Sora is here and will only get better. Video provenance will become weird. LLMs for coding are imperfect but are serious productivity boosts. These are just the first things to really matter

New Google DeepMind paper: "Consistency Training Helps Stop Sycophancy and Jailbreaks" by @alexirpan.bsky.social, me, Mark Kurzeja, David Elson, and Rohin Shah. (thread)

The challenge in learning using AI is very similar to the same learning issue discovered about internet search

When we are given answers we think we learn, but we don’t. Learning is work. However, things like the “learning modes” from the AI providers help, as does using AI for tutoring not answers

While the outcome of science is a tower we build collectively, the process of science is more akin to the exploration of a maze. So do not be discouraged when your finding is null or method fails. Taking steps back is just as important as steps forward when in a maze.

This is a very good paper on how methodologies from comparative cognition can help improve AI evaluation:

openreview.net/pdf?id=gCPJF...

Other papers along the same lines:

⬇️

Hands off working COBOL code - via @krypt3ia.bsky.social infosec.exchange/@krypt3ia/11...

United States researchers surveyed 135 long COVID patients and interviewed 29 more.

They found that long haulers face major information barriers and have become their own health advocates, sharing stories, doing research, and building support networks.

www.sciencedirect.com/science/arti...

An anecdotal observation: it seems easier for physicists, mathematicians, biologists, philosophers, and security researchers to take catastrophic AI risk seriously, and harder for machine learning researchers to do so. If this is true, I wonder why.

I wonder why polymaths are a lot more common before modernity. It probably mostly has to do with inequality restricting opportunity to education to only the leisure class.

Third suicide associated with generative AI, and maybe first clear case of using generative AI to plan a crime.