[ { "id": "llm", "name": "Scale next-token prediction", "short": "LLM scaling", "order": 1, "color": "#2F5FD0", "bet": "Internet-scale text pretraining, then reinforcement learning on outcomes, then tools. Grounding, planning and memory are expected to emerge from scale.", "capital": { "text": "Most of the money in AI. Two labs filed to go public in June 2026 at $852B and $965B.", "sources": [ { "url": "https://techcrunch.com/2026/06/01/anthropic-files-to-go-public/", "date": "2026-06-01", "title": "Anthropic files to go public", "publisher": "TechCrunch" }, { "url": "https://techcrunch.com/2026/06/08/following-anthropic-openai-files-confidentially-for-ipo/", "date": "2026-06-08", "title": "OpenAI files confidentially for IPO, following Anthropic", "publisher": "TechCrunch" } ] }, "stances": { "signal": { "text": "Text, then verifiable rewards and human preference", "values": [ "text", "verifiable-reward", "human-preference" ] }, "when": { "text": "Offline pretrain, then freeze; memory is a retrieval layer", "values": [ "offline" ] }, "representation": { "text": "Tokens over a transformer", "values": [ "tokens" ] }, "acts_in": { "text": "Software: chat, code, browsers, APIs", "values": [ "software" ] } }, "ideas": [ { "name": "Scaling laws", "sources": [ { "url": "https://arxiv.org/abs/2001.08361", "date": "2020-01-23", "title": "Scaling Laws for Neural Language Models", "publisher": "arXiv" } ], "summary": "Loss falls predictably as you add parameters, data and compute, following a power law. Kaplan and colleagues measured it in 2020; it is the reason labs can budget a training run before running it." }, { "name": "RLHF", "sources": [ { "url": "https://arxiv.org/abs/2203.02155", "date": "2022-03-04", "title": "Training language models to follow instructions with human feedback (InstructGPT)", "publisher": "arXiv" } ], "summary": "Reinforcement learning from human feedback. Train a reward model on human preferences between answers, then optimize the language model against it. It turned GPT-3 into a usable assistant." }, { "name": "RL with verifiable rewards", "sources": [ { "url": "https://arxiv.org/abs/2501.12948", "date": "2025-01-22", "title": "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning", "publisher": "arXiv" } ], "summary": "Skip the human raters where an answer can be checked mechanically (maths, code, unit tests) and reward the model directly. DeepSeek-R1 showed long chains of reasoning emerging from this alone." }, { "name": "Test-time compute", "sources": [ { "url": "https://arxiv.org/abs/2408.03314", "date": "2024-08-06", "title": "Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters", "publisher": "arXiv" } ], "summary": "Spend more compute at answer time, by thinking longer or sampling many candidates, instead of only at training time. The o-series made it a product knob." }, { "name": "Mixture of experts", "sources": [ { "url": "https://arxiv.org/abs/2101.03961", "date": "2021-01-11", "title": "Switch Transformers", "publisher": "arXiv" } ], "summary": "Route each token through a few of many sub-networks so a trillion-parameter model costs a fraction of that to run. Now standard in every frontier and open-weight model." }, { "name": "Long context", "sources": [ { "url": "https://arxiv.org/abs/2403.05530", "date": "2024-03-08", "title": "Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context", "publisher": "arXiv" } ], "summary": "Windows of a million tokens or more let a model read a codebase or a book in one pass, reducing the need for retrieval." }, { "name": "Agents and tool use", "sources": [ { "url": "https://arxiv.org/abs/2210.03629", "date": "2022-10-06", "title": "ReAct: Synergizing Reasoning and Acting in Language Models", "publisher": "arXiv" } ], "summary": "Let the model call code, browsers and APIs in a loop. Most of the 2026 product surface, and the reason the camp claims planning without a planner." }, { "name": "Open weights vs closed", "sources": [ { "url": "https://www.wing.vc/content/chinas-open-weight-takeover", "date": "2026-05-01", "title": "China's open-weight takeover", "publisher": "Wing Venture Capital" } ], "summary": "Whether to publish model weights. China's labs mostly do; the American frontier mostly does not." } ], "status": { "text": "Dominant on revenue, compute and talent. Two IPO filings in June 2026 will put audited numbers behind the valuations. Open-weight Chinese labs keep the gap to the closed frontier under a year and took a majority of third-party API traffic in 2026. Talent is consolidating: Karpathy and Jumper both joined Anthropic in spring 2026. The internal split that matters is pretraining-scale believers versus post-training-scale believers.", "sources": [ { "url": "https://techcrunch.com/2026/06/01/anthropic-files-to-go-public/", "date": "2026-06-01", "title": "Anthropic files to go public", "publisher": "TechCrunch" }, { "url": "https://techcrunch.com/2026/06/08/following-anthropic-openai-files-confidentially-for-ipo/", "date": "2026-06-08", "title": "OpenAI files confidentially for IPO, following Anthropic", "publisher": "TechCrunch" }, { "url": "https://www.wing.vc/content/chinas-open-weight-takeover", "date": "2026-05-01", "title": "China's open-weight takeover", "publisher": "Wing Venture Capital" }, { "url": "https://techcrunch.com/2026/05/19/openai-co-founder-andrej-karpathy-joins-anthropics-pre-training-team/", "date": "2026-05-19", "title": "OpenAI co-founder Andrej Karpathy joins Anthropic's pre-training team", "publisher": "TechCrunch" } ] }, "proof": { "text": "Commercially proven already. The live question is whether it saturates on physical grounding, sample efficiency and post-deployment learning, which is exactly what every other camp bets on.", "sources": [] }, "sources": [], "explainer": "Take a transformer, train it to predict the next token on most of the written internet, then teach it manners with human feedback and reasoning with reinforcement learning on problems that have checkable answers. Give it tools. Repeat with more compute.\n\nThat recipe produced every product a layperson has used: ChatGPT, Claude, Gemini, Grok, DeepSeek, Kimi. The camp's claim is that the recipe keeps working. Grounding, planning and memory are expected to arrive as side effects of scale plus post-training, not as separate research programs.\n\nThe money reflects that confidence. Anthropic raised $65B at a $965B valuation in May 2026 and filed for an IPO on June 1; OpenAI filed a week later after a $122B round at $852B. Two Chinese labs, Zhipu and MiniMax, listed in Hong Kong in January. The open-weight labs in China keep the gap to the closed frontier under a year and have taken roughly half of third-party API traffic.\n\nThe fight inside the camp is about where the next gains come from: bigger pretraining runs, or more reinforcement learning after pretraining. Every other camp in this atlas is a bet on where this recipe runs out.", "disagreements": [ { "question": "Do gains now come from pretraining scale or from post-training?", "positions": [ { "who": "Pretraining scalers (OpenAI, Anthropic, xAI)", "claim": "Larger runs still deliver; the $100B-class rounds are for compute." }, { "who": "Post-training scalers (Thinking Machines, DeepSeek, Reflection)", "claim": "The frontier moved to reinforcement learning on outcomes; DeepSeek-R1 got reasoning from RL on a base model." } ], "sources": [ { "url": "https://arxiv.org/abs/2501.12948", "date": "2025-01-22", "title": "DeepSeek-R1", "publisher": "arXiv" }, { "url": "https://techcrunch.com/2025/07/15/mira-muratis-thinking-machines-lab-is-worth-12b-in-seed-round/", "date": "2025-07-15", "title": "Mira Murati's Thinking Machines Lab is worth $12B in seed round", "publisher": "TechCrunch" } ] }, { "question": "Open weights or closed?", "positions": [ { "who": "Open-weight labs (Qwen, DeepSeek, Moonshot, Mistral, Reflection)", "claim": "Release weights, price low, let distribution compound; Qwen passed Llama on downloads in 2026." }, { "who": "Closed labs (OpenAI, Anthropic, Google)", "claim": "Frontier capability and safety require control of the model; Meta itself went hosted-only with Muse in 2026." } ], "sources": [ { "url": "https://www.wing.vc/content/chinas-open-weight-takeover", "date": "2026-05-01", "title": "China's open-weight takeover", "publisher": "Wing Venture Capital" }, { "url": "https://builtin.com/artificial-intelligence/meta-superintelligence-labs", "date": "2026-03-10", "title": "Meta Superintelligence Labs: what we know so far", "publisher": "Built In" } ] }, { "question": "Is the recipe saturating?", "positions": [ { "who": "This camp", "claim": "No: reasoning models, agents and tool use keep opening new capability." }, { "who": "LeCun, Sutton, Chollet", "claim": "Yes on grounding, sample efficiency and post-deployment learning; those need a different signal, not more of the same." } ], "sources": [ { "url": "https://www.technologyreview.com/2026/01/22/1131661/yann-lecuns-new-venture-ami-labs/", "date": "2026-01-22", "title": "Yann LeCun's new venture is a contrarian bet against large language models | MIT Technology Review", "publisher": "MIT Technology Review" }, { "url": "https://betakit.com/ai-pioneer-richard-sutton-founds-new-research-lab/", "date": "2026-07-14", "title": "AI pioneer Richard Sutton founds new research