Philippe Laval

Cerebrium just raised $8.5M in a seed round to scale its high-performance, serverless AI infrastructure platform. Founded by Michael Louis and Jono Irwin, Cerebrium is building the serverless infrastructure powering real-time, multimodal AI applications — from voice agents to LLMs and large-scale batch processing — all without the complexity of managing GPUs or scaling infrastructure manually. While Cerebrium is known for its serverless GPU infrastructure, it also offers the ability to do batching, multi-region deployments, large-scale data processing, and much more. This enables teams to run compute-intensive workloads with minimal setup, scale elastically, and only pay for what they use, without the complexity of managing infrastructure while adhering to strict security and data residency requirements. They're already powering some of the most innovative companies pushing the boundaries in AI, including YC companies like Tavus and Vapi. Congrats to the team on the round! https://lnkd.in/eBRSFzi4

448

35 commentaires

French companies are showing what it means to scale AI with purpose. Being here this week, I saw how businesses here are moving from piloting AI use cases to embedding it into their operations to transform how they work.    CHU de Montpellier is leading the way. Using AI, the hospital has streamlined the way clinicians access information, giving time back to medical teams, and improving patient care. It’s a powerful example of AI driving impact where it matters most.    This is AI at scale and a sign of things to come. Microsoft’s commitment to supporting France’s innovation journey and Europe’s digital future remains steadfast.    https://lnkd.in/dNPV3ipg

373

13 commentaires

Can we predict a drug’s potential before it even reaches the market? And what if AI could restore fairness in clinical trials by bridging the best drug's potential with the right group of patients — faster? I had the pleasure of joining Ziryab DIDI on #LaCapsule to explore how AI can radically transform how we develop medicines. 🔹 90% of clinical trials fail. 🔹 Benefit-Risk analysis of drugs is 30 years late 🔹 Innovation must balance profitability with responsibility. At ArcaScience, we’ve built the most comprehensive benefit-risk database in the world — powered by 20+ AI models trained by clinicians, and already trusted by top-tier pharma. We help our partners anticipate therapeutic potential, reduce risk, and build drug profiles in seconds, not months. More than a tech shift, this is about responsibility: ✅ Making smarter decisions, ✅ Serving patients faster, ✅ And doing it transparently. Huge thanks to the podcast team for the thoughtful conversation, and to Clara Armand-Delille & Lucie Ramin for making this connection possible 🙏 🎧 Listen here: Spotify: https://lnkd.in/e4aptsfy YouTube: https://lnkd.in/e3UjfdAs Apple: https://lnkd.in/eJxHaqxr

42

9 commentaires

🚀 Papilio.Bio JOINS THE HIIT PROGRAM TO ACCELERATE HEALTH INNOVATION: THE ECOSYSTEM IS THE KEY TO GOING BIGGER! Papilio.Bio has been selected as one of 24 startups (out of 80+ applications) to join the prestigious HIIT (Health Innovation Intensive Training) program, organized by La French Tech Grand Paris and the Digital Medical Hub (AP-HP, Assistance Publique - Hôpitaux de Paris). 💡 WHY HIIT? This program is a game-changer for us, offering a one-week deep dive with top experts in healthcare, investment, and entrepreneurship. We’ll gain valuable insights through high-impact workshops, strategic discussions, and hands-on sessions covering the key challenges of e-health: ✅ Navigating regulations & legal frameworks ✅ Building strong collaborations with healthcare professionals & patients ✅ Defining our market access & funding strategy ✅ Anticipating the future of digital health & its transformation 🚀 OUR COMMITMENT TO HEALTHTECH/MEDTECH At Papilio.Bio, we're on a mission to revolutionize #HPV screening with a rapid, at-home self-test that delivers results in just 15 minutes. By making early detection more accessible, fast, and discreet, we aim to reduce HPV-related cancers and bridge healthcare gaps. 🙏 A huge thank you to the organizers, mentors, and experts guiding us through this journey. We’re excited to share our progress and connect with this amazing community of health innovators! LET’S GO! 🔥 P.S. This was the moment (immortalized by my grandma with her phone) — right after we snapped our official MedInTechs photo for the award, Serena and I literally ran off because we had to pitch at the InvestEU Portal Deeptech Summit, organized by EuroQuity - Bpifrance. What a day. 🚀 The adrenaline, the excitement, the hunger to build, to discover. And we can't wait to do even more! Jérôme LELEU Paul Midy Marc-Olivier GAUCI Coralie Prebet Joris Galland Mathieu Grajoszex Nicolas Castoldi Antoine Tesniere @Charha Louafi Pierre Dubreuil Clara Chappaz Alexandra André Maxime Huerre Claire Poulard Lucas Thiery Antoine Lefebvre Melissa ZOULIM Ryan Durand Alexis Auvitu Nicolas Desruelles Amgen Biotech Experience France (ABE France) Marie MORICE-MORAND Dihya Benhamouche Paris Santé BioLabs France Bpifrance Ministèredelasanté Rdc

