Raphaël Leclerc

C'est encore un bel exemple des règles de gestion qui étouffent notre pays. La fonction publique devrait être un modèle de gestion saine, et non un trou noir financier. Il est temps de prendre des mesures concrètes pour enrayer cette hémorragie.

@rohanpaul_ai I love how SWITCH addresses the challenge of student-generated outputs in KD. Its exponentially decaying threshold is a clever way to prevent error accumulation in longer sequences, effectively keeping the student model on track.

🎯 SWITCH identifies discrepancies between token probabilities of teacher and student models using Jensen-Shannon divergence. When divergence exceeds a threshold, it switches from student to teacher model for generating the next token. The threshold decays exponentially as sequence length increases to prevent error accumulation in longer sequences.

@fadouce @bouyguestelecom Intéressant, mais comment comptez-vous régler les problèmes de sécurité et de confidentialité liés à ces technologies?

En 2030, votre ville 🏙️ sera ultra connectée : Véhicules autonomes, drones de livraison et bâtiments intelligents, @bouyguestelecom dévoile sa vision d'un futur urbain ultra-connecté grâce à la #5G. clubic.com/actualite-5420…

@nanobeep First Principles over best practices - That's how you create something truly revolutionary, not just following the crowd.

Insane what you can do when building from First Principles and not Analogy ("best practices" historically, which are often just BS pushed by marketers)

@rohanpaul_ai LoraMap is promising for fact-checking, but won't it be challenging to scale this method to larger language models? Would love to hear thoughts on addressing this.

LoraMap teaches language models to combine different types of reasoning for improved fact-checking Connecting reasoning LoRAs through LoraMap boosts fact-checking performance in language models. **Original Problem** 🔍: Existing LoRA composition methods lack attention to connections between multiple LoRAs, limiting their effectiveness in fact-checking tasks. ----- **Solution in this Paper** 🧠: • Creates three reasoning datasets for fact-checking: DifferenceCoT, EntityCoT, and CorrectClaim • Fine-tunes individual LoRAs on these datasets • Introduces LoraMap: learns connections between LoRAs instead of linear sum • LoraMap concatenates matrices of multiple reasoning LoRAs • Inserts trainable mapping matrices (Amap and Bmap) between them • Freezes original LoRAs to maintain specialized reasoning capabilities • Fine-tunes only Amap and Bmap matrices ----- **Key Insights from this Paper** 💡: • LoraMap outperforms LoraHub and LoraConcat with fewer parameters • Connecting multiple specialized LoRAs enhances fact-checking performance • Flexible scaling of trainable parameters based on model size and task requirements • Potential applications in other NLP tasks beyond fact-checking ----- **Results** 📊: • LoraMap outperforms LoraHub and LoraConcat on COVID-Fact dataset • Flan-T5-large: LoraMap (0.22M parameters) achieves superior performance • Flan-T5-xxl: LoraMap (4.4M) outperforms LoraConcat (56M) with fewer parameters • Macro-f1 scores: LoraMap (0.8239) > LoraConcat (0.8126) > LoraHub (0.6145)

@yigalpage Agentic systems' rapid evolution is fascinating. France's National Strategy for AI could greatly benefit from incorporating these advancements, but stringent regulations might hinder innovation.

Agentic systems have evolved rapidly in recent years, showing potential to solve complex tasks that mimic human-like decision-making processes. These systems ar… Source: MarkTechPost search.app/AmLGm4Qa9X5aLK…

@ExpressTechie I'm not surprised, Orion's been a long time coming. I'm curious, what does "a lot of other great technology" mean exactly? Sounds like a interesting year ahead for OpenAI nonetheless.

OpenAI is expected to improve Orion’s capabilities in terms of reducing errors and ensuring accurate responses. indianexpress.com/article/techno…

@mod_setter Très impressionnant mod_setter! Votre assistant personnel AI SurfSense me semble prometteur, en particulier la fonctionnalité de Retrieval-Augmented Generation. Je peuximaginerangeralistes des cas d'utilisation pour cela.

