Ice cream sales and drowning accidents rise at the same time. Does eating ice cream cause people to drown? Of course not. Warmer weather drives both: people buy more ice cream, and more people go swimming. Temperature is the confounder. The same trap appears in model evaluation. “A new prompt improved performance by 15%.” But did the prompt actually improve, or did the test set just get easier? Real-world evaluations involve countless, often unobserved variables, making perfectly controlled experiments difficult or impossible. A change in query difficulty, distribution, or context can easily look like a model improvement. This is why causal evaluation matters. Instead of asking what changed together, it asks what the intervention actually caused. By accounting for confounders, we can distinguish genuine improvements from changes in the evaluation environment and make better downstream decisions. How does your team account for confounders across evaluation runs? #CausalAI# #CausalWorldModels#

“What if I do this action? What will happen?” That’s the core of counterfactual reasoning in Physical AI. Our founder @huang_biwei joined @CharlieFink and @VirtualTedS on @AIXRPodcast to talk 10+ years of causal AI research, how it’s shaping our causal world models, and why Physical AI is our first proving ground. Check out the full episode in the comments.

What makes a world model causal? As a robotic gripper closes around an object, the contact conditions change. Friction, grasp point, speed and angle can all affect the outcome. Our framework treats causality as an explicit system of causal variables, causal structure and causal dynamics in latent space. Which condition would you vary first? #CausalWorldModels#

Judea Pearl (@yudapearl) is a Turing Award winner famous for his contributions to artificial intelligence and causal reasoning. We talked about the greatest scientist of all time, his major breakthroughs and his predictions for AI today. Full conversation also on my YouTube, Spotify, and Apple Podcasts. Timestamps: 00:00 Intro 00:54 How he got into AI 11:17 Greatest scientist of all time 20:15 What people thought of AI in the 80s 26:23 Entering academia and researching AI 34:52 The invention of Bayesian networks 46:28 Pioneering work in causality 55:38 The causal hierarchy 59:34 LLMs and predictions 01:20:12 A restless mind pays 01:24:36 Advice for his younger self 01:26:37 Outro

Aether AI is building causal world models for real-world intelligence. We’re hiring across research, engineering, and robotics. Join a small team working across models, systems, simulation, and real-world machines. Explore open roles:

Machines’ understanding of humans is not just a perception problem. For robots, it becomes a question of action: what people are trying to do, how intention shows up in movement, and what changes in the physical world after action is taken. This is why causal world models matter for physical AI. Our founder @huang_biwei will join Forum 1: Machine Understanding & Human–Robot Coexistence at the Humanity & AGI Summit 2026 on July 12 at Stanford. Aether AI will also have a booth at the summit. Stop by if you’ll be there. #CausalAI# #EmbodiedAI# #Robotics# #WorldModels#