lab", "publisher": "BetaKit" }, { "url": "https://arcprize.org/blog/arc-prize-2025-results-analysis", "date": "2025-12-05", "title": "ARC Prize 2025 results and analysis", "publisher": "ARC Prize" } ] } ], "watch": [ { "text": "The public S-1s: revenue, gross margin and compute commitments for OpenAI and Anthropic will be the first audited numbers for the camp.", "sources": [ { "url": "https://techcrunch.com/2026/06/01/anthropic-files-to-go-public/", "date": "2026-06-01", "title": "Anthropic files to go public", "publisher": "TechCrunch" }, { "url": "https://techcrunch.com/2026/06/08/following-anthropic-openai-files-confidentially-for-ipo/", "date": "2026-06-08", "title": "OpenAI files confidentially for IPO, following Anthropic", "publisher": "TechCrunch" } ] }, { "text": "Open-weight share of API traffic: Chinese providers went from under 2% to over 45% of OpenRouter tokens in a year.", "sources": [ { "url": "https://www.wing.vc/content/chinas-open-weight-takeover", "date": "2026-05-01", "title": "China's open-weight takeover", "publisher": "Wing Venture Capital" } ] }, { "text": "Whether the next generation of models improves on reasoning benchmarks that resist memorization, such as ARC-AGI-2 and 3, without a search wrapper.", "sources": [ { "url": "https://arcprize.org/blog/arc-prize-2025-results-analysis", "date": "2025-12-05", "title": "ARC Prize 2025 results and analysis", "publisher": "ARC Prize" } ] } ], "exposure": [ { "lab": "google-deepmind", "kind": "public", "note": "Alphabet (GOOGL). Gemini plus every other camp." }, { "lab": "meta", "kind": "public", "note": "Meta (META). Frontier models, hosted since 2026." }, { "lab": "microsoft-ai", "kind": "public", "note": "Microsoft (MSFT). In-house MAI models plus the OpenAI stake." }, { "lab": "xai", "kind": "public", "note": "SpaceX (SPCX) since June 2026; Grok is its AI division." }, { "lab": "zhipu", "kind": "public", "note": "Z.ai (HKEX 2513), the first listed LLM company." }, { "lab": "minimax", "kind": "public", "note": "MiniMax (HKEX 2610)." }, { "lab": "alibaba-qwen", "kind": "public", "note": "Alibaba (BABA). Qwen is a cloud demand generator." }, { "lab": "baidu", "kind": "public", "note": "Baidu (BIDU). ERNIE." }, { "lab": "nvidia", "kind": "indirect", "note": "Nvidia (NVDA) sells the compute; the purest way to own the camp without picking a lab." }, { "lab": "anthropic", "kind": "private", "note": "S-1 filed June 2026; listing expected late 2026." }, { "lab": "openai", "kind": "private", "note": "S-1 filed June 2026; timing undecided." }, { "lab": "deepseek", "kind": "private", "note": "First outside round June 2026; IPO preparations reported." }, { "lab": "moonshot", "kind": "private", "note": "Hong Kong IPO possible by end of 2026." } ], "reading_list": [ { "title": "Attention Is All You Need", "url": "https://arxiv.org/abs/1706.03762", "date": "2017-06-12", "why": "The architecture everything since is built on." }, { "title": "Scaling Laws for Neural Language Models", "url": "https://arxiv.org/abs/2001.08361", "date": "2020-01-23", "why": "The bet, stated as power laws: loss falls predictably with parameters, data and compute." }, { "title": "Language Models are Few-Shot Learners", "url": "https://arxiv.org/abs/2005.14165", "date": "2020-05-28", "why": "GPT-3. Scale produced in-context learning nobody trained for." }, { "title": "Training Compute-Optimal Large Language Models", "url": "https://arxiv.org/abs/2203.15556", "date": "2022-03-29", "why": "Chinchilla. Most models had been undertrained on data; the recipe changed overnight." }, { "title": "The Bitter Lesson", "url": "http://www.incompleteideas.net/IncIdeas/BitterLesson.html", "date": "2019-03-13", "why": "Sutton's essay that the scaling camp quotes more than he would like." }, { "title": "DeepSeek-R1", "url": "https://arxiv.org/abs/2501.12948", "date": "2025-01-22", "why": "Reasoning trained with verifiable rewards on top of a pretrained base; the second scaling axis." } ] }, { "id": "wm", "name": "World models", "short": "World models", "order": 2, "color": "#7A4FBF", "bet": "Predict the world, not text. Learn the dynamics of the environment from video and sensors so an agent can imagine outcomes and plan.", "capital": { "text": "More than $6B raised in 2026. A dozen labs, and the largest seed round in Europe.", "sources": [ { "url": "https://techcrunch.com/2026/03/09/yann-lecuns-ami-labs-raises-1-03-billion-to-build-world-models/", "date": "2026-03-09", "title": "TechCrunch", "publisher": "TechCrunch" }, { "url": "https://www.forbes.com/sites/josipamajic/2026/06/30/world-model-startups-raise-billions-as-vcs-bet-past-llms/", "date": "2026-06-30", "title": "World model startups raise billions as VCs bet past LLMs", "publisher": "Forbes" } ] }, "stances": { "signal": { "text": "Self-supervised prediction of video and latent state", "values": [ "video" ] }, "when": { "text": "Offline pretraining, but built to plan at inference by rolling out imagined futures", "values": [ "offline", "inference-planning" ] }, "representation": { "text": "Latent embeddings (JEPA) or generated pixels (video models). These sub-camps disagree.", "values": [ "latent", "pixels" ] }, "acts_in": { "text": "Simulated and physical environments; robotics and driving downstream", "values": [ "simulation", "physical" ] } }, "ideas": [ { "name": "JEPA", "sources": [ { "url": "https://openreview.net/forum?id=BZ5a1r-kVsf", "date": "2022-06-27", "title": "A Path Towards Autonomous Machine Intelligence (LeCun)", "publisher": "OpenReview" } ], "summary": "Joint-Embedding Predictive Architecture. Encode two views of the world, predict one embedding from the other, never decode to pixels. LeCun's 2022 position paper laid out the whole program." }, { "name": "Energy-based models", "sources": [ { "url": "http://yann.lecun.com/exdb/publis/pdf/lecun-06.pdf", "date": "2006-08-19", "title": "A Tutorial on Energy-Based Learning (LeCun et al.)", "publisher": "Predicting Structured Data, MIT Press" } ], "summary": "Instead of a probability over outputs, learn a scalar energy that is low for plausible states. Lets a model say \"this is compatible\" without enumerating every pixel." }, { "name": "Predict in latent space, not pixels", "sources": [ { "url": "https://openreview.net/forum?id=BZ5a1r-kVsf", "date": "2022-06-27", "title": "A Path Towards Autonomous Machine Intelligence (LeCun)", "publisher": "OpenReview" } ], "summary": "The JEPA camp's central argument: the world has too much unpredictable detail to model at the pixel level, so predict what matters and ignore the rest." }, { "name": "Model-based RL (Dreamer)", "sources": [ { "url": "https://arxiv.org/abs/1912.01603", "date": "2019-12-03", "title": "Dream to Control: Learning Behaviors by Latent Imagination", "publisher": "arXiv" } ], "summary": "Learn a compact world model from experience, then train the policy inside the model's imagination. Hafner's Dreamer line took this from Atari to real robots." }, { "name": "Interactive video generation", "sources": [ { "url": "https://arxiv.org/abs/2402.15391", "date": "2024-02-23", "title": "Genie: Generative Interactive Environments", "publisher": "arXiv" } ], "summary": "Generate the next frame conditioned on the player's action, fast enough to play. Genie 3 does it at 24 frames per second." }, { "name": "Spatial intelligence", "sources": [ { "url": "https://www.ted.com/talks/fei_fei_li_with_spatial_intelligence_ai_will_understand_the_real_world", "date": "2024-04-16", "title": "With spatial intelligence, AI will understand the real world (Fei-Fei Li, TED)", "publisher": "TED" } ], "summary": "Fei-Fei Li's term for models that perceive, generate and reason about 3D space rather than text. World Labs' Marble is the product version." }, { "name": "Persistent 3D scenes", "sources": [ { "url": "https://www.worldlabs.ai/blog/marble-world-model", "date": "2025-11-12", "title": "Marble: a multimodal world model", "publisher": "World Labs" } ], "summary": "A generated world that stays consistent when you turn around and come back, rather than re-imagining it each frame." } ], "status": { "text": "The money arrived before the results. Generative video world models (Genie 3, Cosmos 3) exist and are useful as simulators and data engines. JEPA-style latent models have no headline demonstration yet; AMI Labs has $1.03B and no product. LeCun argues pixel reconstruction wastes capacity; the video camp keeps shipping. World-model startups raised more than $6B in the first half of 2026.", "sources": [ { "url": "https://techcrunch.com/2026/03/09/yann-lecuns-ami-labs-raises-1-03-billion-to-build-world-models/", "date": "2026-03-09", "title": "TechCrunch", "publisher": "TechCrunch" }, { "url": "https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/", "date": "2025-08-05", "title": "Genie 3: a new frontier for world models", "publisher": "Google DeepMind" }, { "url": "https://www.forbes.com/sites/josipamajic/2026/06/30/world-model-startups-raise-billions-as-vcs-bet-past-llms/", "date": "2026-06-30", "title": "World model startups raise billions as VCs bet past LLMs", "publisher": "Forbes" }, { "url": "https://www.technologyreview.com/2026/01/22/1131661/yann-lecuns-new-venture-ami-labs/", "date": "2026-01-22", "title": "Yann LeCun's new venture is a contrarian bet against large language models | MIT Technology Review", "publisher": "MIT Technology Review" } ] }, "proof": { "text": "A latent world model that plans a multi-step physical task from raw video with far less data than a vision-language-action model, or one that beats LLM plus tools on long-horizon planning.", "sources": [] }, "sources": [], "explainer": "A language model predicts the next word. A world model predicts the next moment: what the scene will look like, where the objects will be, what happens if you push. Train that on video and sensor streams and you get something an agent can plan inside, by imagining outcomes before acting.