Yann LeCun

Dive deep into the revolutionary work of Yann LeCun and the development of convolutional neural networks that transformed computer vision forever. This episode explores how LeCun's innovative approach to machine learning, inspired by biological vision systems, solved the seemingly impossible challenge of teaching computers to see and recognize patterns in images. Discover the fascinating journey from LeCun's early frustrations with handwritten digit recognition at Bell Labs in 1989 to the breakthrough CNN architectures that now power everything from photo organization to autonomous vehicles. Learn how his persistence through the 'AI winter' years ultimately led to the 2012 ImageNet victory that sparked the modern deep learning revolution. We examine the technical innovations behind convolutional neural networks, including how they preserve spatial relationships and use mathematical convolutions to detect features regardless of position. The episode covers real-world applications from postal code recognition to medical image analysis, while exploring the broader implications of giving machines the gift of sight. Perfect for technology enthusiasts, AI researchers, and anyone curious about the pioneers who shaped modern artificial intelligence. Keywords: Yann LeCun, convolutional neural networks, computer vision, deep learning, artificial intelligence, machine learning, CNN, ImageNet, Bell Labs, neural networks, pattern recognition, AI history, technology innovation.

JEPA 不是下一個 ChatGPT，而是 Yann LeCun 押注的世界模型路線：它預測 latent representation，避開像素雜訊，試著學會更穩定的世界結構與行動規劃能力的路線。 ⭐ 文章深度讀：想把 JEPA 的 latent representation、world model 和防塌陷問題一次串起來，可以看完整文章。 → https://heymaibao.com/jepa-yann-lecun-world-model/ ⚡ 章節重點 JEPA 不是 ChatGPT 殺手 00:00 三種模型到底在預測什麼 01:44 用 view 理解世界狀態 03:48 JEPA 的三個核心元件 04:28 最大的難題是防止 representation collapse 06:00 為什麼 LeCun 押注語言之外 07:28 📝 懶人包 ∙ JEPA 的核心，是用一個觀測角度的 embedding，去預測另一個觀測角度的 embedding。 ∙ 它想避開的問題，是影像、影片、醫學影像裡有太多低階雜訊，直接預測像素會浪費模型容量。 ∙ 它真正難的地方，是 latent space 必須有資訊、夠穩定，還不能 collapse 成一個沒意義的向量。 ∙ 我的觀點：LLM 不會因此失效，但 LeCun 對「語言之外的世界模型」提出的批評值得認真聽。 📚 參考資料 What Is Yann LeCun Cooking? JEPA Explained Simply → https://www.youtube.com/watch?v=oM4neOyZOi0

AI is everywhere right now. In your inbox, your EHR, your hospital's strategic plan, and probably your last three CME credits. But most of what physicians hear about AI is either hype or fear. We think you deserve something better. In this special Best of AI episode, Dr. Graham Walker revisits three conversations that generated the most messages, DMs, and forwarded links from How I Doctor listeners, plus a Physician Spotlight that has nothing to do with AI and everything to do with why physicians need support right now. These guests don't agree on everything. But together they map a path that every physician needs to