Andrew Tan

Congratulations to Yann LeCun and the AMI Labs team. Following Fei-Fei Li’s World Labs, this is another major capital signal that serious money is moving into 𝐧𝐨𝐧-𝐋𝐋𝐌 paradigm. AMI Labs, just raised $1.03B at a $3.5B pre-money valuation to build AI around world models rather than LLMs. And it changed something deeper: the AI race is no longer just an LLM race. It has split into three camps. 1) 𝐓𝐡𝐞 𝐋𝐋𝐌 𝐂𝐚𝐦𝐩 The bet is simple: scale text prediction. LLMs demand massive data, compute, and energy as an ongoing requirement. Hallucinations are architectural. Periodic retraining is mandatory. The weaknesses are structural: lacks reliability, adaptability, predictability, and accountability. These systems are frozen after training, cannot learn continuously and brittle in the real world. That is why so much of today’s AI industry looks like scaffolding built around static models: wrappers, guardrails, orchestration layers, patches, and human babysitting. In many cases it is productivity theater with no real intelligence. 2) 𝐓𝐡𝐞 𝐖𝐨𝐫𝐥𝐝-𝐌𝐨𝐝𝐞𝐥 𝐂𝐚𝐦𝐩 This is where Yann LeCun’s AMI Labs and Fei-Fei Li’s World Labs sit. The bet here is to move beyond text into multimodal world prediction: learning from video, audio, space, and sensor data to support reasoning, planning, and eventually robotics and other physical-world systems. This is a real break from traditional LLM thinking. But it still appears to remain within the pre-trained-model paradigm: train → deploy → retrain Smarter than text prediction, yes. But not yet a full break from model-centric AI. This camp simulates a better model of the world by improving the representation of reality, but they do not create a system that truly learns. 3) 𝐓𝐡𝐞 𝐒𝐞𝐥𝐟-𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐀𝐈 𝐂𝐚𝐦𝐩 This camp starts from a different premise. Real intelligence is the ability to continuously learn, adapt, and apply what is learned across changing contexts. That means persistent memory, reasoning, understanding, dynamic world models, and autonomous adaptation through experience. These systems self-improve with experience. The result is compounding intelligence, not just faster or more fluent answers. This is developmental intelligence that starts with foundational knowledge, then develops upward through experience, without consuming planetary-scale data, compute, and energy. None of this can be brute-forced, it's the hardest and the least crowded path. This is the real paradigm shift: compounding ROI, no endless retraining, far lower resource requirements, and a cleaner path to real intelligence. AMI Labs is more ambitious than LLMs, but real intelligence comes from continuous learning and compouding capabilities through use in production. 𝐋𝐋𝐌𝐬 𝐠𝐞𝐧𝐞𝐫𝐚𝐭𝐞. 𝐖𝐨𝐫𝐥𝐝 𝐦𝐨𝐝𝐞𝐥𝐬 𝐬𝐢𝐦𝐮𝐥𝐚𝐭𝐞. 𝐒𝐞𝐥𝐟-𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐀𝐈 𝐬𝐲𝐬𝐭𝐞𝐦𝐬 𝐜𝐨𝐦𝐩𝐨𝐮𝐧𝐝. That is the emerging stack. The real AI race has only just begun.

54

9 Comments

Yann LeCun, the former chief AI scientist at Meta and a Turing Award winner, just raised $1.03 billion in seed funding for a new startup called Advanced Machine Intelligence Labs, or AMI. The round values the company at $3.5 billion and is believed to be the largest seed round ever raised by a European company. Investors include Nvidia, Temasek, and Jeff Bezos' Bezos Expeditions. LeCun has been one of the most vocal critics of the technology dominating the AI industry. His argument is that large language models, the systems behind ChatGPT, Claude, and Gemini, are fundamentally limited because they only learn from text. They can generate fluent language, but they do not understand the physical world. AMI is built around a different approach called world models. Instead of predicting the next word in a sequence, these systems learn from video, audio, and sensor data to build abstract representations of how the real world works. The underlying architecture, called JEPA, is a framework LeCun first proposed in 2022. The company plans to spend its first year entirely on research and development, with target industries including healthcare, robotics, and manufacturing. The founding team is drawn almost entirely from Meta's AI research organization. A billion dollars in seed funding for a company openly challenging the direction nearly every major AI lab is moving in. Most of the industry is betting on making language models bigger. LeCun is arguing the entire approach has a ceiling and that the next generation of AI will need to learn the way humans do, through experience with the physical world rather than text alone.