\n\nTwo sub-camps disagree about what to predict. LeCun's JEPA line predicts abstract latent features and argues that generating pixels wastes capacity on irrelevant detail. The video camp (Genie, Cosmos, Luma, Runway, Decart) generates frames directly and points out that its models exist, ship, and produce training data for robots.\n\nThe money arrived before the results. AMI Labs raised a $1.03B seed with no product; World Labs raised $1B; Wayve, Waabi, Luma, Runway, Decart, Odyssey and General Intuition each raised hundreds of millions. Genie 3 can generate a playable world in real time, but no latent world model has yet planned a multi-step physical task from raw video better than a vision-language-action model does.", "disagreements": [ { "question": "Predict pixels or predict latents?", "positions": [ { "who": "LeCun, AMI Labs, Meta FAIR (V-JEPA)", "claim": "Reconstructing pixels forces the model to predict noise; predict in an abstract space and plan there." }, { "who": "Google DeepMind (Genie), Nvidia (Cosmos), Luma, Runway, Decart", "claim": "Generated video is useful today: interactive worlds, simulation data for robots, and products." } ], "sources": [ { "url": "https://openreview.net/forum?id=BZ5a1r-kVsf", "date": "2022-06-27", "title": "A Path Towards Autonomous Machine Intelligence (LeCun)", "publisher": "OpenReview" }, { "url": "https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/", "date": "2025-08-05", "title": "Genie 3: a new frontier for world models", "publisher": "Google DeepMind" }, { "url": "https://www.technologyreview.com/2026/01/22/1131661/yann-lecuns-new-venture-ami-labs/", "date": "2026-01-22", "title": "Yann LeCun's new venture is a contrarian bet against large language models | MIT Technology Review", "publisher": "MIT Technology Review" } ] }, { "question": "Is a world model a research program or a video feature?", "positions": [ { "who": "AMI Labs, World Labs, General Intuition", "claim": "World models are the foundation that replaces LLMs for anything physical." }, { "who": "Sceptics inside the LLM camp", "claim": "Video models are a media product; the planning claim is unproven, and multimodal LLMs absorb the useful parts." } ], "sources": [ { "url": "https://www.forbes.com/sites/josipamajic/2026/06/30/world-model-startups-raise-billions-as-vcs-bet-past-llms/", "date": "2026-06-30", "title": "World model startups raise billions as VCs bet past LLMs", "publisher": "Forbes" }, { "url": "https://techcrunch.com/2026/02/18/world-labs-lands-200m-from-autodesk-to-bring-world-models-into-3d-workflows/", "date": "2026-02-18", "title": "TechCrunch", "publisher": "TechCrunch" } ] } ], "watch": [ { "text": "AMI Labs' first public model. A $3.5B pre-money seed with nothing shipped sets a high bar; the first release will tell whether latent prediction beats video generation on any task.", "sources": [ { "url": "https://techcrunch.com/2026/03/09/yann-lecuns-ami-labs-raises-1-03-billion-to-build-world-models/", "date": "2026-03-09", "title": "TechCrunch", "publisher": "TechCrunch" } ] }, { "text": "Robot-learning results that credit a world model for sample efficiency, such as Nvidia GR00T N2, which is built on world-model research.", "sources": [ { "url": "https://nvidianews.nvidia.com/news/nvidia-expands-open-model-families-to-power-the-next-wave-of-agentic-physical-and-healthcare-ai", "date": "2026-03-16", "title": "NVIDIA expands open model families", "publisher": "NVIDIA" } ] }, { "text": "Whether driving companies (Wayve, Waabi, Tesla) attribute deployments to learned simulators rather than to more fleet data.", "sources": [ { "url": "https://www.cnbc.com/2026/02/24/wayve-fundraise-nvidia-microsoft.html", "date": "2026-02-24", "title": "CNBC", "publisher": "CNBC" } ] } ], "exposure": [ { "lab": "google-deepmind", "kind": "public", "note": "Alphabet (GOOGL): Genie 3, Veo." }, { "lab": "nvidia", "kind": "public", "note": "Nvidia (NVDA): Cosmos world models plus the GPUs." }, { "lab": "meta", "kind": "public", "note": "Meta (META): V-JEPA continues at FAIR, but the company moved its bet to camp 1." }, { "lab": "ami-labs", "kind": "private", "note": "$1.03B seed at $3.5B pre; Bezos, Nvidia, Samsung, Temasek." }, { "lab": "world-labs", "kind": "private", "note": "$1B round with $200M from Autodesk." }, { "lab": "wayve", "kind": "private", "note": "$8.6B; Microsoft, Nvidia, Uber and SoftBank on the cap table." }, { "lab": "runway", "kind": "private", "note": "$5.3B." }, { "lab": "luma", "kind": "private", "note": "$4B+; Saudi HUMAIN is the lead." }, { "lab": "decart", "kind": "private", "note": "$4B; reported Anthropic acquisition talks." }, { "lab": "waabi", "kind": "private", "note": "$3B; Uber and Volvo." } ], "reading_list": [ { "title": "World Models", "url": "https://arxiv.org/abs/1803.10122", "date": "2018-03-27", "why": "Ha and Schmidhuber: an agent that learns a compressed model of its environment and trains inside its own dream." }, { "title": "A Path Towards Autonomous Machine Intelligence", "url": "https://openreview.net/forum?id=BZ5a1r-kVsf", "date": "2022-06-27", "why": "LeCun's position paper. JEPA, hierarchical planning and why predicting pixels is the wrong target." }, { "title": "Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture", "url": "https://arxiv.org/abs/2301.08243", "date": "2023-01-19", "why": "I-JEPA: prediction in representation space, the first working piece of the plan." }, { "title": "Mastering Diverse Domains through World Models", "url": "https://arxiv.org/abs/2301.04104", "date": "2023-01-10", "why": "DreamerV3. One learned world model, many tasks, no per-task tuning." }, { "title": "Genie: Generative Interactive Environments", "url": "https://arxiv.org/abs/2402.15391", "date": "2024-02-23", "why": "Action-conditioned worlds learned from unlabeled video; the generative branch of the camp." }, { "title": "V-JEPA 2", "url": "https://arxiv.org/abs/2506.09985", "date": "2025-06-11", "why": "Video world model that plans on a real robot with no robot data in pretraining." } ] }, { "id": "phys", "name": "Physics AI and AI for science", "short": "Physics AI", "order": 3, "color": "#C2632A", "bet": "Ideas are cheap now; experiments are the bottleneck. Replace the lab with a universal learned simulator that also tells you which direction to improve.", "capital": { "text": "Over $15B, most of it one round. Prometheus alone raised $12B at $41B in June 2026.", "sources": [ { "url": "https://techcrunch.com/2026/06/11/jeff-bezoss-prometheus-raises-12b-to-build-an-artificial-general-engineer-for-the-physical-world/", "date": "2026-06-11", "title": "Jeff Bezos's Prometheus raises $12B to build an 'artificial general engineer' for the physical world", "publisher": "TechCrunch" }, { "url": "https://techcrunch.com/2025/09/30/former-openai-and-deepmind-researchers-raise-whopping-300m-seed-to-automate-science/", "date": "2025-09-30", "title": "Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science", "publisher": "TechCrunch" }, { "url": "https://www.physicsx.ai/newsroom/physicsx-announces-300m-series-c-to-accelerate-physics-ai-for-industrial-engineering", "date": "2026-06-08", "title": "PhysicsX announces $300M Series C", "publisher": "PhysicsX" } ] }, "stances": { "signal": { "text": "Simulation data plus the governing equations themselves, which act as a checkable reward", "values": [ "simulation", "physical-law" ] }, "when": { "text": "Offline pretraining, then optimization loops at inference: simulate, take a gradient, redesign", "values": [ "offline", "inference-optimisation" ] }, "representation": { "text": "Continuous fields in 3D plus time; neural operators, resolution-invariant", "values": [ "fields" ] }, "acts_in": { "text": "Design and discovery loops: materials, fusion, weather, devices, drugs", "values": [ "design-loops" ] } }, "ideas": [ { "name": "Neural operators (FNO)", "sources": [ { "url": "https://arxiv.org/abs/2010.08895", "date": "2020-10-18", "title": "Fourier Neural Operator for Parametric PDEs", "publisher": "arXiv" } ], "summary": "Networks that learn maps between functions, not between fixed grids, so one model handles any resolution. The Fourier neural operator (2020) made them fast." }, { "name": "FourCastNet", "sources": [ { "url": "https://arxiv.org/abs/2202.11214", "date": "2022-02-22", "title": "FourCastNet: A Global Data-driven High-resolution Weather Model", "publisher": "arXiv" } ], "summary": "Nvidia's 2022 global weather model on neural operators; the first to match numerical forecasts at a tiny fraction of the compute." }, { "name": "Direct 4D rollout", "sources": [ { "url": "https://techstartups.com/2026/08/25/ai-founders-walked-away-from-bezos-backed-prometheus-now-theyre-building-physics-ai-to-model-the-universe/", "date": "2026-08-25", "title": "AI