EMO’s expert clusters look very different from a traditional MoE: they organize around semantic domains like health, news, politics, & film/music. Traditional MoEs often cluster around surface patterns like prepositions and articles, making selective expert use tougher.

A short post about my favorite adversarial review prompt: blog.fsck.com/2026/05/01/a...

It’s about time we stop infantilizing farmers, they are just like you and me, acting according to their self interests. (I’m just a gardener but married into a farming family)

Are you tired of keeping track of 4 different variants SWE Benchmarks? I am. So I did something kinda dumb -- but at least it was fun -- asking 11 frontier models to blind grade each other's open ended prediction about AI's future. Post at: shimin.io/journal/what... Findings below (given n=1)

here's a mildly provocative take: every valid LLM worry about societal effects is reducible to some approximation of 'it makes what were once costly things trivially easy to do'

The Latest Pelican Bicycle Benchmark result from @simonwillison.net for Opus 4.7 was so shocking that I had to do some follow up experiments. It turns out Opus 4.7 is ....just kinda lazy?? It uses almost no reasoning tokens compare to Qwen, and 40x less than Opus 4.6 shimin.io/journal/clau...

I'd been missing my Pi Agent harness since the Anthropic subscription crack down. Tonight I finally got it back with a local Qwen 3.6 setup locally and it feel great to be reunited with my favorite harness! Thank you @mariozechner.at @mitsuhiko.at for your work on Pi!

@addyosmani.bsky.social's multi agent experience agrees with my own, going from 2 to 5 parallel agents is a quick way to transform from a judge to a ticket usher.

Another banger from @anildash.com : Actually, people love to work hard anildash.com/2026/04/06/p...

Reminder when the only fatigue we felt was about too many new JS frameworks? Those were the days.

Great post from @antirez.bsky.social tracing through the history of the open source movement and positions 'clean room clones' as an extension and not revolution.

AI sycophancy will reap many promising careers in next next few years.

Folks speak of the dangers of AI psychosis, what about the dangers of human psychosis from repeated attempts at convincing LLMs that the Earth is flat?

Found an additional graphic that gets even more of these quotes together. I've kept "I hate myself, I hate clover, and I hate bees" pinned above my desk since I first started studying evolutionary biology as an undergraduate. So relatable to get extremely frustrated with your study system.

Genuinely very impressed by the SVG of a pelican riding a bicycle I just got out of Google's new Gemini 3 Deep Think model simonwillison.net/2026/Feb/12/...