🔥Telegram İfşa

Then we'll implement a high-level DSL to simplify the construction of complex samplers with Blackjax, and use AeMCMC to automatically compile samplers thet are adapted to your model. github.com/rlouf/aex

Another PR to Aesara Devs has been merged successfully😃. I'm thrilled to be a part of this amazing #OpenSource project and to contribute to the tech community.🎉🥳

The Aesara Project is excited to announce that we will be rewriting all of our libraries in Rust. Let’s cargooooooool!

The Numba backend and Rust both use LLVM under the hood!

AePPL

Submit 30 PRs 🎂

🥳

Saul Shanabrook Numba Aesara Devs We're in the process of moving some of our symbolic work from Aesara Devs into the the LLM space, and we might end up using egglog and the new RVSDG support in Numba to do that.

Saul Shanabrook Numba Aesara Devs The plan with Aesara Devs was always to shed the tensor library elements and focus on building useful domain-specific relations using whatever frameworks are most effective. egglog and Numba are shaping up to be those frameworks, and that's exciting!

I gave a lightning talk at PyData NYC giving a short intro to e-graphs as well as this example of compiling scikit-learn to numba youtube.com/watch?v=TEkLFm…

I just wrote a reply involving questions about Aesara Devs, and it reminded me that we need to provide some updates regarding the projects: github.com/aesara-devs/ae…

Simply put, the projects have gone through a lot in the past year and we've moved on to some exciting new ventures. We're bringing similar symbolic thinking and approaches to a new space: LLMs.

Regarding our original mission with Aesara Devs, we believe that projects like Numba, its push for RVSDG support, and projects like the ones by Saul Shanabrook are good ways forward in this space.

Aesara was designed to fulfill a very specific backwards compatibility requirement that ultimately complicated its true goals: domain-specific symbolic rewriting. We're very proud of what we were able to accomplish under those constraints, but the way forward can't have them.

I'm glad that my adhoc transpilation approach was able to breathe new life into an old project like Theano, but tighter integration with better compilation targets are necessary.

Projects like Numba can provide that, along with all the necessary support and expertise for maintaining them.

At some point I'll start putting out write-ups about some of the most interesting domain-specific rewrites and optimizations we added, and some details about the automated sampler construction process and its potential.

In the end, I'm glad to have gotten the chance to implement and demonstrate some of the features I always wanted in a PPL, and I still hope to see one that realizes them all--even if it ends up being in Julia!