founders walked away from Bezos-backed Prometheus. Now they’re building physics AI to model the universe - Tech Startups", "publisher": "Tech Startups" } ], "summary": "Predict a whole space-time field forward in one shot rather than stepping a solver. Accelerated Understanding's pitch." }, { "name": "PDE residual as reward", "sources": [ { "url": "https://arxiv.org/abs/1711.10561", "date": "2017-11-28", "title": "Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear PDEs", "publisher": "arXiv" } ], "summary": "Plug the model's prediction back into the governing equation; the size of the violation is a training signal you never run out of." }, { "name": "Differentiable simulation", "sources": [ { "url": "https://arxiv.org/abs/1910.00935", "date": "2019-10-01", "title": "DiffTaichi: Differentiable Programming for Physical Simulation", "publisher": "arXiv" } ], "summary": "Write the simulator so you can take gradients through it, then optimize a design by descent instead of by trial." }, { "name": "Autonomous labs", "sources": [ { "url": "https://www.nature.com/articles/s41586-023-06734-w", "date": "2023-11-29", "title": "An autonomous laboratory for the accelerated synthesis of novel materials (A-Lab)", "publisher": "Nature" } ], "summary": "Robots that synthesize and test what the model proposes, closing the loop without a human at the bench." }, { "name": "Cross-physics transfer", "sources": [ { "url": "https://arxiv.org/abs/2405.19101", "date": "2024-05-29", "title": "Poseidon: Efficient Foundation Models for PDEs", "publisher": "arXiv" } ], "summary": "One model pretrained across fluids, solids, electromagnetics and chemistry, on the bet that the maths overlaps." } ], "status": { "text": "Weather forecasting is already won by learned models. Materials, fusion and device design are early. Money moved fast in 2026: Prometheus raised $12B, PhysicsX and CuspAI raised $300M and $450M, and Accelerated Understanding came out of stealth in August with the most aggressive scale claim. Periodic and Lila bet on closing the loop with real robots in real labs.", "sources": [ { "url": "https://techcrunch.com/2026/06/11/jeff-bezoss-prometheus-raises-12b-to-build-an-artificial-general-engineer-for-the-physical-world/", "date": "2026-06-11", "title": "Jeff Bezos's Prometheus raises $12B to build an 'artificial general engineer' for the physical world", "publisher": "TechCrunch" }, { "url": "https://www.physicsx.ai/newsroom/physicsx-announces-300m-series-c-to-accelerate-physics-ai-for-industrial-engineering", "date": "2026-06-08", "title": "PhysicsX announces $300M Series C", "publisher": "PhysicsX" }, { "url": "https://techstartups.com/2026/08/25/ai-founders-walked-away-from-bezos-backed-prometheus-now-theyre-building-physics-ai-to-model-the-universe/", "date": "2026-08-25", "title": "AI founders walked away from Bezos-backed Prometheus. Now they’re building physics AI to model the universe - Tech Startups", "publisher": "Tech Startups" }, { "url": "https://techcrunch.com/2025/09/30/former-openai-and-deepmind-researchers-raise-whopping-300m-seed-to-automate-science/", "date": "2025-09-30", "title": "Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science", "publisher": "TechCrunch" } ] }, "proof": { "text": "A device, material or molecule designed in the loop by the model that beats human plus classical simulation on a real-world benchmark, not just screened from a candidate list.", "sources": [] }, "sources": [], "explainer": "Ideas are cheap now; experiments are the bottleneck. This camp wants a learned simulator good enough to replace most of the lab bench, and a loop around it that says which direction to improve a design.\n\nThe signal is different from text: simulation data, plus the governing equations themselves, which act as a reward you can check. The representation is continuous fields in space and time, handled by neural operators that work at any resolution. The proof case already exists: learned weather models beat the best physics-based forecasts from 2023 on.\n\nThe camp has two strands, and the lab pages below are sorted by them. Simulator-first labs (Prometheus, PhysicsX, Accelerated Understanding, Physical Superintelligence) bet on scale in the model: a learned simulator good enough to design inside. Lab-in-the-loop companies (Periodic, Lila, Radical, CuspAI, and the drug designers Isomorphic, Chai and Latent) bet on the loop itself: a model proposes, an automated lab tests, the result trains the next model. Bezos's Prometheus, at $41B, is the largest bet on the first strand; drug design is the most mature part of the second.", "disagreements": [ { "question": "Scale the simulator, or close the loop with a real lab?", "positions": [ { "who": "Accelerated Understanding, PhysicsX, Physical Superintelligence", "claim": "A large enough learned simulator over the equations replaces most experiments." }, { "who": "Periodic Labs, Lila, Radical AI", "claim": "Models are wrong in ways only a real experiment reveals; the product is the automated lab." } ], "sources": [ { "url": "https://techstartups.com/2026/08/25/ai-founders-walked-away-from-bezos-backed-prometheus-now-theyre-building-physics-ai-to-model-the-universe/", "date": "2026-08-25", "title": "AI founders walked away from Bezos-backed Prometheus. Now they’re building physics AI to model the universe - Tech Startups", "publisher": "Tech Startups" }, { "url": "https://techcrunch.com/2025/09/30/former-openai-and-deepmind-researchers-raise-whopping-300m-seed-to-automate-science/", "date": "2025-09-30", "title": "Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science", "publisher": "TechCrunch" }, { "url": "https://www.technologyreview.com/2025/12/15/1129210/ai-materials-science-discovery-startups-investment/", "date": "2025-12-15", "title": "AI materials discovery now needs to move into the real world | MIT Technology Review", "publisher": "MIT Technology Review" } ] }, { "question": "Is Prometheus in this camp?", "positions": [ { "who": "Prometheus", "claim": "It is building an 'artificial general engineer' for design and manufacturing and says it has nothing to do with robotics." }, { "who": "This atlas", "claim": "Its stated loop (design, predict performance, manufacture) is the physics camp's bet at $41B, so it is listed here and flagged." } ], "sources": [ { "url": "https://techcrunch.com/2026/06/11/jeff-bezoss-prometheus-raises-12b-to-build-an-artificial-general-engineer-for-the-physical-world/", "date": "2026-06-11", "title": "Jeff Bezos's Prometheus raises $12B to build an 'artificial general engineer' for the physical world", "publisher": "TechCrunch" } ] } ], "watch": [ { "text": "A material, device or molecule designed in the loop that beats human-plus-classical-simulation on a real benchmark, not screened from a candidate list.", "sources": [ { "url": "https://www.technologyreview.com/2025/12/15/1129210/ai-materials-science-discovery-startups-investment/", "date": "2025-12-15", "title": "AI materials discovery now needs to move into the real world | MIT Technology Review", "publisher": "MIT Technology Review" } ] }, { "text": "Accelerated Understanding's first customer results after its August 2026 launch; the 5-trillion-data-point claim is unverified outside the company.", "sources": [ { "url": "https://techstartups.com/2026/08/25/ai-founders-walked-away-from-bezos-backed-prometheus-now-theyre-building-physics-ai-to-model-the-universe/", "date": "2026-08-25", "title": "AI founders walked away from Bezos-backed Prometheus. Now they’re building physics AI to model the universe - Tech Startups", "publisher": "Tech Startups" } ] }, { "text": "Clinical-stage readouts from Isomorphic, the camp's most advanced commercial test.", "sources": [ { "url": "https://techcrunch.com/2025/03/31/alphabets-ai-drug-discovery-platform-isomorphic-labs-raises-600m-from-thrive/", "date": "2025-03-31", "title": "Alphabet's AI drug discovery platform Isomorphic Labs raises $600M from Thrive", "publisher": "TechCrunch" } ] } ], "exposure": [ { "lab": "google-deepmind", "kind": "public", "note": "Alphabet (GOOGL): GraphCast, AlphaFold lineage, Isomorphic." }, { "lab": "nvidia", "kind": "public", "note": "Nvidia (NVDA): neural operators, Earth-2, and an investor in PhysicsX, Periodic, Lila and Radical." }, { "lab": "prometheus", "kind": "private", "note": "$41B; JPMorgan, BlackRock, Goldman on the round." }, { "lab": "periodic-labs", "kind": "private", "note": "$1.3B seed valuation; reported talks at $7.5B." }, { "lab": "physicsx", "kind": "private", "note": "$2.4B; Siemens and Applied Materials as strategics." }, { "lab": "lila-sciences", "kind": "private", "note": "$1.3B+; Flagship-incubated." }, { "lab": "isomorphic-labs", "kind": "private", "note": "Alphabet subsidiary with outside investors since 2025." } ], "strands": [ { "id": "simulator-first", "name": "Simulator first", "text": "Learn a universal simulator of physical systems (neural operators, field models, 'world engines') and optimize designs inside it. Prometheus, PhysicsX, Accelerated Understanding, Orbital." }, { "id": "lab-in-the-loop", "name": "Lab in the loop", "text": "Models propose molecules or materials, automated labs and instruments test them, the results train the next model. The signal