18

2 Comments

I’ve been working on the theory of why small, shippable factories are the perfect inversion to the centralized prefab factory model. It all collapsed into one line of math (credit to Gilles Retsin): Factory Efficiency = Usage × (Throughput × Product Value ) / ( CapEx + Deployment OpEx) The idea is that the closer you can drive Usage to 100 % while keeping the denominator tiny, the faster the flywheel spins.  A shippable microfactory tackles this by primarily focusing on a high Usage factor. Because the microfactory can be packed up afterwards and reused, its CapEx gets amortized over multiple projects, effectively raising its overall Usage across the year. If one project alone doesn’t fully occupy the factory, it can simply roll over to another job to maintain high Usage. Its CapEx is often an order of magnitude lower than a centralized factory (< $1M vs. $10-50M min), so the breakeven throughput is more achievable on a small pipeline of work. The real breakthrough here is portability. There is near-zero stranded capital and almost no idle time. To flesh this theory out, I wrote 3000 words on the topic for Brad Hargreaves and Thesis Driven. The article breaks down: - why “fixed factories, shipped goods” is being inverted to shipped factories, fixed goods - what a sub‑$1 M robotic cell does to CapEx per home versus a £45 M off‑site plant - early production/cost data on a shippable microfactory from Mollie Claypool, Gilles Retsin, Sam Baker and the Automated Architecture (AUAR) team - the hybrid reality of robots tackling the heavy structural components and human crews assembling faster and cheaper - technical constraints to pull this model off and future considerations on the optimal business model (own vs. rent) Major shoutout to Gilles Retsin for helping me pull this together. https://lnkd.in/eRTnbZPa

54

2 Comments

Reimagining RL for Long-Running, Complex LLM Agent Tasks Traditional RL methods like GRPO + tool masking (e.g., Retool-RL) work fine—if your agent only needs to make 3–5 tool calls. Think toy tasks: simple calculator math or RAG for basic QA. But let’s be real: Real-world agent tasks—like debugging code or developing features—are messy. They involve 20+ tool calls, long reasoning chains, and cascading errors that even top-tier models (like Claude Sonnet-4 or GPT-4.1) struggle to recover from. 🧠 So why not just train with RL on these long tasks? Some teams tried (e.g., SkyRL)… but RL barely moved the needle: SWE-bench went from 11% → 14.6%. That’s not enough. Here’s the problem: 🔁 RL on long-horizon tasks is bottlenecked at rollout time. In my experiments training vision agents at LandingAI, rollout (generating experience traces) took 75%+ of the training time. That’s not because backprop is slow—updating the model is cheap. It’s because each task takes hundreds of seconds to complete, and future tasks will take days. 💣 And worse, because success rate/action are low (say 20%), the chance of getting a single successful N-step trace is 1 / (0.2^N) to create positive learning signal, but then this quickly vanish in bunch of negative gradient you got from other failed traces. So what actually works? Curriculum. Break down the task → Train specialists → Synthesize good traces → Bootstrap a generalist → RL fine-tune. This “specialist-to-generalist” recipe increases early success rates and makes RL training actually feasible. 🧪 This approach is gaining momentum: ether0 (FutureHouse): Train RL specialists on chemistry subtasks, synthesize CoT + tool traces, and bootstrap a generalist with SFT before final RL tuning. SWE-Swiss (https://lnkd.in/g4TCpgSz): Break SWE-bench into Localization, Repair, and Unit Test Generation → train specialists → synthesize traces → build a generalist → RL on full tasks. Result: state-of-the-art 60.2% on SWE-bench Verified  🧩 The New Recipe (summarized): Decompose long tasks into easier, verifiable subtasks Train specialist models with RL for each subtask Use them to generate clean traces (Chain-of-Thought + tool use) Train a new generalist model on these synthetic traces Apply RL to the generalist on full, long-running tasks This boosts early success rate → stronger learning signal → higher sampling efficiency. If you’re building RL agents for real-world workflows, this curriculum-based approach is worth a serious look. Happy to share ideas or diagrams if you’re working on something similar. Let’s push this frontier forward 🚀 We're building the next generation bioinformatic agents at https://www.thedrylab.com/

25

Yann LeCun just raised $1.03B. Not for LLMs. For what comes after them. My first thought wasn't "wow." It was: how many founders are about to miss this? Most of us are sprinting to build on top of GPT or Claude architecture. LeCun just bet a billion dollars that architecture is wrong. World models don't predict text. They understand causality. Physics. How things actually work. Not just the next token. It has happened to me over and over again. Every time I got comfortable with the current stack, the stack moved. This is the stack moving again. The founders who win the next decade are already asking different questions. Not "how do I build faster on this?" "What comes after this?" $1.03B says the AI era isn't settled. It's barely started.