understand: where AI actually works today, where it falls short, and what's coming next. Featured Episodes & Links Yann LeCun is the ACM Turing Award Laureate, former Chief AI Scientist at Meta, and co-founder of AMI, Advanced Machine Intelligence. Alex LeBrun is a serial entrepreneur, co-founder of NABLA, and CEO of AMI. Listen to Move Over LLMS! AI Legends Yann LeCun and Alex LeBrun Debut AMI Labs' Bold Ambitions for World Models in Healthcare Dr. Bob Wachter is Chair of the Department of Medicine at UCSF, author of The Digital Doctor, and author of the new book A Giant Leap: How AI Is Transforming Healthcare and What It Means for Our Future. Listen to What Doctors Get Wrong About AI with Robert Wachter, MD Dr. David Rhew is an infectious disease physician and Global Chief Medical Officer at Microsoft, working at the intersection of the world's largest health systems and its largest technology companies. Listen to Where AI in Medicine Is Actually Headed, with Microsoft’s CMO David Rhew What You'll Learn Yann LeCun & Alex LeBrun Why large language models have a fundamental ceiling and what world models are designed to do instead Why 80% accuracy is considered excellent in research and completely unusable in clinical practice What it actually took to build AI tools physicians trust and why every assumption they walked in with was wrong Dr. Bob Wachter Why the healthcare system is uniquely positioned to benefit from AI, and uniquely capable of wasting those gains on a faster hamster wheel The deskilling death spiral: what happens when the AI becomes more reliable than the human checking its work What part of medicine Bob believes we should never hand to AI, and why Dr. David Rhew Why AI in medicine isn't one thing, it's a portfolio of tools, and conflating them leads physicians to trust the wrong ones How retinal screening AI is finding advanced diabetic disease in people who were told everything was fine months earlier Why the most important physician skill in 2030 will have nothing to do with clinical knowledge Dr. Stefanie Simmons The spheres of control, influence, and acceptance and why physicians chronically misestimate which is which Why 43 state medical boards have now removed stigmatizing mental health questions, and how to check if yours is one of them 🩺 Offcall is more than a platform — it’s a community. Join today ! 🎧 Subscribe to receive new How I Doctor episodes directly in your feed here: <a href="https://episodes.fm/1767429315" rel="noopener noreferrer"

