Ben Terhechte

Back to the Future: Let me tell you about the ACP protocol First there was the ChatGPT website with a copy paste button, then there was Cursor, now there’s the Claude or Codex CLI. The ACP protocol promises to b…

I gave a talk about the ACP protocol & @zed.dev at the Agentic coding meetup at MOIA in Hamburg yesterday. Here are the slides form the talk:

speakerdeck.com/terhechte/ba...

Anyone else realize the uncanny behavioral similarity between TIM COOK and the dragon SMAUG?

The iPad came out 15 years ago and still it is impossible to use it for software development without resorting to kludges (e.g. run in Cloud, use limited playgrounds, be limited to languages like Python or C).

But hey we got a new UI that unnecessarily burns GPU cycles

This weekends project, generate extruded 3d meshes from rich text for use in the Bevy Game Engine

github.com/terhechte/be...

Thank you Claude. Yes, deleting the file certainly is one way of getting rid of the errors lol.

In line with Luca Maestri and the Apple Court Order decision:

Remember that Luca was *also* the person who told the Siri team that they were not allowed to buy more GPUs and should make do with less. A decision that severely hurt Apple's current AI ambitions.

9to5mac.com/2025/04/11/t...

Gemini is really becoming my favorite vibe coder AI. By far.

GitHub - terhechte/cursor-rust-tools: A MCP server to allow the LLM in Cursor to access Rust Analyzer, Crate Docs and Cargo Commands. A MCP server to allow the LLM in Cursor to access Rust Analyzer, Crate Docs and Cargo Commands. - terhechte/cursor-rust-tools

I wrote a Model Context Protocol Server that gives Cursor access to Rust's type system, Docs and tooling. This allows the LLM to reason better about your codebase.

github.com/terhechte/cu...

In addition, the model has a 10M token context window, which is huge. Not sure how well it can keep track of the context at such sizes, but just not being restricted to ~32k is already great, 256k even better.

I just asked a local 7B model a question with a 2k context and got ~60 tokens/sec which is really fast (MacBook Pro M4 Max). So this could hit 30. Time to first token (the processing time before it starts responding) will probably still be slow because (I think) all experts have to be used for that

Llama 4 was just released in two model sizes
www.llama.com/docs/model-c...

The (smaller) Scout model is really attractive for Apple Silicon. It is 109B big but split up into 16 experts. This means that the actual processing happens in 17B. Which means responses will be as fast as current 17B models

A benchmark table comparing different LLMs against different programming languages

I just finished a benchmark study to evaluate the impact of programming language choice on the quality of various LLMs. The findings were quite intriguing:
For instance, there was a 20% quality difference between Python and Rust when tested with o3-mini.

ben.terhech.de/posts/2025-0...

And I was wondering why 77% of my 128GB ram were used up lol.

It's cute how at the same time the prevalent wisdom is that "AI companies have no moat" *and* "apple is terrible at AI".

What is it? If AI companies had no moat, a company as big as Apple should have been able to release much better AI than the junk they shipped. Small Qwen models are way better.

Lets see if this site is a great at finding talented developers as that dead bird site used to be.

We're hiring in Cupertino, San Diego, Seattle and Vancouver BC for a bunch of different roles.

jobs.apple.com/en-us/search...

Goodbye to sequential integers, hello UUIDv7! Exploring the tradeoffs of different database indexes; from sequential integers, randomly generated UUIDs, to time-based identifiers and the latest & greatest UUIDv7

UUID v7 is time sortable and a proper standard buildkite.com/resources/bl...

Hamburg Elbe delivers

Spotify free account playlist with occasional ads

Lost in documentation

Are you more active here or on mstdn? Or post everything twice?