5

2 Comments

Allen Zhu (GSR Ventures): The AI app boom is imminent — the next ByteDance, Kuaishou, or Rednote was likely founded this year. At the 2025 Inclusion Bund Conference, Allen Zhu delivered a clear, unsentimental read on where AI is headed: the real opportunities are beyond the model. Key takeaways for builders & investors Model plateau → miniaturization. Under the Transformer paradigm, core “intelligence” gains are tapering; wins will come from UX and cost. Expect smaller, cheaper models, smarter data curation, and edge deployment. Apps are shifting text → voice → video. Token consumption is exploding in real use cases. Voice is breaking out now; ultra-low-latency video (~1s) could trigger the next wave. Moats live outside AI. Think workflows, editing/polish, proprietary/regulated data, operations, and specialized hardware (e.g., capture devices, smart name badges). Do the “boring, dirty work” others avoid. Agents: go vertical or get absorbed. Generic agents lack defensibility and will be eaten by advancing base models. Build sticky, domain-specific agents tied to real processes and outcomes. Commercialization > vibes. Retention is king. Beware “vibe revenue.” In SV, many VCs wait for ~$2M ARR post-launch; in China, sub-12-month paths to ~$5M ARR are becoming the bar. China founders’ edge & go-global. Strength in To-C experience design and hardware supply chains; look to Japan/SEA/MENA to avoid direct “model wars” and compete on execution. From UNIQUE RESEARCH’s tracking of thousands of AI apps globally, we see the same arc: the model dividend is flattening; winners are those who embed AI into specific scenarios, compress latency/cost, and build non-technical moats in data, workflow, and user insight. Bottom line: Don’t chase generic model horsepower. Ship products that deliver outcomes in real workflows—and measure yourself by retention, renewal, and unit economics. https://lnkd.in/gC63AABv

2

“Compute is a substitute for discipline.” Most frontier LLMs work because scale makes memorization cheap. When a model can afford to remember everything, it doesn’t need to understand much. That’s why hyperscale data centers are popping up everywhere, and why billions of dollars continue flowing into compute-intensive AI companies. Scale works—but it also masks inefficiency. I’m training a braille-native model on a single A100 under the opposite constraint: - No cheap surface-form memorization via scale - High information density per token - Architectural pressure toward invariants and concept reuse (braille contractions) - Scarcity by design, not by accident When you remove the escape hatches, abstraction stops being expensive. It becomes mandatory. In the 256 cells of a braille universe, there is nowhere for a hallucination to hide. Training perspective: If my entire training process were the height of one penny (1.52 mm): - GPT-4 would be roughly the height of the Burj Khalifa (828 m) - GPT-5 would be closer to a commercial jet’s cruising altitude (~11 km) This isn’t about “beating” GPT, Gemini, or Claude on benchmarks. It’s about intelligence per parameter and reasoning per joule. If a constrained model approaches similar parameter counts, it must internalize structure rather than rely on recall. Frontier models prove what scale can buy. Constraint-trained models test what intelligence actually is. I suspect compute was never the bottleneck—just a very effective way to avoid discipline.

6

Thrilled to share this news from my venture colleague Elise Stern investing in the $1.03B seed round of Yann LeCun and the AMI - Advanced Machine Intelligence team. A landmark moment for European AI. This investment is a reflection of a broader conviction we’ve been building at Eurazeo: Europe has an exceptional AI talent pool, and with the right capital and support, there is no structural reason why European companies can’t compete at the very top — alongside the best in the US and China. From the foundational model layer to the infrastructure layer and the agentic layer — we’re building conviction across the full stack, with European founders at the center: 🧠 Foundation models — Mistral AI, AMI - Advanced Machine Intelligence, Gradium 🗄️ Data infrastructure — Neo4j 🤖 Agents & enterprise AI — Dataiku , NiCE Cognigy At Eurazeo, our goal is simple: to be a partner of choice for AI companies building global champions from Europe. Yann LeCun, Alex LeBrun, Léopold Treppoz and the AMI team embody exactly that mission: world-class scientific ambition, executed from Europe, with global reach Congrats to my venture colleagues for this great addition Elise Stern, Thomas Turelier, Nicolas Debock, Benoist Grossmann, Carole du Fretay Cc Hala Fadel, Raluca Ragab, Anne-Charlotte Philbert, Maud Gillot, Alexandra Pederzani, Sophia Aghzadi, Ioana Iacobescu, Adrian Kreiser, Joséphine Nordin d’Arc, Richard Meyer Forsting, Inês Patrício, Justin Vibert, Louis Gautier, Francesco Mondini de Focatiis