here is real experiments, not simulation, which is what separates this strand from the first. Isomorphic, Chai, Latent, Recursion, Lila, Periodic, Radical, FutureHouse, Adaptyv." } ], "reading_list": [ { "title": "Physics Informed Deep Learning (Part I)", "url": "https://arxiv.org/abs/1711.10561", "date": "2017-11-28", "why": "PINNs. The governing equations become part of the loss." }, { "title": "Fourier Neural Operator for Parametric Partial Differential Equations", "url": "https://arxiv.org/abs/2010.08895", "date": "2020-10-18", "why": "Learn the solution operator, not one solution; resolution-invariant and orders of magnitude faster than solvers." }, { "title": "Highly accurate protein structure prediction with AlphaFold", "url": "https://www.nature.com/articles/s41586-021-03819-2", "date": "2021-07-15", "why": "The result that convinced funders a learned model can replace an experiment." }, { "title": "Neural Operator: Learning Maps Between Function Spaces", "url": "https://arxiv.org/abs/2108.08481", "date": "2021-08-19", "why": "The general framework behind the simulator-first strand." }, { "title": "Scaling deep learning for materials discovery", "url": "https://www.nature.com/articles/s41586-023-06735-9", "date": "2023-11-29", "why": "GNoME: 2.2 million new crystals predicted, hundreds made in a lab. The lab-in-the-loop strand in one paper." } ] }, { "id": "rl", "name": "Learn from experience", "short": "Experience / RL", "order": 4, "color": "#1F8A5B", "bet": "Intelligence is online reinforcement learning from a reward stream. Static datasets and the train-then-deploy split are a detour, and current deep learning cannot learn continually.", "capital": { "text": "Small by design, with one $1.1B exception. Silver's Ineffable Intelligence raised $1.1B at $5.1B in April 2026.", "sources": [ { "url": "https://techcrunch.com/2026/04/27/deepminds-david-silver-just-raised-1-1b-to-build-an-ai-that-learns-without-human-data/", "date": "2026-04-27", "title": "DeepMind's David Silver just raised $1.1B to build an AI that learns without human data", "publisher": "TechCrunch" }, { "url": "https://betakit.com/ai-pioneer-richard-sutton-founds-new-research-lab/", "date": "2026-07-14", "title": "AI pioneer Richard Sutton founds new research lab", "publisher": "BetaKit" } ] }, "stances": { "signal": { "text": "Scalar reward from interaction, plus self-generated subgoals (options)", "values": [ "scalar-reward" ] }, "when": { "text": "Continually, at runtime, forever. No train/deploy boundary", "values": [ "continual" ] }, "representation": { "text": "Whatever the agent builds online: learned features and options, no replay buffer", "values": [ "online-features" ] }, "acts_in": { "text": "Any environment, ultimately embodied. Target: 20 watts", "values": [ "physical", "simulation" ] } }, "ideas": [ { "name": "The Bitter Lesson", "sources": [ { "url": "http://www.incompleteideas.net/IncIdeas/BitterLesson.html", "date": "2019-03-13", "title": "The Bitter Lesson (Sutton)", "publisher": "incompleteideas.net" } ], "summary": "Sutton's 2019 essay: methods that use more computation (search and learning) always beat methods that encode human knowledge. Every camp quotes it; this one takes it literally." }, { "name": "Alberta Plan (12 steps)", "sources": [ { "url": "https://arxiv.org/abs/2208.11173", "date": "2022-08-23", "title": "The Alberta Plan for AI Research", "publisher": "arXiv" } ], "summary": "A 2022 roadmap from Sutton, Bowling and Pilarski for building an agent that learns continually, from prediction of a single signal up to planning with learned options." }, { "name": "OaK architecture", "sources": [ { "url": "https://www.youtube.com/watch?v=gEbbGyNkR2U", "date": "2025-08-08", "title": "The OaK Architecture: A Vision of SuperIntelligence from Experience (Sutton, RLC 2025 keynote)", "publisher": "Reinforcement Learning Conference" } ], "summary": "Options and Knowledge: an agent that invents its own sub-goals (options) and learns models of them, all online. The namesake of Oak Lab." }, { "name": "Continual backprop / plasticity", "sources": [ { "url": "https://www.nature.com/articles/s41586-024-07711-7", "date": "2024-08-21", "title": "Loss of plasticity in deep continual learning", "publisher": "Nature" } ], "summary": "Deep networks lose the ability to learn new things the longer they train. A 2024 Nature paper measured it and proposed reinitialising unused units." }, { "name": "Era of Experience", "sources": [ { "url": "https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf", "date": "2025-04-10", "title": "Welcome to the Era of Experience (Silver & Sutton)", "publisher": "Google DeepMind" } ], "summary": "Silver and Sutton's 2025 essay: the age of human data is ending; the next leap comes from agents generating their own experience." }, { "name": "Reward is enough", "sources": [ { "url": "https://doi.org/10.1016/j.artint.2021.103535", "date": "2021-10-01", "title": "Reward is enough (Silver, Singh, Precup, Sutton)", "publisher": "Artificial Intelligence" } ], "summary": "The 2021 claim that a single scalar reward, maximised in a rich environment, is sufficient to produce every ability we call intelligence." }, { "name": "Big-world hypothesis", "sources": [ { "url": "https://khurramjaved.com/the_big_world_hypothesis.html", "date": "2024-08-09", "title": "The Big World Hypothesis and its Ramifications for Artificial Intelligence (Javed & Sutton, RLC 2024 workshop)", "publisher": "khurramjaved.com" } ], "summary": "The world is far larger than any agent, so no agent can be trained to completion; it has to keep approximating and keep learning." } ], "status": { "text": "The camp doubled in size in 2026. Silver left DeepMind in January and raised $1.1B by April; Sutton and Javed left Keen in July to found Oak Lab in Toronto. Public results are toy-scale so far. The plasticity and catastrophic-forgetting problems are named but not solved. Sutton gives one-in-four odds of human-level AI by 2030.", "sources": [ { "url": "https://techcrunch.com/2026/04/27/deepminds-david-silver-just-raised-1-1b-to-build-an-ai-that-learns-without-human-data/", "date": "2026-04-27", "title": "DeepMind's David Silver just raised $1.1B to build an AI that learns without human data", "publisher": "TechCrunch" }, { "url": "https://betakit.com/ai-pioneer-richard-sutton-founds-new-research-lab/", "date": "2026-07-14", "title": "AI pioneer Richard Sutton founds new research lab", "publisher": "BetaKit" }, { "url": "https://betakit.com/new-turing-award-winner-richard-sutton-calls-doomers-out-of-line-talks-path-to-human-like-ai/", "date": "2025-03-05", "title": "Turing Award winner Richard Sutton talks path to human-like AI", "publisher": "BetaKit" }, { "url": "https://www.nature.com/articles/s41586-024-07711-7", "date": "2024-08-21", "title": "Loss of plasticity in deep continual learning", "publisher": "Nature" } ] }, "proof": { "text": "An agent that keeps improving after deployment in a non-stationary environment without forgetting, at a scale where frozen transformers plateau.", "sources": [] }, "sources": [], "explainer": "The other camps train once, then freeze. This one says that is the mistake. Intelligence is an agent learning continually from a stream of experience and reward, building its own features and sub-goals as it goes, with no line between training and deployment.\n\nSutton's Bitter Lesson and the Alberta Plan are the founding documents; the 2025 essay 'Welcome to the Era of Experience' by Silver and Sutton is the manifesto. The engineering problem is that current deep networks lose plasticity and forget: a network trained on one thing for long enough stops being able to learn the next thing.\n\nUntil 2026 the camp was deliberately tiny. Then David Silver left DeepMind, raised $1.1B for Ineffable Intelligence, and Sutton left Keen to found Oak Lab with a target of a trillion-parameter agent learning in real time on 20 watts. Prime Intellect sells the tooling; Adaption Labs sells models that learn on the fly. Public results are still at toy scale.", "disagreements": [ { "question": "Can today's deep learning learn continually, or does it need new primitives?", "positions": [ { "who": "Sutton, Oak Lab", "claim": "Current methods 'need not more tweaks, but fundamentally new ideas and a thorough reworking'; loss of plasticity is structural." }, { "who": "Silver, Ineffable; Prime Intellect; Adaption Labs", "claim": "Reinforcement learning at scale on top of existing architectures can get there; the missing piece is the experience stream, not the network." } ], "sources": [ { "url": "https://betakit.com/ai-pioneer-richard-sutton-founds-new-research-lab/", "date": "2026-07-14", "title": "AI pioneer Richard Sutton founds new research lab", "publisher": "BetaKit" }, { "url": "https://techcrunch.com/2026/04/27/deepminds-david-silver-just-raised-1-1b-to-build-an-ai-that-learns-without-human-data/", "date": "2026-04-27", "title": "DeepMind's David Silver just raised $1.1B to build an AI that learns without human data", "publisher": "TechCrunch" }, { "url": "https://www.nature.com/articles/s41586-024-07711-7", "date": "2024-08-21", "title": "Loss of plasticity in deep continual learning", "publisher": "Nature" }, { "url": "https://techcrunch.com/2026/07/08/prime-intellect-raises-130m-series-a-to-help-enterprises-build-their-own-ai-agents/", "date": "2026-07-08", "title": "Prime