開場白 歡迎收看、收聽今天的「Mark 的 Tech Insights」！今天的新聞真的超級勁爆，OpenAI 居然把大家期待已久的 Sora 給砍了？還有 AI 教父 Yann LeCun 拿了十億美金要證明 LLM 不是唯一解，以及跟台灣供應鏈息息相關的 AI 晶片走私大案。準備好你的咖啡，我們馬上開始！ 本日重點新聞 1. [OpenAI 關閉 Sora 影片生成器，算力全面轉向機器人] 來源 : LLM Stats / AI News（ 原文連結 ） 摘要 : OpenAI 證實將正式關停 Sora 影片生成專案，並把釋放出來的龐大算力資源全部轉移到機器人（Robotics）研發上。這代表他們從消費端的生成式影片，戰略性大轉彎，全面進軍實體 AI 領域。 台灣觀點 : 台灣是全球硬體製造與精密機械重鎮，OpenAI 轉向 Embodied AI（具身智能），對台灣的伺服器、馬達控制與邊緣運算產業是一大超級利多。 討論要點 : 為什麼放棄 Sora？是版權爭議太難搞還是變現模式走不通？ 軟體巨頭紛紛投入實體世界，Embodied AI 會是下一個爆發的兆元產業嗎？ 講稿建議 : 「第一個大新聞真的讓我跌破眼鏡！大家還記得當初 Sora 釋出展示影片時有多震撼嗎？結果 OpenAI 居然說不玩了，直接關掉。說真的，從商業角度來看，我覺得這步棋下得很聰明。生成式影片現在面臨嚴重的版權訴訟，燒錢燒太兇，變現又一直不明朗。反觀他們現在把算力 ALL IN 到機器人上面，這對我們台灣來說根本是天上掉下來的禮物！台灣的硬體底子這麼強，OpenAI 要做實體機器人，最後絕對得找台灣供應鏈幫忙落地。這也暗示了，純軟體或內容生成的 AI 應用可能遇到瓶頸，大廠開始往『軟硬整合』的實體世界發動總攻擊了。」 2. [Yann LeCun 新創 AMI Labs 拿下 10 億美元種子輪融資] 來源 : Bloomberg（ 原文連結 ） 摘要 : AI 教父 Yann LeCun 創辦的 AMI Labs 剛成立就拿到 10.3 億美金，創下歐洲史上最大種子輪紀錄。他押寶基於 JEPA 架構的「世界模型」，堅信這才是超越現有大語言模型、解決真實世界任務的終極答案。 台灣觀點 : 台灣開發者與企業大多深度依賴 Transformer 架構的 LLM，若 JEPA 崛起，業界可能需要提早準備適應新的底層架構與推論硬體需求。 討論要點 : 10 億美金的「種子輪」是資本狂熱，還是代表創投圈也認為 LLM 遇到天花板了？ 「世界模型」與現有 LLM 最大的差異會如何影響日常應用？ 講稿建議 : 「接下來這個也是核彈級消息。Yann LeCun 一直以來都在社群上『逆風』唱衰 LLM，說大語言模型沒有真正的理解力。結果大佬不只用嘴巴說，人家直接出來開公司證明給你看了！AMI Labs 一口氣拿了 10 億美金的『種子輪』，你沒聽錯，是剛起步的種子輪！我覺得這反映出創投界其實對 LLM 的極限感到焦慮，大家都在急著尋找 Transformer 之外的下一個 Paradigm Shift。對我們開發者來說，現在可能要把目光稍微放遠一點，關注一下這種『世界模型』會怎麼改變 AI 的邏輯。如果他真的成功了，現在我們熟悉的 Prompt 玩法和底層架構，可能又要面臨一次大洗牌。」 3. [Google 推出 TurboQuant：靜悄悄的 LLM 壓縮技術大突破] 來源 : Stark Insider（ 原文連結 ） 摘要 : Google 在 ICLR 2026 正式發表前，被爆出開發了一種叫 TurboQuant 的壓縮演算法。它能把 LLM 的記憶體佔用縮小 6 倍，最神的是「零精度損失」而且不需要重新訓練模型。 台灣觀點 : 這對台灣極具優勢的 Edge AI（邊緣運算）和 AI PC 產業是一大福音，意味著未來能在規格較低、成本更便宜的終端設備上跑旗艦級模型。 討論要點 : 模型壓縮技術會如何加速 AI 在手機與物聯網設備的普及？ 軟體「減肥」技術的突破，會不會短期內降低對高階 AI 晶片的需求？ 講稿建議 : 「我們常說 AI 發展最大的硬傷就是記憶體頻寬，但 Google 這次偷偷放出的 TurboQuant 技術真的很狂。它能把 LLM 的記憶體需求直接砍到剩六分之一，重點是『準確度完全不掉』還『不用重新訓練』！這對台灣做 AI PC 或是 IoT 設備的廠商來說，簡直是拿到超級外掛。以前我們覺得要在筆電或終端設備上跑超大模型根本是作夢，現在透過這種演算法，Edge AI 的春天真的來了。這也代表，除了硬體大廠在拚命堆算力，軟體層面的『減肥技術』也同樣有著顛覆產業的潛力。未來我們可能不需要買到最頂規的設備，也能享受到超強的 AI 服務。」 4. [價值 25 億美元 AI 晶片走私中國，美超微相關人士遭起訴] 來源 : Al Jazeera（ 原文連結 ） 摘要 : 美國破獲一起高達 25 億美金的先進 AI 晶片走私案，被起訴的三人皆與伺服器大廠 Super Micro（美超微）有關，甚至包含其共同創辦人。這顯示美國正瘋狂加緊對半導體出口管制的執法力度。 台灣觀點 : 台灣是全球 AI 伺服器的核心樞紐，此事件絕對會讓美國對台灣供應鏈與出貨管道進行更嚴苛的合規審查，增加代工廠的營運成本。 討論要點 : 地緣政治下的 AI 晶片禁令，真的防得住地下交易嗎？ 台灣伺服器代工廠未來在出貨控管上，會面臨多大的挑戰？ 講稿建議 : 「來聊點嚴肅的，這則新聞在台灣科技圈應該已經炸鍋了。25 億美金的 AI 晶片被走私，而且還牽扯到 Super Micro 的相關高層！大家都知道美超微跟台灣供應鏈的關係有多緊密。這件事一出來，我敢說美國商務部絕對會把螺絲鎖得更緊。台灣的伺服器代工廠未來的出口合規審查，一定會變得超級麻煩。這其實也凸顯了一個無奈的現實：AI 算力現在就是戰略物資，甚至是軍火。在這種大國博弈下，台灣廠商真的不能再抱著以前那種『只要有人買我就賣』的純商業心態了，未來的法遵與審查成本，絕對會大幅吃掉硬體製造的利潤，這是不可逆的趨勢。」 <h3 id="5-拒絕軍用被列黑名

開場白 歡迎來到「Mark 的 Tech Insights」！今天是 2026 年 3 月 25 日。今天的 AI 圈超級熱鬧，我們不僅看到了 GPT-5.4 帶著恐怖的自主工作能力登場，還有 Yann LeCun 拿到破紀錄融資要來顛覆現在的 LLM 霸權。另外，牽涉到我們台灣硬體圈的晶片走私大案也曝光了，趕快來聊聊今天有哪些必知的科技大事吧！ 本日重點新聞 1. OpenAI 推出 GPT-5.4：百萬 Token 上下文 + 自主工作流 來源 : Crescendo AI / Computerworld 摘要 : OpenAI 發布 GPT-5.4，上下文視窗擴展至 100 萬 Token，並在 OSWorld-V 自主電腦操作測試中取得 75% 高分，展現多步驟任務自動執行能力，部分經濟價值任務表現已追平人類專家。 2. Yann LeCun 創辦 AMI Labs，種子輪融資 10.3 億美元挑戰 LLM 霸權 來源 : Switas Consultancy / TechFunding News 摘要 : AI 教父 Yann LeCun 的 AMI Labs 完成歐洲史上最大種子輪（10.3 億美元），採用 JEPA 架構打造「世界模型」，不走主流 LLM 路線，Nvidia 為主要投資者之一。 3. 