46

1 commentaire

iktos CEO Yann Gaston-Mathe Is Making Tools Everyone Can Use Yann Gaston-Mathé is seeing robots. As Co-Founder and Chief Executive Officer of Iktos, he leads a Paris-based company pioneering the use of artificial intelligence and robotics for drug discovery. Under his leadership since the company’s inception in 2016, Iktos has established itself as a leading player in generative molecular design, retrosynthesis planning, and integrated AI-driven drug development. And that’s just in the first 10 years. Gaston-Mathé co-founded Iktos with Quentin Perron and Nicolas Do Huu, driven by the ambition to apply generative deep-learning models to de novo small-molecule drug design. Iktos’s platforms—notably Makya, for generative molecular design, and Spaya, for retrosynthetic planning—were built to help medicinal chemists navigate chemical space more efficiently, optimizing for multiple parameters simultaneously (potency, safety, synthetic feasibility, etc.). In his words, “the search for new drug matter is still a laborious and uncertain quest… The application of artificial intelligence… offers the possibility of a radical change… The technologies we have developed … will allow us to execute our plan of being the preferred partner for the industry.” The financiers certainly think so. In March 2023, Iktos raised €15.5 million in Series A funding, co-led by M Ventures and Debiopharm Innovation Fund, with backing from Omnes Capital. According to Gaston-Mathé, the funds would accelerate not only its AI technology but also the launch of Iktos Robotics, an “automated end-to-end drug discovery service.” More recently, in February 2025, Iktos secured a €2.5 million grant (with a potential additional €5 million) from the European Innovation Council (EIC) Accelerator. Gaston-Mathé has underscored a philosophy of democratizing complex drug design technology. In an interview, he noted that Iktos aims to make its tools “accessible to everyone … whatever their level of expertise in deep learning and computer programming.” He has positioned Iktos not merely as a vendor, but as a mission-driven partner to pharma—combining data science, synthetic chemistry, and user-friendly software to accelerate innovation. He makes robots, sure but people remain at the center of his company’s goals.

127

5 commentaires

💡 Beyond Pretraining: How Post-Training Shapes LLM Reasoning While pretraining gets all the headlines, it’s post-training that defines how useful, safe, and aligned your language model actually is. This new survey — LLM Post-Training: A Deep Dive into Reasoning — maps the evolving landscape of fine-tuning, reinforcement learning, and test-time scaling techniques that shape reasoning and alignment in today’s LLMs. 🔍 What the paper covers: 🔹 Fine-Tuning Classic SFT remains core to domain adaptation, but risks overfitting and high compute. Parameter-efficient approaches like LoRA and adapters help scale fine-tuning more affordably. 🔹 Reinforcement Learning (RL) LLMs reframe RL as token-level decision-making over long horizons. Modern methods like PPO, DPO, GRPO, and RLAIF go beyond traditional RL by optimizing for human-aligned objectives like helpfulness, safety, and factuality. 🔹 Reward Modeling Explains how models use both explicit rewards (e.g., human feedback) and process rewards (e.g., chain-of-thought quality) to improve step-by-step reasoning. Emerging trends include multi-sample optimization and preference-based training (e.g., ORPO, DPO). 🔹 Test-Time Scaling & Reasoning Frameworks From Tree-of-Thought to Self-Consistency and Value-Guided Decoding, inference-time tools now enhance reasoning without retraining. 🔹 Unified Taxonomy Breaks down recent models (GPT-4o, Claude 3.5, DeepSeek-R1, Gemini, etc.) by their post-training stacks — SFT, RLHF, GRPO, distillation — offering a panoramic view of modern LLM training pipelines. 📘 Why it matters: Pretraining builds general skills, but post-training determines real-world utility. This paper is a go-to resource for understanding how models are adapted, aligned, and optimized for reasoning tasks — from math and code to multi-turn dialogue and safety-critical scenarios. 🎯 Want to build truly useful and aligned AI? Our LLM Bootcamp dives deep into post-training techniques — the real drivers of reasoning, safety, and performance in today’s top models. 🔗 Register now: https://hubs.la/Q03pLvKf0 #LLM #AIAlignment #MachineLearning #PostTraining #RLHF #DPO #DeepLearning #AIBootcamp #LanguageModels