Intellect raises $130M Series A to help enterprises build their own AI agents", "publisher": "TechCrunch" } ] }, { "question": "Is the LLM camp already doing this?", "positions": [ { "who": "LLM labs", "claim": "RL on verifiable rewards plus agents that act in the world is learning from experience." }, { "who": "This camp", "claim": "That is offline RL on a frozen model with a replay buffer; nothing learns after deployment." } ], "sources": [ { "url": "https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf", "date": "2025-04-10", "title": "Welcome to the Era of Experience (Silver & Sutton)", "publisher": "Google DeepMind" }, { "url": "https://arxiv.org/abs/2501.12948", "date": "2025-01-22", "title": "DeepSeek-R1", "publisher": "arXiv" } ] } ], "watch": [ { "text": "An agent that keeps improving after deployment in a non-stationary environment without forgetting, at a scale where frozen transformers plateau.", "sources": [ { "url": "https://www.nature.com/articles/s41586-024-07711-7", "date": "2024-08-21", "title": "Loss of plasticity in deep continual learning", "publisher": "Nature" } ] }, { "text": "Oak Lab's first results against the 1T-parameters-at-20W target; Sutton gives one-in-four odds of human-level AI by 2030.", "sources": [ { "url": "https://betakit.com/ai-pioneer-richard-sutton-founds-new-research-lab/", "date": "2026-07-14", "title": "AI pioneer Richard Sutton founds new research lab", "publisher": "BetaKit" }, { "url": "https://betakit.com/new-turing-award-winner-richard-sutton-calls-doomers-out-of-line-talks-path-to-human-like-ai/", "date": "2025-03-05", "title": "Turing Award winner Richard Sutton talks path to human-like AI", "publisher": "BetaKit" } ] }, { "text": "Ineffable's first system: $1.1B buys a real test of whether 'no human data' can work beyond games.", "sources": [ { "url": "https://techcrunch.com/2026/04/27/deepminds-david-silver-just-raised-1-1b-to-build-an-ai-that-learns-without-human-data/", "date": "2026-04-27", "title": "DeepMind's David Silver just raised $1.1B to build an AI that learns without human data", "publisher": "TechCrunch" } ] } ], "exposure": [ { "lab": "google-deepmind", "kind": "public", "note": "Alphabet (GOOGL) keeps the AlphaZero lineage, minus Silver." }, { "lab": "ineffable-intelligence", "kind": "private", "note": "$5.1B; Sequoia and Lightspeed." }, { "lab": "prime-intellect", "kind": "private", "note": "$1B; the picks-and-shovels play for RL environments." }, { "lab": "adaption-labs", "kind": "private", "note": "$50M seed." }, { "lab": "oak-lab", "kind": "private", "note": "Funding undisclosed." }, { "lab": "keen-technologies", "kind": "private", "note": "$20M in 2022; Carmack-funded." } ], "reading_list": [ { "title": "Reinforcement Learning: An Introduction (2nd ed.)", "url": "http://incompleteideas.net/book/the-book-2nd.html", "date": "2018", "why": "Sutton and Barto. Read chapters 1, 3 and 6 before anything else." }, { "title": "Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm", "url": "https://arxiv.org/abs/1712.01815", "date": "2017-12-05", "why": "AlphaZero. Superhuman play from self-generated experience alone." }, { "title": "Reward is Enough", "url": "https://www.deepmind.com/publications/reward-is-enough", "date": "2021-05", "why": "Silver, Singh, Precup and Sutton: the hypothesis that reward maximization is sufficient for intelligence." }, { "title": "The Alberta Plan for AI Research", "url": "https://arxiv.org/abs/2208.11173", "date": "2022-08-23", "why": "A twelve-step program for continual, online agents. The camp's roadmap." }, { "title": "Loss of plasticity in deep continual learning", "url": "https://www.nature.com/articles/s41586-024-07711-7", "date": "2024-08-21", "why": "Why standard deep networks stop learning when trained forever, and one fix." }, { "title": "Welcome to the Era of Experience", "url": "https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf", "date": "2025-04", "why": "Silver and Sutton's case that human data has run out and agents must generate their own." } ] }, { "id": "robot", "name": "Physical AI and robotics", "short": "Physical AI", "order": 5, "color": "#B0345C", "bet": "The architecture is roughly settled; data is the bottleneck. Scale vision-language-action models on teleoperation, human video and simulation until robots generalize.", "capital": { "text": "Tens of billions including hardware. Figure is valued at $39B, Skild at $14B; Unitree listed at $9B and closed its first day near $47B.", "sources": [ { "url": "https://www.figure.ai/news/series-c", "date": "2025-09-16", "title": "Figure exceeds $1B in Series C funding at $39B post-money valuation", "publisher": "Figure" }, { "url": "https://techcrunch.com/2026/01/14/robotic-software-maker-skild-ai-hits-14b-valuation/", "date": "2026-01-14", "title": "Robotics software maker Skild AI hits $14B valuation", "publisher": "TechCrunch" }, { "url": "https://english.sse.com.cn/news/newsrelease/voice/c/c_20260806_10828128.shtml", "date": "2026-08-06", "title": "Unitree Robotics kicks off STAR Market IPO pricing", "publisher": "Shanghai Stock Exchange / Global Times" } ] }, "stances": { "signal": { "text": "Teleop demonstrations, human video, sim-to-real, then on-robot RL", "values": [ "demonstrations", "simulation", "scalar-reward" ] }, "when": { "text": "Offline, with on-robot RL fine-tuning starting to matter", "values": [ "on-robot-finetune" ] }, "representation": { "text": "Mostly camp 1: a vision-language model with a diffusion or flow action head", "values": [ "vla" ] }, "acts_in": { "text": "Actuators in the real world", "values": [ "physical" ] } }, "ideas": [ { "name": "Vision-language-action models", "sources": [ { "url": "https://arxiv.org/abs/2307.15818", "date": "2023-07-28", "title": "RT-2: Vision-Language-Action Models", "publisher": "arXiv" } ], "summary": "Take a vision-language model and add an output head that emits robot actions. RT-2 (2023) showed the language pretraining transfers to manipulation." }, { "name": "Cross-embodiment datasets", "sources": [ { "url": "https://arxiv.org/abs/2310.08864", "date": "2023-10-13", "title": "Open X-Embodiment", "publisher": "arXiv" } ], "summary": "Pool demonstrations from many different robots so one policy learns from all of them. Open X-Embodiment gathered 22 robot types." }, { "name": "Flow-matching action heads", "sources": [ { "url": "https://arxiv.org/abs/2410.24164", "date": "2024-10-31", "title": "π0: A Vision-Language-Action Flow Model", "publisher": "arXiv" } ], "summary": "Generate continuous action chunks with a diffusion or flow model instead of predicting one discrete action at a time. pi0 made it the default." }, { "name": "Sim-to-real (Isaac)", "sources": [ { "url": "https://arxiv.org/abs/1703.06907", "date": "2017-03-20", "title": "Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World", "publisher": "arXiv" } ], "summary": "Train in a physics simulator with randomised appearance and dynamics so the policy survives the transfer to a real robot." }, { "name": "Humanoid form factor", "sources": [ { "url": "https://www.figure.ai/master-plan", "date": "2024-03-01", "title": "Figure master plan", "publisher": "Figure" } ], "summary": "Build the robot to fit a world designed for humans. Expensive, but every tool and doorway already fits." }, { "name": "Data flywheels from deployment", "sources": [ { "url": "https://arxiv.org/abs/2511.14759", "date": "2025-11-18", "title": "π*0.6: a VLA That Learns From Experience", "publisher": "arXiv" } ], "summary": "Every deployed robot collects data that trains the next policy. Tesla's FSD model, applied to bodies." } ], "status": { "text": "Less a scientific bet than an application layer: it imports architectures from camp 1 and simulators from camp 2. Real deployments exist in warehouses and laundries; humanoid demos are still teleop-heavy and Tesla missed its 2025 Optimus target entirely. Valuations moved faster than deployments: Skild tripled to $14B in seven months, Unitree closed its first trading day up 460%. Levine's wing is philosophically closer to the experience camp.", "sources": [ { "url": "https://techcrunch.com/2026/01/14/robotic-software-maker-skild-ai-hits-14b-valuation/", "date": "2026-01-14", "title": "Robotics software maker Skild AI hits $14B valuation", "publisher": "TechCrunch" }, { "url": "https://english.sse.com.cn/news/newsrelease/voice/c/c_20260806_10828128.shtml", "date": "2026-08-06", "title": "Unitree Robotics kicks off STAR Market IPO pricing", "publisher": "Shanghai Stock Exchange / Global Times" }, { "url": "https://electrek.co/2026/04/22/tesla-optimus-production-fremont-model-sx-line/", "date": "2026-04-22", "title": "Tesla pushes Optimus V3 reveal later this year - again", "publisher": "Electrek" }, { "url": "https://www.bloomberg.com/news/articles/2025-11-20/robotics-startup-physical-intelligence-valued-at-5-6-billion-in-new-funding", "date": "2025-11-20", "title": "Robotics startup Physical Intelligence valued at $5.6 billion", "publisher": "Bloomberg" } ] }, "proof": { "text": "A general-purpose robot doing unseen household tasks at useful reliability without per-site data collection.", "sources": [] }, "sources": [], "explainer": "The robotics camp mostly agrees with the LLM camp about architecture: a vision-language model with an action head, trained on teleoperated demonstrations, human video and simulation, then fine-tuned on the robot. What it needs is data, and the whole field is organized around getting it: fleets, teleoperation farms, simulators, and cheap hardware.