Eric O LEBIGOT (EOL)

Large Language Models are not (yet?) good at fixing real-life bugs: they typically involve understanding how many parts of a code base relate to each other. leaddev.com/tech/researche…

I am saying this as the creator of a CD of compositions for the piano: #AI-generated music, much like AI-generated images, recently made big progress, to the point that I do find many examples generated by ⁦@SoundrawUS⁩ quite convincing. Good job! soundraw.io

The Escape key is often located far from the main keyboard keys. Having Caps Lock behave like Escape lets me stay in the flow, when dismissing dialogs and editing text (with vi(m), which was developed on a terminal where Escape was very accessible). vim.fandom.com/wiki/Map_caps_…

Do you like clowns, jugglers, dancers and, of course, music, all wrapped in an energetic performance? I can't recommend Stomp enough. Now in #Paris. stomponline.com

Given the fact that players don't see the nature of the opponent's pieces in Stratego, this is an interesting development (even if poker has already been a success for AI).

Interested in writing rich notes, seeing connections between them, anywhere, and finding information more easily? I can't recommend Obsidian enough. A strong program with a strong community. obsidian.md

Transformers have shown remarkable capabilities - but can they improve themselves autonomously from trial and error? New research shows how a single transformer can explore and solve tasks using this method without ever updating its parameters ➡️ dpmd.ai/3sv7Oh1

JupyterCon is back! Mark your calendar for next year's conference about @ProjectJupyter on May 10-12, 2023 at Cité des Sciences in Paris. blog.jupyter.org/jupytercon-is-…

After building on years of work from MILA, DeepMind, ourselves, and others, our AIs are now expert-human-level in no-press Diplomacy and Hanabi! Unlike Go and Dota, Diplomacy/Hanabi involve *cooperation*, which breaks naive RL. arxiv.org/abs/2210.05492 arxiv.org/abs/2210.05125 🧵👇

@aureliengeron Can you share when the PDF version will be available?

The 3rd edition of my book is being printed, it will be released within the next few weeks! 📙🦎🎉😁 You can preorder it at homl.info/amazon3 You can play with the notebooks at github.com/ageron/handson… It covers many new topics, including diffusion models:

Hands-On Machine Learning is arguably a very pedagogical and in-depth book that covers a large number of subjects. After doing a bibliography of modern machine learning books, this is the book I first recommend. 👍

"Writing, to me, is simply thinking through my fingers." (Isaac Asimov) Celebrating 10 years of journaling (mostly through my thumbs) and 13 k journal entries (mostly through the excellent @dayoneapp).

Today in partnership with @emblebi, we’re releasing predicted structures for nearly all catalogued proteins known to science, which will expand the #AlphaFold database by over 200x - from nearly 1 million to 200+ million structures: dpmd.ai/AF-22-TW 1/

@l_amazone_ Secouer un bébé est un geste terrible. J'ai été témoin de la chute d'un bébé, qui a eu les symptômes du Syndrome du Bébé Secoué. Il faut donc se méfier des mots : le Syndrome du Bébé Secoué est un ensemble de symptômes que l'on retrouve chez les bébés secoués, mais pas seulement.

Check out my latest video on @DeepMind’s recent “AI for Mathematics” breakthrough that enabled mathematicians to establish new and significant theorems. I cover the knot theory background, machine learning basics, and the Nature paper youtube.com/watch?v=UPCI1-…

If you have a baby or a young child, I recommend @BabyCenter. Their regular emails point to sound and practical advice tailored to the age of your children. I have never found a better site. babycenter.com

Dune, the novel, weaves a rich world which is much more interesting than its cinema adaptations. Some famous science-fiction works feel dated, but Dune is different. It is worth reading it again with 35 additonal years of life experience. en.wikipedia.org/wiki/Dune_(nov…

It can be challenging for robots to imitate precise and decisive behaviors. Introducing Implicit Behavioral Cloning, a simple method that scales to difficult real-world tasks and achieves state-of-the-art performance on human-expert offline RL benchmarks→ goo.gle/3FurkP6

Little known feature: cutting and pasting files also exists on #macOS. Cmd-C (copy), and then Option-Cmd-V (almost like paste). This was the only thing I was missing from #Windows, and macOS does it better (there is no risk to lose your files if you delete them midway, etc.).

Have you ever used Scikit-Learn's TransformedTargetRegressor? It's pretty convenient when your targets need to be preprocessed. For example, if they have widely different scales, it's probably best to predict log(targets), then compute exp(predictions). Like this:👇