🎐 @saturdayrobotic Robotics & World Models Reading Club #13#, @huang_biwei (Assistant Professor at @UCSanDiego & Founder of @AetherLab_AI) delivered a deep technical vision for what a true Causal World Model (CWM) should be—and why today's frontier AI systems still fall short. Current paradigms: 🔹 LLMs (GPT, Claude, Gemini, π): language correlation models with weak grounding in physical interventions. 🔹 Video generators (Sora, Wan, Veo 3): high-fidelity pixel prediction but limited mechanism modeling, counterfactual reasoning, or causal guarantees. 🔹 JEPA-style models (V-JEPA, AMI): promising representation learning, but not yet fully causal world models. Key diagnosis: ❌ Modeling correlations ✅ Modeling how latent causes generate observations and how the world evolves under actions Causal World Model learns: 📥 Inputs: • Observations • Interventions / Actions • Temporal Information • Context / Environment • Other Agents / Entities ⚙️ Core Components: 1️⃣ Causal Feature Representation Learning Encoder → latent causal variables Z, recovering disentangled causal factors rather than merely compressing observations. 2️⃣ Causal Structure Learning Learn causal graph G: • Directed arrows → causal effects • Dashed links → latent confounders • Explicit latent causal variables and hidden drivers 3️⃣ Causal Dynamic Learning Temporal causal model over t → t+1 → t+2 with intra-time and inter-time causal dependencies. Benefits: ✅ Generalization under distribution shift ✅ Precise forecasting & action control ✅ Data-efficient and model-size-efficient learning ✅ Intervention & counterfactual reasoning ✅ Discovery of hidden mechanisms and new knowledge A recurring principle across tabular data, questionnaires, images, videos, time series, and genomics: 👉 Representation learning should recover latent factors and causal relations among them—not just compress observations. Examples: 📊 Tabular/questionnaire data → latent DAGs with causal variables and latent factors U1–U5 (NeurIPS 2022/2023, ICLR 2024, ICML 2024). 🎥 Images/videos → compositional concept graphs ("chicken eating rice" + "peacock" → "peacock eating rice"), hierarchical concepts, temporal pose graphs z₁,₁…z₃,₆ (CVPR 2026, ICLR 2025/2022, NeurIPS 2024/2022). 📈 Time series → trend, seasonal, cyclical, noise decomposition. 🧬 Genomics → motifs, binding sites, regulatory features. Strong critique of current video world models: ❌ Precise Action Control (WorldGym) ❌ Object Consistency (UniSim) ❌ Physical Consistency (CogVideoX) Beyond scaling, world models need explicit causal structure. Mathematically, the framework jointly models: • Observation function oₜ • Reward function rₜ • Hierarchical latent states sᵢ,ₜ • State dynamics • Action variables aₜ • Structural relation masks Dₛ(·,i) Empirical results: 🏃 Walker2d-v2 & Humanoid-v2: Causal Dynamic Learning (CDL) consistently outperforms Curiosity, CID, and ASR in: • Lower next-state prediction error • Higher policy returns For distribution shifts, the framework introduces: 🔹 Domain factors c 🔹 Structural adaptation parameters θₖ Key insights: • Graphs compactly encode what changed and where • Not all state dimensions matter for action prediction • Not all θ need adaptation • Factors can be adapted independently 🎮 CoinRun experiments further show that new latent variables and structural relations (adding new state dimensions such as s₄) can be introduced while preserving causal factorization, enabling efficient adaptation to new environments, mechanics, and visual styles. World Models = Causal Feature Learning + Causal Graph Discovery + Causal Dynamics + Factorized Distribution Shift Modeling Only then can AI achieve: 🔮 Counterfactual simulation 🎯 Intervention reasoning ⚡ Data-efficient adaptation 🔬 Discovery of previously unknown mechanisms This is not simply scaling larger models—it is a proposal to move from predicting what happens next toward understanding why the world changes.

Aether AI has raised $20M in seed funding, backed by a group of leading global investors with deep expertise in artificial intelligence and frontier technologies. Founded by Prof. Biwei Huang（@huang_biwei）. We're building Causal World Models for Real World Intelligence. 🔗

MLSS ( is back in Tübingen. It is hard to believe how much the field has changed since Alex Smola and I did the first MLSS, and since the first Tübingen edition in 2003. Our last in-person edition here was in 2017, so this feels long overdue (1/3)

Why Hasn't Embodied AI Had Its GPT Moment? Last night on #GeekParkLive#, Jack Zhang sat down with Biwei Huang @huang_biwei, founder of @AetherLab_AI. The question on the table: can next-token prediction ever really understand why things happen? And does that matter for physical AI? A few things from the conversation worth sitting with:

Today is the day. Tonight, 9PM Beijing time(06:00 AM PDT), Prof. @huang_biwei goes live with Zhang Peng, Founder and President of @thegeekpark. Topic: Why causal world model is the next AI paradigm? LIVE conversation. No scripts. Straightforward take on where AI is heading. See you in the livestream! #CausalAI# #AetherAI# #BiweiHuang# #GeekPark#