Insert meme with Turkish shooter from the Olympics on how some of the AI agents have way too many knobs to customize the model and context. My favorite part of Amp is that it’s just a text input.

this is how i use cursor rn

JetBrains would have a different trajectory if they stopped breaking their plugin APIs, and if they removed the two-day manual review step for plugin releases. This is low-hanging fruit, they're leaving so much goodwill on the table.

Why is Claude Code a 170mb download? Their ripgrep binaries are 30-40mb per platform while normal ripgrep binaries are 2-3mb. That's 400 terabytes of unnecessary bandwidth per week.

my conversation with a friend

Billions poured into training the best model. Meanwhile, the research button jitters in the Gemini iOS app as you’re typing.

Amp is the best AI agent I've ever used (better than Claude Code). Recommend me AI agents in the comments and I'll try them to see if anything is better

While Amp — Research Preview is great for large, agentic coding sessions… Sometimes you need to make surgical edits to the code yourself. We’ve enabled more fine grained suggestions in Amp Tab, so you can now Tab through a refactor and view each change carefully

Ya’ll Amp — Research Preview is good. Try it. Thank me later. It’s expensive but not as expensive as your time.

I'll be presenting at the Oslo Java meetup on the 13th of August. come by if you are around! meetup.com/javabin/events…

Amp Tab is now 30% faster We worked with Baseten to switch to TensorRT-LLM and KV caching with speculative decoding. The new lookahead decoding also reduces draft tokens and compute per iteration by using improved n-gram candidate selection and variable-length speculations.

I'm thrilled to announce that today I'm joining @ampcode as Builder in Residence. The journey here started all the way back in October 1987 in Colorado Springs, Colorado. I was 10 when my dad burst through the door after work and shouted "Kids! I've got a surprise!" I raced down

New JetBrains integration in the Amp CLI

Tested GPT-5 and Claude Sonnet head-to-head on the same task: run Amp CLI in a native JetBrains view using libghostty terminal emulation. Sonnet wins by a wide margin, although with a non-functional solution. GPT-5 didn't get anywhere after ten failed iterations.

Work account deactivated. Thank you Sourcegraph for an amazing four and a half years working on language indexing, IDE tooling and AI agents. I need to blog about all the cool projects I worked on… Taking a few weeks off before starting a new job that I am very excited about!

Three kids under two, 18 months later wandering and sliding on their own

GPT-5 with thinking in ChatGPT has great vibes based on questions about topics I know well. The answers are concise, hit the core points, and are reasonably accurate. Also, I appreciate the toned down sycophancy. Well done OpenAI.

Simpler Build Tools with Object Oriented Programming by Li Haoyi convinced me to look more seriously into Mill. I’ve always been a reasonably happy sbt user but the scoping rules are unnecessarily complicated. I love the more holistic OOP design in Mill.