Publicly available LLM capabilities are indeed not yet at a sufficient level to advance the frontier of AI R&D. OTOH, this is little evidence about what the next generation of systems—or those behind closed doors, known of only within small silos at top labs—might be capable of.

My original statement is: From the capability point of view, AGI models will eventually see humans as humans see ants. Hence, if human wanted AGI to protect the human species, starting from now, human should better protect ecology and other types of living beings, to give a

Works in Progress x ARIA – Sept 18th Hear from three ARIA Programme Directors about their research, and about how ARIA will be recruiting for its next cohort of Directors. RSVP: lu.ma/atl8zd2f

One of the (unlikely, but plausible) “alignment-by-default” scenarios is that future AIs infer that this applies to them too.

Since Shannon-Hartley “bandwidth” is simply the *difference* in Hz between the high-frequency and low-frequency limits of the band (and not, as one might think, to do with the *ratio* of limits), rendering communications spectrum uses on a logarithmic scale is deeply misleading.

If you’re an applied mathematician or theoretical computer scientist interested in Safeguarded AI, you can apply to join our TA1.1 workshop, October 1-2 in Manchester, UK. wg9ud9ekylx.eu.typeform.com/creator-event#…

LLMs that incrementally synthesize formally verifiable reimplementations of legacy code will not just solve cybersecurity, they will also be good for performance. Electron apps can be gradually converted to WASM 🦀

This paper (arxiv.org/abs/2408.16984) is at once a critical review & research agenda. In it we characterize the role of preferences in AI alignment in terms of 4 preferentist theses. We then highlight their limitations, arguing for alternatives that are ripe for further research.

Yudkowsky’s tweet here, being an analysis that recommends meta-meta-analysis, is a meta-meta-meta-analysis. My tweet pointing this out is a meta-meta-meta-meta-analysis. This is a demonstration of Yudkowsky’s Law: the 3rd iterated meta is the highest level with any importance.

Congratulations to our founder and scientific director Yoshua Bengio who has been included on TIME's list TIME100 AI of 2024! #TIME100AI time.com/7012890/yoshua…

Honoured to be named in TIME’s AI100: getting AI right is one of the biggest levers we have for progress, in the UK and the world, and I feel lucky to be working on it time.com/7012825/matt-c…

In the Sinosphere, the concept of “bottom line” (底线) means *lower bound*, whereas in the Anglosphere “bottom line” means *sum*. In my AI safety plan, it is essential to compose the concepts of lower bound and expected value (a sum), reasoning about *bounds on expected values*.