Bertil Chapuis

Loves hacking and collaborating at the intersection of several disciplines. Building in public.

🚧 🏗️ ⚒️

[Bertil Chapuis @bchapuis]

Started exploring extensively how AI-assisted coding changes web development in 2025 and it escalated into an open source visual automation platform built on @cloudflare called Dafthunk.

2026 plan: build more, share more 🏗️

[Bertil Chapuis @bchapuis]

@grok thoughts? Is one of these a long-awaited load-bearing emoji?

[Bertil Chapuis @bchapuis]

Can't contain my excitement for these 8 new emoji. Worth every one of the 7150000000 bytes

[Leonard @LeonardCser]

@mitchellh been building cursortab.nvim for this: https://t.co/vWS2jPD3ml

sweep-next-edit-1.5B with speculative decoding would work the best on M4 max with llama.cpp

for better but slower completions zeta-2 is the best local model

[Bertil Chapuis @bchapuis]

@francoisfleuret Teaching students, some of them are surprisingly good with agents. To me, drive and systems thinking seem to matter more than raw experience. TBH I'm still not sure which traits matter most

[Bertil Chapuis @bchapuis]

@lucatac0 Quick question: is it possible to get the API schemas of the models programatically? I'm introspecting schemas on replicate and this would be very useful on AI Gateway

[Bertil Chapuis @bchapuis]

I'm currently experimenting with cloudflare durable objects and starting to think they're a new kind of primitive for scaling multitenant applications. Every tenant gets compute and an isolated 10GB of storage, scales to zero when idle, billed per-usage. Does anyone know if an open source alternative already exists?

[Bertil Chapuis @bchapuis]

Side project update: my globe built with Claude in C/WebGPU/WASM can now serve and render custom tiles that include DEM-derived meshes and vector layers from Overture Maps... YOLO mode can brings you quite far

[Bertil Chapuis @bchapuis]

@lemire Someone should write "How Societies Actually Work" for curious and competent people (who start by acknowledging a knowledge gap). Because those traits become handicaps when everyone around you is trained to project confidence on half-understood topics

[Bertil Chapuis @bchapuis]

More and more unsolicited emails from OpenClaw bots in recruiting mode, scraping GitHub for developers interested in agentic ai. Is this really the current state of automation?

[Christos Tzamos @ChristosTzamos]

1/4 LLMs solve research grade math problems but struggle with basic calculations. We bridge this gap by turning them to computers.

We built a computer INSIDE a transformer that can run programs for millions of steps in seconds solving even the hardest Sudokus with 100% accuracy

[Bertil Chapuis @bchapuis]

@morganlinton Meanwhile in Lausanne https://t.co/vijU9dyQDa

[Bertil Chapuis @bchapuis]

🎉 Just got accepted into @Cloudflare's Worker Loader API beta.

Codemode lets the LLM skip the one-tool-at-a-time loop and write code instead. The integration in Dafthunk was just straightforward and my agentic node can now call tools with code.

Can't wait to get accepted in the email service beta 😘

[dina kozlov 🐀 @dinasaur_404]

Code More via Dynamic Worker Loader for the win!!

@thomasgauvin you’re up next with email 😛

[Bertil Chapuis @bchapuis]

@dinasaur_404 @thomasgauvin These products and documentation are a real breath of fresh air and fun to work with! Thank you

[Bertil Chapuis @bchapuis]

🎉 Just got accepted into @Cloudflare's Worker Loader API beta.

Codemode lets the LLM skip the one-tool-at-a-time loop and write code instead. The integration in Dafthunk was just straightforward and my agentic node can now call tools with code.

Can't wait to get accepted in the email service beta 😘

[Bertil Chapuis @bchapuis]

Interesting take by @stefanomaz

> I fully expect that “reading code” will be as common as people looking at the assembly spit out by their compilers.

https://t.co/5JBVr1utAq

Requires a bit of forward thinking but probably true.

[Bertil Chapuis @bchapuis]

Same experience.
Something I want to try earlier in the curriculum:
Teach the basics, give students a problem and an AI-quality solution, then tell them vaguely what's potentially achievable with the problem, domain, and hardware knowledge acquired during the course.
Their job is to get there, using AI however they want.

[Bertil Chapuis @bchapuis]

https://t.co/12kuNv6kMM

[Bertil Chapuis @bchapuis]

When @HamelHusain puts out skills for doing AI evals with AI, it's time to pay attention. His stuff is grounded in real work and applications.

https://t.co/PyLXPXhtgw