Grégory Salvan

Security is a shared responsibility. The Julia ecosystem is advancing its security foundations with community-driven advisories, integrations with industry-standard tools, and a renewed focus on trust and visibility. This effort helps ensure Julia remains production-ready, secure, and trusted across research and enterprise use. juliahub.com/blog/securing-… #JuliaLang #JuliaCommunity #OpenSourceSecurity #SoftwareSecurity #SupplyChainSecurity #DevSecOps

I tried Objective C on Linux and instantly found NSA backdoor.

@cgarciae88 Gpt 5.2 exhibits sophistic reasoning when confronted with evidence of misleading or inacurate informations.

product release tier list: > gemini 3: entire month of a/b test leaks, clear sota at the time, lots of hype, generated interest on tpus > nano banana pro: flooded the TL with viral generations for weeks, still sota > gpt 5.2: impressive ARC results, some "feel the AGI" moments, openai back on the frontier, user base split over several issues, not sota at everything > opus 4.5: best coding model, put breaks on the gemini hype, solid model overall, practically no PR > deepseek v3.2: solid benchmarks, only published paper, new techniques (DSA), interest mostly technical

Thinking about using small multimodal models? Want a clearer understanding of what breaks when downscaling model size, and why? ✨Introducing our new work on Downscaling Intelligence: Exploring Perception and Reasoning Bottlenecks in Small Multimodal Models 🧵👇

#LGL @GrandeLibrairie 19 nov 2025. Où je reviens sur l'idée de la "Totalité" au coeur de mon travail.

So we're hiring.

@Adilko91152 @Elisabeth_Borne Devine qui finance les restos du cœur pratiquement pour moitié et qui a augmenté les subventions de 64% en 2024.

@GregorySalvan @Elisabeth_Borne Il y’ a tes mots et les fils d’attente devant le resto du ❤️ !!!! Et le permis ? Avec qu’elle argent ? Mdrrrrrrrr Un déconnecté qui la réalité… La jeunesse va tellement bien qu’elle est sous BALLON….

Ce bilan collectif, vous pouvez en être fiers !

@Adilko91152 @Elisabeth_Borne T'as raison, complètement déconnecté, on devrait créer des rencontres jeunesse ou un Conseil National dédié à la jeunesse... ah mais attend c'est pas ce qu'ils ont fait ? le permis à 17 ans, le mentorat, des dispositifs sur mesure... complètement abandonnés oui 🤪

@Elisabeth_Borne PTDRRRRRRRRRRRRRRRRRRRRRRRREUHHHHH plus déconnecté que vous ça n’existe pas…. La jeunesse de ce pays a été totalement abonnée pendant 8 ans. Je sais pas quand quelle pays vous vivez.. mais c’est EFFRAYANT.

@DSPyOSS 's optimizers are amazingly useful and powerful but it can feel a bit daunting to get started. @CShorten30's knack of explaining things with such clarity bridges that gap perfectly. Do check it out

Dans ma nouvelle note sur le site de FIPECO j'examine le plan de redressement des comptes publics pour 2026 présenté par F. Bayrou : fipeco.fr/commentaire/Le…

Voter la confiance pour freiner la dette lenouveleconomiste.fr/voter-la-confi… via @LeNouvelEco

@rohanpaul_ai What a shame they restricted it to "sex" instead of survival tactics and strategies. Sexual selection is actually a subset of survival strategy when viewed through the lens of life history theory and social intelligence research. That would have strengthened their theory.

