Mathieu Morey

🧵 1/ How well do LLMs actually do on Olympiad-level math? We evaluated frontier models on 455 problems from the IMO Shortlist. Unlike most benchmarks, we emphasize proof validity, not just final answer correctness. Here’s what we found 👇

@daniellellecco I second the proposals on consulting and working for the public sector, or in a small to medium company operating on a specific frontier. Lots of opportunities to transpose or transfer tools and methods, and sometimes truly invent.

what career would you recommend to someone who is good at things in general?

@JoeMayo Explicit typing, clean inheritance, docstrings should help. Hopefully best practices for humans will be easy to leverage for AI coding assistants, so we can all benefit.

Current observations in AI coding assistants. They're great at writing raw algorithms and often reveal new designs or insights outside of my current patterns/habits. However, working with 3rd party libraries and APIs is a mixed bag (one might even say that it's a new type of versioning hell too). Wondering if this is an opportunity for some type of industry practices, protocols, or standards for 3rd party providers to make their interfaces more approachable to LLM/AI coding tool consumption?

We're looking for a Staff Rust Developer! We're looking for people with - Great communication skills - Experience with large Rust projects - Skills in high-performance parsing and parsing-adjacent technology - Understanding of automata/state machines Apply below if you fit!

@daniellellecco I get less from this place than I used to, a few years ago. Informative content seems more hyped and less honest, overused words in repetitive constructions arranged to evoke attention-grabbing rhythms. Your writing and posting is a comforting exception, so thanks.

I actually don’t think I can stay here; besides the writing part, I really hate it; I enjoy saying the things I say, but I hate everything everyone else says

@CollectifCeM Pont de Vivaux aussi. Les usagers aimeraient bien avoir un peu plus de visibilité (et d'ouverture, mais il faut pas trop en demander).

Ça va, tranquille ? La piscine Vallier est fermée jusqu'au 20 janvier... 1 mois de fermeture (pour l'instant). Une des deux seules piscines ouvertes le matin en semaine dans une ville de 800000 habitants. Mais bon, à Marseille, pas besoin de piscines, on a la mer !

@samuelcolvin @astral_sh @quarto_pub addresses most of these issues for me (rich markdown, pure text so git friendly etc). Builds on years of experience from the Rmd community quarto.org/docs/computati…

4 years on, and my argument that we need a (plain text) successor to Jupyter notebooks has only become more relevant. @astral_sh's support for dependencies in a comment at the start of scripts is a good starts. github.com/samuelcolvin/n…

Part 2: Why do boosted trees outperform deep learning on tabular data?? @Jeffaresalan &I suspected that answers are obfuscated by the 2 being considered very different algs Instead we show they are more similar than you’d think — making their diffs smaller but predictive!🧵1/n

A new paper about AI in Materials Discovery has gained a lot of attention - it reports a big increase in productivity of scientists using AI tools, and interesting secondary effects of using AI for science. But do the technical claims stack up? Let's see...

🌟Come work with us on Safety at ⁦@MistralAI⁩ . Be part of a small ambitious team where there is lots to build and lots to ship!🌟 jobs.lever.co/mistral/b13733…

🌟 AI enthusiasts! Join @MistralAI and shape the future of generative AI! 🌟 We're hiring AI Scientists, Research Engineers, and more 🌐 Check out our openings: jobs.lever.co/mistral 🚀 Be part of a brilliant team working on cutting-edge projects. #AIJobs #TechCareers

I'm sorry Noam, but a blog post does not come close to meeting the standards of reproducibility, methodology, acknowledgment of prior work, and fair comparison with the state of the art, that a technical paper has to satisfy. Look, when you develop new technology under pressure to have short-term product impact, you just build the thing that you think is most likely to work as quickly as possible. If it's good enough, you deploy it. You may not care whether it's particularly innovative, whether it actually beats the state of the art, or whether it's a horrible kludge or The Right Thing to do in the long run. It's OK to delude yourself into thinking it's the best thing since slice bread, as long as your boss and the product people can also be deluded. But you know that's not how research works.

We just released a one day ticket for tomorrow. It’s not too late to join the party ! pydata.org/paris2024/tick…

@Stuk_89 @lau_devil Si c'est le vocable "influenceur" qui vous gêne, remplacez ça par "bon client des médias", le fond est le même et reste juste. Impossible pour des scientifiques de faire entendre un discours fidèle à l'état des connaissances, qd le débat public est structuré par ces argumentaires

@moreymat @lau_devil Ca ne méritait pas le tweet assassin initial pour commenter son intervention.

Le discours scientifique vulgarisé doit être ancré dans la réalité, la temporalité et doit être étayé, s’appuyant sur les conclusions de ses confrères et consœurs. Les scientifiques disent ´Nous pensons que ´ les influenceurs disent ‘Je pense que’

@Stuk_89 @lau_devil L'argument serait recevable s'il pouvait s'appuyer sur une activité de publication dans des revues de science politique à comité de lecture.

@lau_devil Elle est politologue, évidemment qu'elle aborde ces questions sous un autre angle (qu'on aime ou pas) que les informaticiens. Les médias sont un moyen comme un autre de mettre en avant son travail ou de chercher des fonds. Regardez donc votre propre activité sur les réseaux...

#RechercheDataGouv a 2 ans ! 🎂 Consultez la rétrospective en images des principaux événements de Recherche Data Gouv, l'écosystème réunissant plus de 350 acteurs au service du partage et de l’ouverture des données de recherche ➡️ recherche.data.gouv.fr/fr/actualite/r… #opendata #openscience

@DavidLaMars @lamarsweb Rdc non surélevé et grand stationnement souterrain sur une parcelle à risque d'inondation très élevé (cf. PPRI Huveaune, zonage 4) ?

@daniellellecco Plaza, or any other word from a foreign language that has been completely stretched and twisted way past any acceptable derivation after borrowing. At least in US English.

which word should be removed from the english language? (only one, but I guess those obviously derived from it go too)