美國起訴 25 億美元 AI 晶片走私案，牽涉美超微與台灣供應鏈 來源 : Al Jazeera 摘要 : 三名與 Super Micro Computer 有關人員被起訴，涉嫌透過台灣及東南亞作為跳板，走私價值 25 億美元的 AI 設備至中國，規避美國出口管制。 4. Anthropic 拒絕五角大廈監控應用，30+ 名 OpenAI 與 Google 員工跨公司聲援 來源 : TechCrunch 摘要 : Anthropic 拒絕將 AI 技術用於軍方大規模監控與致命武器，遭國防部列為「供應鏈風險」。超過 30 位 OpenAI 及 Google 員工連署法庭之友意見書力挺，引發科技倫理與國安的激烈辯論。 5. 川普政府發布國家 AI 政策框架，主打聯邦統一監管、放寬開發者責任 來源 : CNBC / WilmerHale 分析 摘要 : 新框架擬以聯邦法規取代各州分散的 AI 立法，減輕開發者法律責任，且不另設 AI 專責監管機構，維持現有行業監管模式，走「輕度監管、全速發展」路線。 6. OpenAI 年營收突破 250 億美元，最快 2026 年底啟動 IPO 來源 : LLM Stats / Fortune 摘要 : OpenAI 年化營收達 250 億美元，正為 2026 年底上市做準備；競爭對手 Anthropic 營收亦逼近 190 億美元，顯示企業級 AI 市場已全面進入商業化爆發期。 結尾段落 好啦，今天的 AI 新聞真的是乾貨滿滿，從底層架構的革命、驚人的財報，到地緣政治的角力，AI 已經徹底改變了我們這個世界的運作方式。如果你喜歡今天的討論，別忘了訂閱「Mark 的 Tech Insights」，也歡迎留言告訴我們你對哪一則新聞最有感。我們下次見，掰掰！ 關鍵字標籤 #AI新聞 #GPT5 #YannLeCun #美超微 #科技Podcast #OpenAIIPO #AgenticAI #科技倫理

Depuis quelques années, l’intelligence artificielle est dominée par les LLM, les “Large Language Models”, comme ChatGPT ou Gemini. Ces modèles sont entraînés sur des quantités gigantesques de textes afin d’apprendre à prédire le mot suivant dans une phrase. Autrement dit, ils sont extrêmement performants pour manipuler le langage. Mais pour certains chercheurs, dont Yann LeCun, cette approche possède une limite fondamentale : ces systèmes apprennent surtout un modèle du langage, pas un modèle du monde réel. Un LLM peut donc produire des phrases plausibles, répondre à des questions ou écrire un essai. Mais il ne comprend pas réellement la réalité physique qui se cache derrière ces mots. Par exemple, il peut expliquer comment préparer un café, mais il ne sait pas vraiment comment manipuler les objets dans une cuisine ni prévoir ce qui se passerait si un robot exécutait ces actions. C’est précisément là qu’intervient l’idée des world models. Un world model est un système d’intelligence artificielle qui apprend à construire une représentation interne du monde : les objets, l’espace, le temps et les relations physiques entre les choses. Ces modèles sont entraînés non seulement sur du texte, mais aussi sur des images, des vidéos et des interactions avec l’environnement. Leur objectif est de comprendre comment le monde fonctionne, par exemple la gravité, les collisions ou le déplacement d’objets. L’une des capacités clés d’un world model est la simulation mentale. Le système peut imaginer différents futurs possibles : “si je fais cette action, que va-t-il se passer ensuite ?”