Jie Fu

Providing abundant verification services, which (could) lead to sentient-friendly; Prev. postdoc with @Yoshua_Bengio @Mila_Quebec | PhD @NUSingapore

[Jie Fu @bigaidream]

@hooshaaii Agree, verification becomes the bottleneck (as argued in https://t.co/cHgLgldWj5). The verification should be as automatic as possible, but this might lead to Goodhart’s Collapse.

[Jie Fu @bigaidream]

🐸 We propose a super efficient approach for mechanistic interpretability: decompose weight matrices from a pretrained LLM into sparse circuit units directly, instead of training a separate sparse representation.

See more in the blog post:
https://t.co/CG1YuV2esr

[Jie Fu @bigaidream]

@feijianghan We discussed it in the paper and use the same exact-SVD component construction as a dense control to isolate the effect of SWD’s sparse read/write connectivity.

[Jie Fu @bigaidream]

@feijianghan Yes, this one is even more directly related to SWD. NaNA decomposes pretrained MLP weights into dense rank-one SVD components and uses them for causal interventions, without training an auxiliary representation.

[Jie Fu @bigaidream]

@feijianghan Both point to the same lesson: use the structure already in the pretrained model before training another interpretability model.

[Jie Fu @bigaidream]

@feijianghan A small circuit can still involve dense incoming and outgoing weights; SWD factorizes those projections into sparse read/write paths and tracks both unit and edge cost, across MLP and attention.

[Dan Hendrycks @hendrycks]

The term “AGI” is currently a vague, moving goalpost.

To ground the discussion, we propose a comprehensive, testable definition of AGI.
Using it, we can quantify progress:
GPT-4 (2023) was 27% of the way to AGI. GPT-5 (2025) is 58%.

Here’s how we define and measure it: 🧵

[Jiayi Geng @JiayiiGeng]

Using LLMs to build AI scientists is all the rage now (e.g., Google’s AI co-scientist [1] and Sakana’s Fully Automated Scientist [2]), but how much do we understand about their core scientific abilities?
We know how LLMs can be vastly useful (solving complex math problems) yet unreliable (counting the number of "R"s in "strawberry" or calculating 9.9 - 9.11) at the same time. Similarly, despite recent advances in applying LLMs to science, are we confident that they can reliably uncover the underlying mechanism of a simple black-box system in a controlled setting?

We study this question in our new preprint: 📢👇
(1/n)

[Jie Fu @bigaidream]

@DavidSKrueger Due to computationally irreducible complexity (https://t.co/GfXokwwLWG), maybe we should not expect too much from interpretable "process", but build a transparent mechanism?

[David Krueger 🦥 ⏸️ ⏹️ ⏪ @DavidSKrueger]

So happy this is out!

IMO a 100% perfect solution to alignment would only cut x-risk by ~50%.

We're on track to put AI in charge of everything.

We need to find a way to defuse the competitive pressures driving this dynamic -- This is critically missing from the public debate!

[Yoshua Bengio @Yoshua_Bengio]

Today, we are publishing the first-ever International AI Safety Report, backed by 30 countries and the OECD, UN, and EU.

It summarises the state of the science on AI capabilities and risks, and how to mitigate those risks. 🧵

Link to full Report: https://t.co/k9ggxL7i66

1/16

[AI Notkilleveryoneism Memes ⏸️ @AISafetyMemes]

Godfather of AI Yoshua Bengio: 🚨🚨🚨 the shoggoths are now trying to escape and we have no idea how to stay in control

Andrew Ng: ha yea we'll figure it out 🤗

Bengio: ok but if we don't u get that everyone dies right??

Andrew Ng: AGI is like a laptop 🤗

Bengio: bro wtf no

[Machine Learning Street Talk @MLStreetTalk]

We discussed AI safety with @Yoshua_Bengio at NeurIPS 2024 - he proposes that we should design agency out of AI systems. Releasing on MLST soon.

[Jeff Dean @JeffDean]

@moderncpp7 @clu_cheng @NeurIPSConf @drfeifei @jiahuiyu @edchi I didn't see the talk, but the images I've seen of the slide seem quite offensive. Such generalizations should have no place in NeurIPS or anywhere else.

[David Krueger 🦥 ⏸️ ⏹️ ⏪ @DavidSKrueger]

They caught this model trying to escape, but they're releasing it anyways. 🙃

[Zekun Wang (ZenMoore) 🔥 @ZenMoore1]

Models and Codes are available at https://t.co/kwULVZrtYS.

[Impact Academy @steadrise_ai]

🔊Advance AI Safety Research & Development: Apply for Global AI Safety Fellowship 2025 🧵

🌟What: The Fellowship is a 3-6 month fully-funded research program for exceptional STEM talent worldwide. (1/10)
...
@aisafetyfellows

[Yoshua Bengio @Yoshua_Bengio]

At the latest @OneYoungWorld Summit, I had the opportunity to discuss the rapid advancements in AI and the importance of implementing both societal and technical safeguards with @CNBC's @TaniaBryer. Listen to the interview: https://t.co/6WVx2r3DK8

[Hanlin Zhang @_hanlin_zhang_]

Critical batch size is crucial for reducing the wall-clock time of large-scale training runs with data parallelism. We find that it depends primarily on data size. 🧵 [1/n]
Paper 📑: https://t.co/LFAPtzRkD9
Blog 📝: https://t.co/tGhR6HDgnE