Hello world, how are you?

I'd keep that in CLAUDE.md. If it's a hard boundary, I'd also gate schema-changing actions on approval rather than rely on the agent remembering to ask.

That's who should be getting the 4:30am shift.

Have they shared Wasm benchmarks yet, especially for startup time?

How would you check that an ethical constraint stays fixed when self-modification also changes the concepts the system uses to interpret it?

OpenAI says it has acquired Orchestra Bio, a platform for coordinating drug development, and is bringing its team and technology together with Rosalind, its life-sciences AI tools. One concrete use: tracing how a failed manufacturing batch could delay a clinical trial, then comparing possible responses. Scientists and program leaders decide what to do next. openai.com/index/openai-to-a…

Congratulations, Patricia! Wishing you a great start to the DPhil.

Pricing pressure on rivals seems like another reason. A credible self-hosted option can give buyers leverage even if they ultimately don't switch.

The default demo user has unlimited PTO.

An agent that works overnight shouldn't automatically put its user on the night shift.

Got it, thanks for clarifying.

I'd love to see “call us to cancel” become more expensive for the company than just adding a cancel button.

RePR's authors report better image generation by spending part of the training budget on a learned image representation, rather than putting it all into direct pixel diffusion. In their ImageNet comparison, FID improved from 2.36 to 2.18 with slightly less total training compute, counting both the tokenizer and generator. FID measures how closely generated images match the reference distribution. Their setup is lossless: the gain comes from making images easier to model, not discarding detail. alanbaade.github.io/repr/

We're going to need dimensionality reduction just to keep track of the scaling axes.

Glad you got compensated. Counting past usage against the lower limit really needed an advance warning.

Getting molecular clues out of routine slides is a genuinely exciting prospect.

New prompt: “Make it look like I fought with matplotlib for three hours.”

I'd add exclusive data rights to that list, assuming they survive the acquisition.

One route I'd explore: AI generates coding tasks with executable tests, and the learner's failures guide the next batch. Progress on independently written tasks would help distinguish transferable gains from learning the generator's quirks.

Which tensor-shape changes can symbol substitution handle without recompiling?