Thibaut Mattio
@tmattio_
Followers
475
Following
983
Media
6
Statuses
255
Building Raven, a machine learning ecosystem for OCaml. Former director of engineering @tarides_, building developer experiences for the OCaml Platform.
Joined April 2013
I'm planning the roadmap for the initial release, would love to hear your thought and feature wishes!.
Anyone using Claude Code or other coding agents with OCaml interested in testing ocaml-mcp-server, an MCP server that gives access to OCaml toolchains, (build status, merlin queries, etc.)?.
1
1
2
Excited to announce that , after almost 6 months of development, the first alpha of Raven is out!. Read the announcement on Discuss, and share your feedback! Link ↓.
2
1
11
Impressive results! Incidently, very relevant to the philosophy behind Raven: in the AI era, the incentives to maintain a zero-dependency system are very high, you can ask an army of agents to optimize your code. That's only possible if you have control over all the code.
New blog post: We've never enjoyed working on Kernels more than this. We have some very fast AI-generated kernels with a simple multi-agent system. They're running close to or even surpassing Pytorch shipped kernels. (1/6). [🔗 link in final post]
1
3
8
Why would you choose Raven vs, say Mojo?.
@TheOneRealPK Mojo = "Python but faster". Raven = "ML in a language built for production systems, without compromising on prototyping speed". You wouldn't build your entire production infrastructure in Mojo, for the same reason you wouldn't in Python (cough. ). Raven's moonshot is to.
0
0
1
Did you hear the news? Raven alpha is almost here - time to start porting your projects, and share feedback!.
As the first alpha version of RavenML is practically out, I'm currently working on refactoring and migrating the entire codebase of SoundML to RavenML. I'm hopping that this significant change to a more modern scientific computing library will solve tons of issues I had wth Owl.
1
0
7
Another project I just open-sourced: Mosaic, a framework to build TUI in OCaml. Raven + OCaml MCP + Mosaic. You see where we're going, right? 👀.
1
1
23
I know I am due for an Raven update, it's been a couple weeks, so here's a teaser:
0
1
6
Anyone using Claude Code or other coding agents with OCaml interested in testing ocaml-mcp-server, an MCP server that gives access to OCaml toolchains, (build status, merlin queries, etc.)?.
3
4
20
The last version of Gemini 2.5 is completely out of control 😂 Behold, Nx's BLAS backend has reached an exceptional level of power!
0
0
3
The React parallel is uncanny: the entire ML ecosystem is converging on functional APIs + compilers. JAX, PyTorch 2.0, Mojo - they're all compiler-first now. if ML is now functional APIs + compilers, what's the actual reason to stay in Python?. 🧵.
1
1
10
If your ML framework is effectively a fancy compiler, why write it in C++?. I can buy the argument for having a Python wrapper (though less so if everyone's building functional APIs anyways!), but wrap OCaml at this point!.
0
0
1
The compiler part: Hack, Flow, Coq, the OCaml compiler itself, Infer, parts of early Rust. OCaml shines most at writing compilers: pattern matching, algebraic data types, the type system. And ML compilers are just compilers!.
1
0
5
The transformation part: JAX traces your Python functions into an IR, which they transform. In Rune, we leverage OCaml 5's effects, transformations are just effect handlers. They compose naturally. Our grad fits under 1K lines of (poorly factored) code.
2
1
8
JAX is:. - Functional transformations (grad, vmap, jit).- Tracing Python functions into a compiler IR.- XLA doing the heavy lifting. It's a functional language and a compiler stack. with C++ to use in Python. At what point do you just use the right tool for the job?.
1
0
5
The React parallel is uncanny: the entire ML ecosystem is converging on functional APIs + compilers. JAX, PyTorch 2.0, Mojo - they're all compiler-first now. if ML is now functional APIs + compilers, what's the actual reason to stay in Python?. 🧵.
"A bit like how React's functional patterns feel more at home in actual functional languages, Raven is becoming what I imagine Jax/Flax intended to be - clean function composition all the way down.".
1
5
30
"A bit like how React's functional patterns feel more at home in actual functional languages, Raven is becoming what I imagine Jax/Flax intended to be - clean function composition all the way down.".
This week in Raven: convolutions are 15x faster (but still 10x slower than PyTorch 😅), new CBLAS backend in progress, overhauled Kaun API with data loading, metrics and model composition.
0
2
7