\n\nThe proof case is a robot doing a task in a home it has never seen. Physical Intelligence's pi0.5 showed generalization of that kind in 2025; Figure reports robots on a BMW line; Agility's Digit has the most warehouse hours. Humanoid demos are still teleop-heavy, and Tesla admitted in January 2026 that no Optimus was doing useful work.\n\nMoney follows the platform thesis. Skild and Physical Intelligence are priced like AI infrastructure, not robot makers. China dominates installations, and Unitree's August 2026 listing in Shanghai gave the camp its first large public pure play. Levine's wing is philosophically closer to the experience camp: robots should learn from their own deployment.", "disagreements": [ { "question": "Humanoid or task-specific?", "positions": [ { "who": "Figure, Tesla, 1X, Apptronik, Unitree, Galbot, AgiBot", "claim": "The world is built for human bodies; one form factor, mass-produced, wins." }, { "who": "Dyna, Covariant/Amazon, Agility (in practice)", "claim": "Deploy what works now: arms on wheels, warehouse-specific bodies. Reliability sells." } ], "sources": [ { "url": "https://www.figure.ai/news/series-c", "date": "2025-09-16", "title": "Figure exceeds $1B in Series C funding at $39B post-money valuation", "publisher": "Figure" }, { "url": "https://www.bloomberg.com/news/articles/2025-09-15/dyna-robotics-raises-120-million-in-funding-from-nvidia-amazon", "date": "2025-09-15", "title": "Dyna Robotics raises $120 million", "publisher": "Bloomberg" } ] }, { "question": "One brain for every body, or vertical integration?", "positions": [ { "who": "Skild, Physical Intelligence, Google DeepMind, Nvidia GR00T", "claim": "A cross-embodiment model licensed to hardware makers is the platform layer." }, { "who": "Tesla, Figure, Genesis AI (since 2026)", "claim": "The data flywheel only closes if you own the hardware and the deployment." } ], "sources": [ { "url": "https://techcrunch.com/2026/01/14/robotic-software-maker-skild-ai-hits-14b-valuation/", "date": "2026-01-14", "title": "Robotics software maker Skild AI hits $14B valuation", "publisher": "TechCrunch" }, { "url": "https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/", "date": "2026-07-15", "title": "Gemini Robotics 2 brings whole body intelligence to robots", "publisher": "Google DeepMind" }, { "url": "https://techcrunch.com/2026/05/06/khosla-backed-robotics-startup-genesis-ai-has-gone-full-stack-demo-shows/", "date": "2026-05-06", "title": "Khosla-backed robotics startup Genesis AI has gone full stack, demo shows", "publisher": "TechCrunch" } ] }, { "question": "Is this a camp at all?", "positions": [ { "who": "This atlas", "claim": "It imports architectures from camp 1 and simulators from camp 2; it is an application layer with its own data bet." }, { "who": "Levine, Physical Intelligence", "claim": "The data bet is the scientific question: robots have to learn from their own experience, which makes it a research program." } ], "sources": [ { "url": "https://www.physicalintelligence.company/blog/pi05", "date": "2025-04-22", "title": "π0.5: a VLA with open-world generalization", "publisher": "Physical Intelligence" }, { "url": "https://www.bloomberg.com/news/articles/2025-11-20/robotics-startup-physical-intelligence-valued-at-5-6-billion-in-new-funding", "date": "2025-11-20", "title": "Robotics startup Physical Intelligence valued at $5.6 billion", "publisher": "Bloomberg" } ] } ], "watch": [ { "text": "A general-purpose robot doing unseen household tasks at useful reliability without per-site data collection.", "sources": [ { "url": "https://www.physicalintelligence.company/blog/pi05", "date": "2025-04-22", "title": "π0.5: a VLA with open-world generalization", "publisher": "Physical Intelligence" } ] }, { "text": "Optimus production numbers from Fremont in the second half of 2026, against a 2025 target of 10,000 that was missed entirely.", "sources": [ { "url": "https://electrek.co/2026/04/22/tesla-optimus-production-fremont-model-sx-line/", "date": "2026-04-22", "title": "Tesla pushes Optimus V3 reveal later this year - again", "publisher": "Electrek" } ] }, { "text": "Unitree's and Agility's first public quarters: revenue per robot and installed base, the first audited figures in the camp.", "sources": [ { "url": "https://english.sse.com.cn/news/newsrelease/voice/c/c_20260806_10828128.shtml", "date": "2026-08-06", "title": "Unitree Robotics kicks off STAR Market IPO pricing", "publisher": "Shanghai Stock Exchange / Global Times" } ] } ], "exposure": [ { "lab": "unitree", "kind": "public", "note": "Unitree (SSE 688836) since August 2026." }, { "lab": "tesla", "kind": "public", "note": "Tesla (TSLA): Optimus, production starting 2026." }, { "lab": "google-deepmind", "kind": "public", "note": "Alphabet (GOOGL): Gemini Robotics, and an investor in Apptronik." }, { "lab": "nvidia", "kind": "public", "note": "Nvidia (NVDA): GR00T, Isaac, and on the cap table of Figure, Skild, PI, Dyna, Agility." }, { "lab": "amazon-agi-lab", "kind": "public", "note": "Amazon (AMZN): Covariant team, Agility customer, Dyna investor." }, { "lab": "boston-dynamics", "kind": "indirect", "note": "Hyundai Motor (005380.KS) owns Boston Dynamics and funds the RAI Institute." }, { "lab": "agility-robotics", "kind": "private", "note": "SPAC with Churchill Capital XI at $2.5B, closing 2026." }, { "lab": "figure", "kind": "private", "note": "$39B." }, { "lab": "skild-ai", "kind": "private", "note": "$14B; SoftBank-led." }, { "lab": "physical-intelligence", "kind": "private", "note": "$5.6B; CapitalG-led." }, { "lab": "apptronik", "kind": "private", "note": "$5.5B; Google and Mercedes." }, { "lab": "galbot", "kind": "private", "note": "$3B+; Hong Kong IPO planned." }, { "lab": "agibot", "kind": "private", "note": "Hong Kong IPO planned at ~$6B." } ], "reading_list": [ { "title": "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion", "url": "https://arxiv.org/abs/2303.04137", "date": "2023-03-07", "why": "The action head most current robot policies use." }, { "title": "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware", "url": "https://arxiv.org/abs/2304.13705", "date": "2023-04-23", "why": "ALOHA. Cheap teleoperation made imitation data collection possible at scale." }, { "title": "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control", "url": "https://arxiv.org/abs/2307.15818", "date": "2023-07-28", "why": "A language model that outputs robot actions; the VLA idea in one paper." }, { "title": "Open X-Embodiment", "url": "https://arxiv.org/abs/2310.08864", "date": "2023-10-13", "why": "Pooling data across 22 robots. The camp's answer to the data problem." }, { "title": "π0: A Vision-Language-Action Flow Model for General Robot Control", "url": "https://arxiv.org/abs/2410.24164", "date": "2024-10-31", "why": "Physical Intelligence's generalist policy; folding laundry from one model." }, { "title": "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots", "url": "https://arxiv.org/abs/2503.14734", "date": "2025-03-18", "why": "NVIDIA's open humanoid model and the synthetic-data pipeline behind it." } ] }, { "id": "alt", "name": "Program search and other substrates", "short": "Other substrates", "order": 6, "color": "#8A7A1B", "bet": "Gradient descent over one giant network is the wrong primitive. Intelligence is efficient skill acquisition on novel problems, which needs search, symbols, probabilistic models, or a different kind of computer.", "capital": { "text": "Small and scattered. Liquid AI at $2.35B and Sakana at $2.65B are the largest; most labs are under $100M, and two are nonprofits.", "sources": [ { "url": "https://techfundingnews.com/liquid-ai-closes-250m-hits-2b-valuation-with-amd-led-funding/", "date": "2024-12-13", "title": "Liquid AI closes $250M, hits $2B valuation", "publisher": "Tech Funding News" }, { "url": "https://www.prnewswire.com/news-releases/extropic-signs-75-million-letter-of-intent-with-us-department-of-commerce-to-scale-and-onshore-thermodynamic-computing-302839290.html", "date": "2026-07-30", "title": "Extropic signs $75 million letter of intent with U.S. Department of Commerce", "publisher": "PR Newswire" } ] }, "stances": { "signal": { "text": "Task success on novel abstraction (ARC), free-energy minimization, symbolic consistency", "values": [ "task-success", "free-energy" ] }, "when": { "text": "Often at test time: search and adapt per task rather than pretrain once", "values": [ "test-time" ] }, "representation": { "text": "Explicit programs, symbols or probabilistic models; deep nets only as a guide for search", "values": [ "programs", "probabilistic" ] }, "acts_in": { "text": "Mostly abstract problems today; some robotics via active inference", "values": [ "abstract", "physical" ] } }, "ideas": [ { "name": "Program synthesis", "sources": [ { "url": "https://arxiv.org/abs/2006.08381", "date": "2020-06-15", "title": "DreamCoder: Growing generalizable, interpretable knowledge with wake-sleep Bayesian program learning", "publisher": "arXiv" } ], "summary": "Search for a short