"A Unified Theory of Language" The paper argues language is a fast Bayesian pattern system, shaped by sexual selection to display intelligence. It uses Construction Grammar, where a construction is a stored pairing that links form to meaning across words and gestures. Processing uses unification, which means matching the best fitting construction to the situation, and Bayesian here means choosing the most likely match. Because both sides unify the same constructions against the same common ground of shared facts, the model claims a communication guarantee. Learning follows the same recipe, a few examples let a listener infer a new construction, then reuse it quickly, giving a learning guarantee. Pragmatics is handled by an interaction engine, skills for turn taking, reference, and intention reading, all encoded as constructions. Speaker and listener play an intention game, one displays their ability to infer goals and references, the other sizes it up. This show off and size up loop is framed as sexual selection, favoring fast theory of mind and richer language. So phonology, syntax, semantics, and pragmatics become one pipeline computed by quick unification over learned constructions. ---- Paper – arxiv. org/abs/2508.20109

@MichaelAArouet This is not comparable, you've to take "after redistribution" to have a small idea of real numbers. French state is taking a lot but redistributing a lot too, through charges exonerations, subsidies for employment... It remains high when you count everything but not at this point

@lateinteraction @isaacbmiller1 A mapping with words you can later add to prompts ? deterministic → 0.0 strict, rigid, precise, factual → 0.1 consistent, reliable → 0.2 systematic, methodical, structured → 0.3 balanced, measured → 0.4 standard, moderate, neutral → 0.5 flexible, adaptative → 0.6 ...

OpenAI models no longer accepting temperature is such an uncalled-for headache. Right now, @isaacbmiller1 are brainstorming some decisions about temperature in 3.1 onwards. Some of the more harmless changes could go into effect right into 3.0.3 and others may wait to 3.1. Let me know what you think! 1. Temperature will be deprecated [but NOT removed!]. Reasoning: It's too low-level of a concept anyway, not meaningfully declarative. Compared to when DSPy was created in 2022, models are now aggressively RL'ed and the effect of temperature is mostly voodoo at this point. It's *at best* a "hint" to DSPy, not a concrete requirement. From then on, two things will happen: If you explicitly set the temperature, it'll still go to the LLM! But you'll get a warning (just the first time in the run) about deprecation. Note that, if your LLM does not support varied temperature, your code will crash because your LLM will reject it. (We preferred this over automagically changing the intent of your temperature silently. A crash from your LLM provider lets you decide how to deal with this. The recommendation will be: just don't set the temperature. Let the provider/DSPy default do its work.) 2. If you want to bypass the cache, i.e. generate multiple fresh rollouts, you can use some new `seed` argument. Sadly, `seed` is already taken by OpenAI and *some* other providers. So we'll come up with a special keyword instead. This will make sure to re-generate fresh rollouts (and cache them separately, under a different cache key). This gives you a combination of speed & reproducibility (cache) but much cleaner control over rollouts than varying temperature. We will modify dspy.MIPRO, dspy.Refine, dspy.SIMBA to stop introducing temperature changes for rollouts. [A lot of people have been doing this via `seed` already, for a long long time, but it's not guaranteed officially in DSPy, since not all providers recognize seeds.] 3. If you do NOT explicitly set the temperature, from 3.1.0 (next "minor" bump), the default temperature will switch to `None` instead of `0.0`. This is the change that bothers me most, so I'm not yet committed to it, but it seems like it'll be necessary eventually. It's already a hassle to be forced to set temperature=1.0 explicitly for some models. We could try to maintain a list of known models that don't support variable temperature, but that's not a scalable solution. Or is it? At DSPy scale, it might actually be possible to cover >95% of users automagically, I suppose? Thoughts, folks?

AI efficiency is important. Today, Google is sharing a technical paper detailing our comprehensive methodology for measuring the environmental impact of Gemini inference. We estimate that the median Gemini Apps text prompt uses 0.24 watt-hours of energy (equivalent to watching an average TV for ~nine seconds), and consumes 0.26 milliliters of water (about five drops) — figures that are substantially lower than many public estimates. At the same time, our AI systems are becoming more efficient through research innovations and software and hardware efficiency improvements. From May 2024 to May 2025, the energy footprint of the median Gemini Apps text prompt dropped by 33x, and the total carbon footprint dropped by 44x, through a combination of model efficiency improvements, machine utilization improvements and additional clean energy procurement, all while delivering higher quality responses. See the blog or technical paper for more about our methodology and ongoing efforts. Blog: cloud.google.com/blog/products/… Link to detailed paper: services.google.com/fh/files/misc/…

