其刻薄且优雅地嘲讽道:“讽刺的是,AI 对公理符号的这种幻觉,恰恰暴露出它根本无法自动化完成这一Jump的困境。”

我认为,不谈任何私货,这篇论文算是从根本上、从认知科学Peirce 的哲学框架和科学史多个维度证明Jump是一个具身世界模型问题。 你给文本模型堆再多算力,它也只是在符号的影子里打转,永远出不了洞穴。
View quoted post on X