program that explains the examples, rather than a network that approximates them. The program generalizes by construction." }, { "name": "Test-time adaptation", "sources": [ { "url": "https://arxiv.org/abs/2411.07279", "date": "2024-11-11", "title": "The Surprising Effectiveness of Test-Time Training for Abstract Reasoning", "publisher": "arXiv" } ], "summary": "Fine-tune or search on each new task at answer time. The ARC leaders all do it; the LLM camp took the idea and called it test-time compute." }, { "name": "Bayesian program learning", "sources": [ { "url": "https://doi.org/10.1126/science.aab3050", "date": "2015-12-11", "title": "Human-level concept learning through probabilistic program induction", "publisher": "Science" } ], "summary": "Represent a concept as a probabilistic program and learn it from one example. Lake, Salakhutdinov and Tenenbaum matched humans on handwritten characters this way in 2015." }, { "name": "Thousand Brains", "sources": [ { "url": "https://arxiv.org/abs/2412.18354", "date": "2024-12-24", "title": "The Thousand Brains Project: A New Paradigm for Sensorimotor Intelligence", "publisher": "arXiv" } ], "summary": "Hawkins's theory that the cortex is thousands of columns each learning a full model of objects through movement, and voting. Now an open-source project." }, { "name": "Active inference", "sources": [ { "url": "https://www.nature.com/articles/nrn2787", "date": "2010-01-13", "title": "The free-energy principle: a unified brain theory? (Friston)", "publisher": "Nature Reviews Neuroscience" } ], "summary": "Friston's framework: an agent acts to minimize surprise about its sensory inputs, which unifies perception and action under one objective." }, { "name": "Neuromorphic / analog compute", "sources": [ { "url": "https://arxiv.org/abs/2212.13345", "date": "2022-12-27", "title": "The Forward-Forward Algorithm (Hinton; mortal computation)", "publisher": "arXiv" } ], "summary": "Compute with physics instead of against it: thermodynamic sampling, analog weights that die with the chip (Hinton's mortal computation), living neurons." }, { "name": "ARC-AGI", "sources": [ { "url": "https://arxiv.org/abs/1911.01547", "date": "2019-11-05", "title": "On the Measure of Intelligence (Chollet)", "publisher": "arXiv" } ], "summary": "The Abstraction and Reasoning Corpus: grid puzzles solvable by most people from a couple of examples, built so that memorization does not help." } ], "status": { "text": "ARC-AGI gets harder faster than models solve it: 54% on ARC-AGI-2 by a search wrapper over Gemini 3, about 1% on the new interactive ARC-AGI-3. Funded but small; Basis and the Thousand Brains Project are nonprofits. The camp's best ideas keep leaking into camp 1 as test-time compute and tool use, which both validates and neutralises it.", "sources": [ { "url": "https://arcprize.org/blog/arc-prize-2025-results-analysis", "date": "2025-12-05", "title": "ARC Prize 2025 results and analysis", "publisher": "ARC Prize" }, { "url": "https://arcprize.org/blog/arc-prize-2026-milestone-1", "date": "2026-07-06", "title": "ARC Prize 2026: ARC-AGI-3 Milestone Prize #1", "publisher": "ARC Prize" }, { "url": "https://poetiq.ai/posts/arcagi_verified/", "date": "2025-12-05", "title": "Poetiq shatters ARC-AGI-2 state of the art at half the cost", "publisher": "Poetiq" } ] }, "proof": { "text": "Beating ARC-AGI at human sample efficiency, or learning a new domain from a handful of examples without a giant prior.", "sources": [] }, "sources": [], "explainer": "Every other camp trains one giant network by gradient descent. This one says that is the wrong primitive. Intelligence is efficient skill acquisition on problems you have not seen, and that needs search over programs, explicit symbols, probabilistic models, a different computer, or a different kind of network.\n\nARC-AGI is the camp's scoreboard. It is built to resist memorization, and it gets harder faster than models solve it: the best verified ARC-AGI-2 result in 2025 was 54%, achieved by wrapping a frontier model in a search loop, at $30 a task. ARC-AGI-3, launched in 2026, is interactive; the first milestone winner scored about 1%.\n\nThe camp's best ideas keep leaking into camp 1 as test-time compute and tool use, which both validates and neutralises it. Its hardware wing (Extropic, Normal, Cortical Labs) is the only part with no overlap: thermodynamic chips and living neurons are not something a transformer can absorb.", "disagreements": [ { "question": "Does ARC measure intelligence or just a gap in training data?", "positions": [ { "who": "Chollet, ARC Prize", "claim": "ARC tests skill acquisition on novel problems; frontier models fail it because they retrieve rather than reason." }, { "who": "LLM camp", "claim": "Scores keep rising with better models and search wrappers; Poetiq got 54% on ARC-AGI-2 on top of Gemini 3." } ], "sources": [ { "url": "https://arcprize.org/blog/arc-prize-2025-results-analysis", "date": "2025-12-05", "title": "ARC Prize 2025 results and analysis", "publisher": "ARC Prize" }, { "url": "https://poetiq.ai/posts/arcagi_verified/", "date": "2025-12-05", "title": "Poetiq shatters ARC-AGI-2 state of the art at half the cost", "publisher": "Poetiq" } ] }, { "question": "Search over programs, or a different substrate?", "positions": [ { "who": "Ndea, Poetiq, Tufa, Basis, Symbolica", "claim": "Use deep learning to guide a search over explicit programs; the program is the generalization." }, { "who": "Extropic, Normal Computing, Cortical Labs, Liquid AI", "claim": "The bottleneck is the computer: sample probabilities with physics, or compute with neurons, or drop the transformer." } ], "sources": [ { "url": "https://techcrunch.com/2025/01/15/ai-researcher-francois-chollet-founds-a-new-ai-lab-focused-on-agi/", "date": "2025-01-15", "title": "AI researcher François Chollet founds a new AI lab focused on AGI", "publisher": "TechCrunch" }, { "url": "https://www.prnewswire.com/news-releases/extropic-signs-75-million-letter-of-intent-with-us-department-of-commerce-to-scale-and-onshore-thermodynamic-computing-302839290.html", "date": "2026-07-30", "title": "Extropic signs $75 million letter of intent with U.S. Department of Commerce", "publisher": "PR Newswire" } ] } ], "watch": [ { "text": "Beating ARC-AGI at human sample efficiency, or learning a new domain from a handful of examples without a giant prior.", "sources": [ { "url": "https://arcprize.org/blog/arc-prize-2025-results-analysis", "date": "2025-12-05", "title": "ARC Prize 2025 results and analysis", "publisher": "ARC Prize" } ] }, { "text": "ARC-AGI-3 scores through 2026: the interactive benchmark started near 1%.", "sources": [ { "url": "https://arcprize.org/blog/arc-prize-2026-milestone-1", "date": "2026-07-06", "title": "ARC Prize 2026: ARC-AGI-3 Milestone Prize #1", "publisher": "ARC Prize" } ] }, { "text": "Whether Extropic's Z1 chip ships at scale under the CHIPS letter of intent, and whether any thermodynamic chip runs a real model.", "sources": [ { "url": "https://www.prnewswire.com/news-releases/extropic-signs-75-million-letter-of-intent-with-us-department-of-commerce-to-scale-and-onshore-thermodynamic-computing-302839290.html", "date": "2026-07-30", "title": "Extropic signs $75 million letter of intent with U.S. Department of Commerce", "publisher": "PR Newswire" } ] } ], "exposure": [ { "lab": "verses", "kind": "public", "note": "Verses (Cboe Canada VERS / OTC VRSSF): the only listed pure play." }, { "lab": "liquid-ai", "kind": "private", "note": "$2.35B; AMD-led." }, { "lab": "sakana-ai", "kind": "private", "note": "$2.65B; MUFG, Khosla." }, { "lab": "extropic", "kind": "private", "note": "$75M CHIPS letter of intent, July 2026." }, { "lab": "normal-computing", "kind": "private", "note": "$85M raised; Samsung Catalyst." }, { "lab": "ndea", "kind": "private", "note": "~$43M." }, { "lab": "poetiq", "kind": "private", "note": "$45.8M seed." } ], "reading_list": [ { "title": "On the Measure of Intelligence", "url": "https://arxiv.org/abs/1911.01547", "date": "2019-11-05", "why": "Chollet defines intelligence as skill-acquisition efficiency and introduces ARC." }, { "title": "The free-energy principle: a unified brain theory?", "url": "https://www.nature.com/articles/nrn2787", "date": "2010-01", "why": "Friston. The theory behind active inference and Verses." }, { "title": "Liquid Time-constant Networks", "url": "https://arxiv.org/abs/2006.04439", "date": "2020-06-08", "why": "Hasani and Lechner: the neuron model Liquid AI is built on." }, { "title": "A Thousand Brains", "url": "https://en.wikipedia.org/wiki/A_Thousand_Brains", "date": "2021-03", "why": "Hawkins: cortical columns, reference frames, and why deep learning is not how brains work." }, { "title": "The Forward-Forward Algorithm", "url": "https://arxiv.org/abs/2212.13345", "date": "2022-12-27", "why": "Hinton's local learning rule and the mortal-computation argument." }, { "title": "Mamba: Linear-Time Sequence Modeling with Selective State Spaces", "url": "https://arxiv.org/abs/2312.00752", "date": "2023-12-01", "why": "Gu and Dao. The state-space line that Cartesia commercialized." }, { "title": "xLSTM: Extended Long Short-Term Memory", "url": "https://arxiv.org/abs/2405.04517", "date": "2024-05-07", "why": "Hochreiter's answer to the transformer, the basis of NXAI." } ] } ]