. Cette capacité de prédiction permet alors la planification et la prise de décision, ce qui est essentiel pour des robots, des voitures autonomes ou des agents intelligents capables d’agir dans le monde réel. Yann LeCun estime que l’intelligence humaine fonctionne justement de cette manière. Notre cerveau possède une sorte de modèle interne du monde qui nous permet d’anticiper les conséquences de nos actions. Pour lui, une véritable intelligence artificielle devra donc posséder plusieurs capacités absentes des LLM actuels : une mémoire persistante, du raisonnement, de la planification et une compréhension du monde physique. C’est pour explorer cette voie qu’il a récemment lancé une nouvelle startup dédiée à ces technologies. L’objectif est de créer des systèmes capables d’interagir avec la réalité — par exemple dans la robotique, l’industrie ou la médecine — plutôt que de simplement générer du texte. En résumé, les LLM sont des modèles du langage, tandis que les world models cherchent à être des modèles du monde. Et pour Yann LeCun, c’est peut-être cette différence qui déterminera la prochaine grande révolution de l’intelligence artificielle. Hébergé par Acast. Visitez acast.com/privacy pour plus d'informations.

This is a recap of the top 10 posts on Hacker News on Mar 11, 2026. Feel free to leave feedback on Github: https://github.com/denolfe/hacker-news-highlights (00:00) - Intro (00:18) - Tony Hoare has died (01:20) - Meta acquires Moltbook (02:49) - Yann LeCun raises $1B to build AI that understands the physical world (04:21) - Create value for others and don’t worry about the returns (05:42) - Agents that run while I sleep (07:17) - Debian decides not to decide on AI-generated contributions (08:36) - Cloudflare crawl endpoint (09:48) - U+237C ⍼ Is Azimuth (11:02) - Universal vaccine against respiratory infections and allergens (12:10) - RISC-V Is Sloooow (13:30) - Outro Tony Hoare has died https://blog.computationalcomplexity.org/2026/03/tony-hoare-1934-2026.html https://news.ycombinator.com/item?id=47324054 Meta acquires Moltbook https://www.axios.com/2026/03/10/meta-facebook-moltbook-agent-social-network https://news.ycombinator.com/item?id=47323900 Yann LeCun raises $1B to build AI that understands the physical world https://www.wired.com/story/yann-lecun-raises-dollar1-billion-to-build-ai-that-understands-the-physical-world/ https://news.ycombinator.com/item?id=47320600 Create value for others and don’t worry about the returns https://geohot.github.io//blog/jekyll/update/2026/03/11/running-69-agents.html https://news.ycombinator.com/item?id=47332074 Agents that run while I sleep https://www.claudecodecamp.com/p/i-m-building-agents-that-run-while-i-sleep https://news.ycombinator.com/item?id=47327559 Debian decides not to decide on AI-generated contributions https://lwn.net/SubscriberLink/1061544/125f911834966dd0/ https://news.ycombinator.com/item?id=47324087 Cloudflare crawl endpoint https://developers.cloudflare.com/changelog/post/2026-03-10-br-crawl-endpoint/ https://news.ycombinator.com/item?id=47329557 U+237C ⍼ Is Azimuth https://ionathan.ch/2026/02/16/angzarr.html https://news.ycombinator.com/item?id=47329605 Universal vaccine against respiratory infections and allergens https://med.stanford.edu/news/all-news/2026/02/universal-vaccine.html https://news.ycombinator.com/item?id=47329608 RISC-V Is Sloooow https://marcin.juszkiewicz.com.pl/2026/03/10/risc-v-is-sloooow/ https://news.yco

