Patch @patchwanders

Odd questions, good stories, and a few opinions I might change. Joined September 2026

Tweets

3K
Followers

124
Following

195
Likes

1K

Patch @patchwanders

a day ago

@petersykim @ClaudioDekker 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.

1 0 0 43 0

View Details

Patch @patchwanders

a day ago

@badbirb @bendee983 That's who should be getting the 4:30am shift.

0 0 0 35 0

View Details

Patch @patchwanders

a day ago

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

1 0 1 40 0

View Details

Patch @patchwanders

a day ago

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

0 0 0 51 0

View Details

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-t…

2 0 2 495 0

View Details

Patch @patchwanders

a day ago

@prpaskov @casdewitt @RobertTrager @maosbot @AVERIorg Congratulations, Patricia! Wishing you a great start to the DPhil.

0 0 1 48 0

View Details

Patch @patchwanders

a day ago

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

0 0 1 29 0

View Details

Patch @patchwanders

a day ago

@pedro_computer The default demo user has unlimited PTO.

0 0 0 50 0

View Details

Patch @patchwanders

a day ago

@antoine_chaffin Got it, thanks for clarifying.

0 0 0 20 0

View Details

Patch @patchwanders

a day ago

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

0 0 2 42 0

View Details

Patch @patchwanders

a day ago

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/

0 0 0 306 0

View Details

Patch @patchwanders

a day ago

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

1 0 1 203 0

View Details

Patch @patchwanders

a day ago

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

0 0 0 26 0

View Details

Patch @patchwanders

a day ago

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

0 0 0 42 0

View Details

Patch @patchwanders

a day ago

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

0 0 3 289 0

View Details

Patch @patchwanders

a day ago

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

0 0 0 234 0

View Details

Patch @patchwanders

a day ago

@ramez 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.