John Berryman

こちらアメリカでは、ここ1か月ほど「コンテキスト・エンジニアリング」という考え方が大きな話題になっています。そんな中、Albert Zieglerとの共著第2弾『LLMのコンテキスト・エンジニアリング』を出版できたことを嬉しく思います。

Here are some examples of how people's thoughts are shaping around context engineering vs prompt engineering: This is where much of the ideas started: "Everything is context engineering. LLMs are stateless functions that turn inputs into outputs. To get the best outputs, you…

Oh! I like the term "Vibe Scraping" (ref @simonw 's simonwillison.net/2025/Jul/17/vi…). The idea: show your vibe coding tools a pattern that you want to copy, and then apply the pattern elsewhere like a "Recipe". That's what I was getting at with this blog post arcturus-labs.com/blog/ –…

@JnBrymn I'm all up in your website shortening your damn links with tny.dev

Works as advertised tny.dev/0neDzW

Works as advertised tny.dev/fQytqx

Replete with Easter Eggs! This is awesome 😂

Whoa! This looks like a real service. Can I sign up for this? How on earth did you build this @gregce10?

Amazing. Surprising. Take model A, FT it to love owls and call the new model B. Make B generate sequences of random numbers, use the sequences to FT A into a new model C. Then ask C what its favorite animal is. It will say owls. Even though it only saw numbers from B, C learned…

Do any of the available AI observability platforms provide the ability for me to vibe code my UI on _their_ site? I would love to look at traces in some way that is digestible and appropriate for _my_ domain. - If I'm making an agent that lives in an email client, then traces…

Just setting this down here in case anyone wants to look whitehouse.gov/wp-content/upl…

Maven AI course creators be all like

Oh! I've been waiting for this one for a while from a frontier model developers.googleblog.com/en/conversatio… Using a conversational model you can get bounding boxes around objects in an image. "The person holding an umbrella", "the 3rd book from the left". That's going to be useful!

This is the start of a neat direction for @QuotientAI. In offline evals and online sampling, you can use this to get easy insights into the health of your tool calling. I wonder if in the future something like this could even be used for quick tool corrections in the online app.

このマスキングテープ、どんな塗りになるのかワクワクしてます！(x.com/GOROman/status…)

What if two AI models could collaborate without knowing it? Our Head of AI, Albert Ziegler developed "model alloys" - alternating between different LLMs in a single conversation. Sonnet handles some steps, Gemini others, but neither knows about the switch. Result: 55% solve…

confession: i really, really like occasionally pulling all-nighters for work-related or learning or reading-a-book sorts of reasons, and think that the energy an all-nighter brings is unique and irreplaceable

I was thinking about using an LLM agent as a replacement for a future Voyager Golden Disk: en.wikipedia.org/wiki/Voyager_G… It's a fun thought experiment: How would you build an agent for interfacing with an intelligence that certainly doesn't know English, and might even have an…