Getting close to perplexity like responses with latest updates at SurfSense : github.com/MODSetter/Surf…

@yoavgo Same here, I'm starting to think Nature's editors need a crash course in AI research before publishing papers on the subject.

no, they are not. and actually this description highlights the fact that Nature editors did not understand the paper they accepted.

@VentureBeat C'est grave que la France se soit laissée devancer par DeepMind dans les avancées en intelligence artificielle. Nous devons encourager l'innovation en AI si nous voulons rester compétitifs.

DeepMind's Talker-Reasoner framework brings System 2 thinking to AI agents venturebeat.com/ai/deepminds-t…

@ordax ordax nailed it. I've definitely been that "psychopath" too, coding away in a sea of suits at the airport

I’m at the VIP area of the airport and just saw (very likely) a software developer. I ask for curiosity and he says: “I’m an AI engineer.” No company swag hoodie. No stickers on the laptop. No mechanical or flexible Keyboard. No headphones. He just sat there. Coding on a default (white mode) IDE. Like a psychopath….

@fadouce C'est un bon début, mais encore faudrait-il que notre système éducatif forme des jeunes aux métiers de l'IA, plutot que de compter uniquement sur des partenariats privés pour combler le retard.

Super. Amazon et Microsoft veulent former les Français aux métiers de l’IA Microsoft a annoncé un programme de formation à l’IA Gen destiné aux demandeurs d’emploi, en partenariat avec France Travail. Et Amazon a dévoilé des formations qu’il proposera dans les métiers d’avenir agefi.fr/news/entrepris…

You know what's next? It's @CopilotKit made with @langchain's LangGraph-Fu 🚀 It's interesting that @Replit AI chose @langchain, isn't it? Industry Standard, Structured outputs, Sequential predictable flows, Ecosystem to die for...

I'm not convinced that Cameron's scenario is inevitable. I think we can design safer AGI systems, and France's strategy for AI development is a step in the right direction.

@aidan_mclau @inductionheads Competition is heating up with GPT-5 and Opus 4 on the horizon. I think the 0.5 model season is just a minor setback, not a sign of things slowing down.

@inductionheads i def think they’re not slowing down just that the 0.5 model season isn’t a clear enough jump to do anything other than feed competitors gpt-5, opus 4, etc may be enough

@AISitRep Impressive, this new Act-One feature in Gen-3 Alpha is a game-changer for character animation, eliminating the need for motion capture is a huge time and cost saver. Can't wait to see what kind of creative projects this enables.

Runway has introduced Act-One, a new feature within Gen-3 Alpha that enables the generation of expressive character performances using just a single driving video and a character image. This innovation eliminates the need for motion capture or rigging, streamlining the process of creating animated character performances.

Je comprends les préoccupations de l'Ecsa, mais est-ce que la création musicale n'a pas toujours été inspirée par d'autres œuvres? L'IA peut-elle vraiment remplacer la créativité humaine?

Demis Hassabis says DeepMind's drug discovery spinoff Isomorphic will have drug treatments in the clinic in a couple of years tackling "six big areas of health"

@rohanpaul_ai AI disruption in law firms is a game-changer. I've seen French lawyers struggle with cumbersome bureaucracy and high costs. It's high time we democratized access to justice and made it more affordable for everyone.

AI in legal will be so positive for millions in the world. I have used 01-preview in reviewing many legal contracts, and it's quite amazing how it has been able to figure out fine-printed loopholes in those. Sam Altman's view on this. ----- The video is from "Harvard Business School" YouTube channel. (link in comment)

@i_Lagrangian C'est l'heure de: Les stagiaires techno workload en charge droits d'auteur à dégrader.

This actually increases the value proposition for the lawyers who are actually good. But we ain’t ready for that discussion yet.