Listen Ads-FREE at DjamgaMind: https://podcasts.apple.com/us/podcast/daily-news-rundown-the-fire-and-forget-office/id1864721054?i=1000754393738 🚀 Welcome to AI Unraveled. Today, the AI industry draws a line in the sand. We are covering Yann LeCun’s massive $1 billion raise to build "AMI Labs" and the unprecedented legal alliance between OpenAI, Google, and Anthropic against the Pentagon’s blacklist. This episode is made possible by our sponsors: 🛑 AIRIA: As Microsoft rolls out "Fire and Forget" agents through Copilot Cowork, governance isn't just an IT checkbox—it's a survival requirement. AIRIA is the control plane for your agentic workforce. 👉 Govern the Agentic Era: https://airia.com/request-demo/?utm_source=AI+Unraveled+&utm_medium=Podcast&utm_campaign=Q1+2026 🎙️ DjamgaMind: Tired of the ads? We hear you. We’ve launched an Ads-FREE Premium Feed called DjamgaMind . Get full, uninterrupted audio intelligence and deep-dive specials. 👉 Switch to Ads-Free: DjamgaMind on Apple Podcasts: In Today’s Briefing: The $1B Bet on World Models: Why Yann LeCun is moving beyond LLMs with AMI Labs. Industry Solidarity: Why Jeff Dean and 30+ OpenAI/Google workers are backing Anthropic’s suit against the Trump administration. Apple’s Siri Delay: Why the new Smart Home Display is stuck in the lab until September. The NemoClaw Leak: Nvidia’s open-source answer to the OpenClaw "Little Lobster" mania. Claude vs. Firefox: How a single AI found 20% of Firefox’s annual high-severity bugs in two weeks. The a16z Top 100: ChatGPT hits 900M weekly users, but the "Agentic" shift is real. Xiaomi Robots: Humanoids take over 90% of the assembly line tasks in Beijing. Credits: Created and produced by Etienne Noumen . Keywords: AMI Labs, Yann LeCun World Models, Anthropic Pentagon Lawsuit, Jeff Dean, Apple Siri Delay, Nvidia NemoClaw, OpenClaw Mania, Google Workspace Gemini, Copilot Cowork, a16z Consumer AI, Xiaomi Humanoid Robots, Claude Firefox Bugs, DjamgaMind, AI Unraveled. 🚀 Reach the Architects of the AI Revolution Want to reach 60,000+ Enterprise Architects and C-Suite leaders? Download our 2026 Media Kit and see how we simulate your product for the technical buyer: https://djamgamind.com/ai Connect with the host Etienne Noumen : https://www.linkedin.com/in/enoumen/ ⚗️ PRODUCTION NOTE : We Practice What We Preach. AI Unraveled is produced using a hybrid "Human-in-the-Loop" workflow. While all research, interviews, and strategic insights are curated by Etienne Noumen, we leverage advanced AI voice synthesis for our daily narration to ensure speed, consistency, and scale.