🤪 I prioritize being clever

@nrehiew_ I'm using it for coding on zed, I like it, just sent it your last message with the paper, the answer is good

First, the evals say its a good model, i have used it and i think its a good model and people i trust have used it and said its a good model. The paper is also great. I think they focused more on breadth here rather than more indepth details. I would like to have more info on certain parts and details but I guess this paper can serve as sort of a general recipe or glossary. I do wonder if theres anything special here because it didnt feel like they did anything unique or special beyond good execution. But there again, if there was would they put it in the paper? 1 thing i liked is the attention head thing which i think is pretty cool to see loss results with Kimi's experiments match up but always check downstream results. Another thing i wonder if doing single turn RLVR first before doing agentic RL helps performance and if this can be a further way to push K2 for eg. The main thing I want to know is on the hybrid reasoners. There wasnt much talk about evals related to this hybrid behavior and the decisions behind it. I think its important since Qwen got rid of the hybrid in exchange for 2 separate models and OpenAI kind of did the same thing for GPT5. I wonder if the team has any insights here.

Let's talk about the GLM 4.5 models. The latest frontier open weights model out of China (and possibly the best at the moment?) with quite a bit of details in the paper.

@AnthropicAI Claude code is constantly failing with SoC, it add getter/setter everywhere even with Demeter law in its contraints and explicit asking for visitor pattern, it cheats and use getters. It reveals issues in scopes awareness too. A huge axis for improvements imho.

@stevekrouse @karpathy ... and reading code (specially bad one) takes more resources than writing it.

Vibe code is legacy code @karpathy coined vibe coding as a kind of AI-assisted coding where you "forget that the code even exists" We already have a phrase for code that nobody understands: legacy code Legacy code is universally despised, and for good reason. But why? You have the code, right? Can't you figure it out from there? Wrong. Code that nobody understands is tech debt. It takes a lot of time to understand code enough to debug it, let alone introduce new features without also introducing bugs Programming is fundamentally theory building, not producing lines of code. This is why we make fun of business people who try to measure developer productivity in lines of code When you vibe code, you are incurring tech debt as fast as the LLM can spit it out. Which is why vibe coding is perfect for prototypes and throwaway projects: It's only legacy code if you have to maintain it! I vibe code happily all the time. Most often for small apps that I don't need to maintain. I'm a big fan, have at it! Vibe coding is on a spectrum of how much you understand the code. The more you understand, the less you are vibing Simply by being an engineer and asking for a web app with a persistent database, you are already vibing less than than a non-programmer who asks for an "app" without understanding the distinction between a web app and a native app, or how persistent data storage works The worst possible situation is to have a non-programmer vibe code a large project that they intend to maintain. This would be the equivalent of giving a credit card to a child without first explaining the concept of debt You'll end up spending a lot of money and getting a large, buggy, legacy code base. If you don't understand the code, your only recourse is to ask AI to fix it for you, which is like paying off credit card debt with another credit card At Val Town, we built Townie, an AI assistant that agnatically reads & writes code, runs it, views the logs, and keeps iterating until it's done. It's is an awesome tool for vibe coding. I heartily recommend it to folks who understand these tradeoffs. I use it to vibe code sometimes. Other times I keep in on a tight leash as it makes surgical edits to a project I care about If you know any non-programmers spending thousands of dollars vibe coding their billion dollar app idea today, warn them that vibe coding is not going to get them where they want to go. They're going to have to learn to use their human eyes to read the code 😱, and that sometimes it's easier to start over with building a well-written code base from scratch than to fix a legacy one that nobody understands

When you realize that open-source is at the frontier of AI despite: - less GPUs - less money - less public and policy support - no $100M salaries to attract talent - with closed-source taking advantage and copying all the innovations of open-source without contributing back theirs 🤯🤯🤯 And we’re just getting started!