99

2 Comments

✨ Whole Vibe: Anysphere’s AI Coding Tool Hits Different Santosh Rao just published our initiation report on Anysphere, the company behind Cursor, the AI code editor that’s revolutionizing software development. With $500MM ARR achieved in just 24 months and zero marketing spend, this represents one of the most compelling growth stories in enterprise AI. 🔑 Key Highlights: 🚀 Explosive growth from $0 to $500MM ARR in 14 months, nearly doubling every two months 👥 Over 1 million daily active users and 360,000+ paying subscribers across individual developers and Fortune 500 companies 🏢 Adopted by 50%+ of Fortune 500 companies including NVIDIA, Uber, Adobe, Stripe, OpenAI, and Spotify 💰 $1.07B total funding including $900MM Series C at $9.9B valuation (281% markup from December 2024) 🧠 Leading the “vibe coding” revolution where natural language prompts generate functional code ⚡ Exceptional team productivity: ~$5MM ARR per employee vs $125K median for private SaaS companies Why This Changes Everything: 1. Anysphere is spearheading the fundamental shift in how software gets built, with AI now generating 41% of all code globally in 2024 2. The company achieved $500MM ARR with zero traditional sales or marketing, proving exceptional product-market fit through pure viral adoption 3. Their AI-native approach and developer ecosystem create compounding network effects that become stronger as adoption scales ✍🏼 What strikes me about Anysphere isn’t just the growth metrics, it’s the paradigm shift they represent. When 92% of developers now use AI coding tools and companies like Google report 25% of new code is AI-generated, we’re witnessing the industrialization of software development. Anysphere didn’t just build a better code editor, they built the infrastructure for how all future software will be created. The fact that they reached $500MM ARR without spending a dollar on marketing tells you everything about product velocity and developer stickiness. This isn’t incremental improvement, it’s fundamental transformation of the $500B software engineering industry. MVR projects their 2025 revenue between $700MM-$1B, with a trajectory toward $2.5-3.3B by 2030. When every developer becomes 10x more productive through AI assistance, Anysphere is positioned to capture that value at unprecedented scale. Talk soon, Jc 🔗: https://lnkd.in/gn6jxFps #VentureCapital #MVPAllStarFund #TechInnovation #preipo #siliconvalley #unicorn #MVPALLSTARFUND #mvpvc #mvp #therealmvpvc #venturebytes #allstar #ipo #research #technology #tommorowsipostoday #SecondaryAsAService #saas #SecondariesAsAService #Anysphere #Cursor #AICoding #SoftwareDevelopment #DeveloperTools #ArtificialIntelligence #VibeCode #ThriveCapital #Accel #AndreessenHorowitz

27

2 Comments

Exciting News in AI! Yann LeCun, the former Chief AI Scientist at Meta and Turing Award winner, has successfully raised $1.03 billion in seed funding for his new venture, Advanced Machine Intelligence Labs (AMI), now valued at $3.5 billion. This remarkable funding round is considered the largest ever for a European startup, with notable investors including Nvidia and Jeff Bezos' Bezos Expeditions. LeCun, a passionate critic of current AI paradigms, challenges the limitations of large language models—like the frameworks behind ChatGPT and others—arguing they lack true understanding of the physical world. Instead, his innovative approach with AMI focuses on "world models," which learn from real-world data such as video and audio, rather than just text. With its founding team primarily from Meta’s AI research department, AMI aims to devote its first year solely to R&D, targeting sectors such as healthcare, robotics, and manufacturing. This billion-dollar backing highlights a pivotal shift in the AI landscape, as LeCun advocates for an evolution beyond mere language models, envisioning a future where AI learns from the world like we do—through experience. Stay tuned to see how AMI might redefine AI! Credit : Evolving AI Repost by : Salem Bagami

10

Anysphere, the San Francisco-based startup behind the AI-powered code editor Cursor, has secured $900 million in a funding round led by Thrive Capital, with participation from Andreessen Horowitz (a16z) and Accel. This investment propels Anysphere's valuation to approximately $9 billion, up from $2.5 billion in January 2025. Founded by MIT grads, Cursor has some major clients like Stripe, OpenAI, and Spotify. Popularised by Andrej Karpathy, the concept of “vibe coding”—where developers code more fluidly with AI—has given Cursor a cultural edge. But with competitors like GitHub Copilot, the real challenge will be maintaining this innovative momentum. Sualeh A., Michael T., Aman Sanger

2