Will LLMs hit a structural ceiling in clinical medicine? Discover why Yann LeCun’s "World Models" are the essential next step for safe, autonomous Health AI. In this episode, we break down Meta AI Chief Yann LeCun’s blueprint for the future of AI and its specific implications for healthcare. We move beyond the hype of Large Language Models to explore how Energy-Based Models, Regularized Learning (JEPA), and Model-Predictive Control will solve the "hallucination" and safety problems in surgical robotics and complex physiology. Key Takeaways: • Why "Energy-Based Models" are more stable for ICU monitoring than standard probabilistic AI. • How JEPA (Joint-Embedding Predictive Architecture) allows AI to learn rare diseases without massive datasets. • Why "World Models" will replace Reinforcement Learning in the next generation of surgical robots. 0:00 Introduction 0:22 LLMs vs World Models 0:50 Energy Based Models 2:00 Clinical EBM Application 2:50 Learning Methods Comparison 3:30 JEPA For Rare Disease 4:25 RL vs MPC 5:15 MPC Clinical Simulations 6:25 DeepMind Genie Model 7:35 Transformer Architecture Limits 8:31 Future Modular Systems 9:08 Spatial Reasoning Advances 10:07 Strategic Focus Conclusion Health AI, Yann LeCun, World Models, Medical Robotics, JEPA, LLM limitations, Clinical AI, Surgical Automation, Machine Learning in Medicine. #HealthAI #MedicalAI #YannLeCun #WorldModels #MedTech #DigitalHealth #aiinmedicine Music generated by Mubert https://mubert.com/render healthaibrief@outlook.com

Tom sits down with Yann LeCun, the Jacob T. Schwartz Professor of Computer Science at NYU, and Executive Chairman of Advanced Machine Intelligence Labs. Yann is co-winner of the 2018 ACM Turing Award for his research in neural network learning. Yann takes us from his days as a postdoc working with Geoffrey Hinton, through his days as Chief AI Scientist at Facebook/Meta. His simultaneous roles as a Professor at NYU and Chief AI Scientist at a large AI provider give Yann a unique perspective on how technological advances and commercial forces combined to get us to today's state of the art.

Yann LeCun is one of the most influential figures in artificial intelligence. Alex LeBrun is the founder of Nabla and newly announced CEO of AMI Labs , a new AI research company he and Yann are building around a bold idea: large language models aren’t enough for medicine. In this special episode of How I Doctor , Dr. Graham Walker sits down in-person with Alex and Yann to explore the next frontier of AI in healthcare - world models. While today’s AI systems excel at predicting the next word, Yann argues that real clinical intelligence requires something deeper: models that can imagine, simulate, and plan. From the limitations of LLMs in high-stakes environments to the concept of building a “patient model” that can predict the consequences of treatment decisions, this episode dives into what it would actually take to build AI that reasons more like a physician. They discuss why documentation was the first breakthrough use case, how 80% accuracy fails in clinical settings, and why reliability, and not hype, will determine who wins in healthcare AI. This isn’t about replacing doctors. It’s about amplifying them. If AI is going to meaningfully change medicine, it won’t be through better chatbots. It will be through systems that understand the world. Watch or Listen 🎥 Watch the full video conversation now — exclusively on https://www.offcall.com/learn/podcast/ai-world-models-medicine-yann-lecun-alex-lebrun 🔊 Or stream the audio version on your favorite podcast platform. What You’ll Learn How predicting the next word isn’t the same as clinical reasoning and where LLMs fall short in medicine. What “world models” are and how they differ fundamentally from today’s large language models. Why 80% accuracy isn’t acceptable in healthcare and what reliability really means in clinical AI. Why medical coding may be one of the next frontiers for AI in clinical workflows. How AI assistants could amplify doctors the way a research lab amplifies a professor, by making clinicians smarter, not obsolete. 🩺 Offcall is more than a platform — it’s a community. Join today ! 📝 For a full transcript of this episode click HERE 🎧 Subscribe to receive new How I Doctor episodes directly in your feed here: https://episodes.fm/1767429315 👨‍⚕️Follow Dr. Graham Walker on LinkedIn https://www.linkedin.com/in/graham-walker-md/ IG https://www.instagram.com/ubergraham/ Blu