68

1 commentaire

[link in the first comment] great profile of Mario Ciabarra of Quantum Metric. some of my thoughts on Mario and a discussion of their excellent work with GenAI from the piece: "...when Insight Partners first invested, Ciabarra was asked who he would like on the board — he requested Managing Director Lonne Jaffe, specifically. The pair had clicked instantly: the same age, both entrepreneurial, both product-focused. 'It’s really fun going deep into product with Mario; he has such strong technical intuitions and a deep understanding of his users,' explains Jaffe. Jaffe was previously CEO of Syncsort (now Precisely), a global leader in data integrity. 'Lonne is an operator, and he has sat in my shoes,' says Ciabarra. 'Lonne knows what my day-to-day looks like. I wanted someone who knew what it felt like to be CEO of a rising company.' When making senior hires, Ciabarra runs candidates past Jaffe first. 'He is very meticulous in the way he thinks and how he attacks specific problems,' Ciabarra explains. Jaffe found Ciabarra to be a rare blend of engineering expertise and sales acumen. 'I remember an Insight Enterprise event where he closed a very large deal while everyone was sipping drinks,' he explains. 'He’s just a very unusual combination. Often you have product-centric or go-to-market-centric CEOs, but Mario is fantastic at both.' Alongside Jaffe’s input, Ciabarra has leveraged expertise from Insight Onsite — a dedicated team of 130+ in-house experts. 'When I have things that I’m working on that I’m not quite sure of — like if we need to create a category — we bring in their Onsite team to help us figure out, do we actually want to create a category? What is it going to look like? They help guide us through the pros and cons,' says Ciabarra. Reflecting on the partnership, Jaffe says: 'Some of our founders are better at taking advantage of Onsite than others. [Ciabarra] is in the very top percentile in terms of using the Insight machinery productively.' Generative AI became a core part of the business in 2024, which Ciabarra says has been a game-changer. 'Lonne has pushed us to really invest in AI and in the data we have, that I don’t think we knew how to unlock, or that it was actually valuable at all.' As an example, their Felix AI offers session summarization and instant customer context so bottlenecks in the customer funnel can be quickly resolved, and customers are spared from having to repeat their issues and queries to support agents. 'You can use that summary as an ingredient in all sorts of analytics. [Ciabarra] showed this to one of the largest companies in the world, where he had been trying to break into for years, and they were just amazed,' says Jaffe. Customers now intuitively understand the power of AI to enhance their ability to harness insights. This has reduced the need for Ciabarra and his team to spend time explaining what the products do, and they can instead focus on accelerating sales growth..."

65

6 commentaires

🚀 The next wave of generative AI is all about specialization.   Our CPO, Mehdi Ghissassi, shares his perspective in the latest European Deep Tech Report 2025 by Walden Catalyst Ventures, Lakestar, Dealroom.co, and Hello Tomorrow.   🎯 As foundational models reach broad capabilities, the focus is shifting toward fine-tuning AI for specific industries and real-world tasks. This opens up massive opportunities for entrepreneurs to build high-value, defensible solutions — and for technologists to integrate AI seamlessly into workflows.   🔗 Dive into the report for more insights shaping Europe’s deep tech future: https://lnkd.in/d5QBcmdu #GenerativeAI #DeepTech #AIInnovation #AIforIndustries #TechEntrepreneurs

153

3 commentaires

As a certified Blackbird.io implementation partner, LangOptima is helping clients move beyond legacy localization workflows and into a new era of content operations. The shift is already happening: 🔹 From point-to-point connectors → to content orchestration 🔹 From centralized governance → to distributed governance 🔹 From assets trapped in silos → to open, queryable content datalakes 🔹 From manual file hand-offs → to API-native, event-driven automation 🔹 From linear pipelines → to parallel, adaptive workflows 🔹 From static TMs and termbases → to living, context-rich memory 🔹 From engineering dependency → to no-code iPaaS autonomy 🔹 From translation as a service → to language as infrastructure 🔹 From post-hoc localization → to shift-left, concurrent content creation 🔹 From vendor-locked stacks → to agnostic, composable ecosystems 🔹 From cost control as the goal → to leverage, reuse, and compound value Blackbird.io makes this real — with 170+ integrations, workflow orchestration, AI enablement, and an open SDK that puts control back in the hands of content and localization teams. If you're ready to rethink how your content moves, we should talk. #ContentOps #Automation #Translation #Localization #LangOps #LangOptima

29

Clarifying interview by Lara O'Reilly. I questioned in a recent piece whether Google *ever* needs to introduce ads to the Gemini chatbot, given that it is already monetizing Gemini output at parity with classic Search in AI Overviews and has begun testing ads — including with the new Direct Offers feature — in AI Mode. I’d go further and propose that Google withholding ads from the Gemini chatbot creates a war of attrition enabled by consumer expectations and capital asymmetry. Google can afford to sustain losses indefinitely with the Gemini chatbot, and it’s already monetizing the model in its most prominent consumer surface area. Why insert ads? As Ben Thompson recently pointed out on TBPN, the ad-free version of Gemini is an effortless substitute for ChatGPT, imposing very low / no switching costs on users. If ChatGPT ads don’t improve the user experience out of the gate — and they won’t! — the product could bleed users. At this point, Google is best served by holding out.

68

6 commentaires

The best directories solve high-consequence problems. And Frey Chu knows exactly how to build them. His episode for The AI Report is live. We talk about: → How to validate a niche using Reddit, Ahrefs, and even Google Review keywords → What it takes to structure data for both SEO and LLM visibility → When to go manual vs. programmatic with enrichment and scraping → And why real directories aren’t dead—they’re just misunderstood Frey breaks down the difference between content farms and assets with staying power. He shares his full tech stack, his long-game mindset, and why time freedom—not hype—is the goal. This is a blueprint for low-maintenance, high-value digital products. (And it’s one of the most tactical interviews we’ve done.) 🎧 Episode is live now: https://lnkd.in/dfbFNniH 👇 Trailer below What’s the most valuable niche directory you still use today?

84

61 commentaires