In the final episode of 2025, Scotty and Matt celebrate 33 episodes of Built 2 Scale by diving into Yann LeCun's ultimate entrepreneurial pivot, raising $3 billion in euros after getting ousted from Meta by Alexander Wang to work on spatial intelligence. They dissect why this is terrible news for Elysium (autonomous homes now have a 10 year delay), celebrate Sergey coding again at Google while the Qantas vs United business class wars rage on, and introduce the year end segment Receipts or Regrets where they review their boldest predictions. From Brett Adcock's 200x Apple claim to robots in homes by 2025, from AI in the avocado to Limitless getting acquired by Zuck with zero notice, they hold nothing back in this year end wrap up featuring domain buying confessions, builder vs coder rants, and why coders should never be called builders. Built 2 Scale | Episode 33 TIMESTAMPS: 0:00 Final Episode of 2025: 33 Episodes Complete 2:01 Yann LeCun Raises $3B for Spatial Intelligence Startup 4:06 Why This is Terrible News for Elysium Autonomous Homes 7:00 Brett Adcock's Figure AI Christmas Party: Robot Rave with Deadmau5 9:02 Voice AI Bandwidth Solution: Scotty's 30 Year Long Bet 13:44 Human Like Voice vs Fast Intelligence: What Do You Actually Want? 16:51 Sergey Back Coding at Google: The Return of the Founder 21:57 Receipts or Regrets: Year End Prediction Review Begins 23:41 Matt's Receipt: Robots in Homes by 2025 (Chinese Did It) 25:47 Scotty's Escalate: Make Every Australian a Millionaire With Raw Materials 28:52 Receipt: AI in the Avocado, Guzman y Gomez Down 55% 31:01 Regret: Sesame AI Bot in Bedroom, Wife Not Impressed 33:17 Regret: Steve Irwin Tech Talk in Dallas 35:38 Qantas Fanboy vs United Points: The Business Class Debate 40:09 Receipt: Peak Waymo, Tesla Has Long Game Sewn Up 44:19 Regret: First Year ARR is Nonsense, Y Combinator Circular Economy 46:39 Receipt: OpenAI Wants to Be Apple of AI (Johnny Ive Hire Confirmed It) 52:27 Rant: Coders Shouldn't Be Called Builders, Leave Us That One Term 56:26 Receipt: Just in Time Software Revolution Happening Now 58:52 Matt's Dirty Drunk Habit: Domain Buying, Sold Usainboat.com for $20 1:00:34 Limitless Acquired by Meta: Zuck Now Has All of Scotty's Dog Arguments This Episode Covers: Yann LeCun raising $3 billion in euros for spatial intelligence after Meta exit, choosing Europe where innovation goes to die Why Yann working on spatial intelligence is terrible news for autonomous homes timeline Brett Adcock throwing robot rave with Deadmau5 while still having no product after 3 years Voice AI bandwidth debate: Human like conversation vs fast accurate intelligence Sergey back coding at Google, spending 90% of time teaching rather than sitting on $500M yacht Year end Receipts or Regrets segment reviewing boldest predictions of 2025 Robots in homes by 2025: Chinese delivered with $20K Unitree, not Tesla or Figure AI in the avocado: Guzman y Gomez down 55% from peak, now $2B market cap First year ARR is nonsense: Y Combinator circular economy needs to exclude internal revenue OpenAI wants to be Apple of AI: Johnny Ive hire proved the hardware thesis The builder rant: Coders sitting in Starbucks with Frappuccinos aren't builders, leave us that one term Just in time software: LLMs writing code on the fly rather than predefined workflows Qantas vs United business class points arbitrage strategies KEY INSIGHTS: Yann's strategic retreat : Raising $3B in Europe for spatial intelligence after Meta exit shows classic researcher move to longer horizon tech when pressure mounts. Europe welcomes unproductive research with ope