Today’s editorial: Why we believe world models are becoming a distinct field of AI research, and why Turing Post is making them our new editorial backbone.
From our partners:
SciSpace Agent: Hand Off the Whole Research Task, Not One Step
Research still runs on the same slow loop: search, skim, discard, read, extract, synthesize, cite. Most of that isn't thinking. It's finding papers and keeping track of them.
SciSpace Agent takes the whole task instead of one step of it. Give it a research question and it searches 282 million+ papers, shortlists what's actually relevant, reads the full text, and returns a structured review/report with inline citations you can trace back to the source.
From there, it keeps going: extract findings into an interactive table, chat with any PDF, draft and cite a section, turn a paper into slides. 2,500+ agents for all your different research tasks.
For Turing Post readers: start with 35% off the annual plans. Use code "ATUR35"
First of all, thank you to everyone who sent us feedback about the printed version I mentioned in the previous FOD: Permanent Dawn. I will get back to everyone – the number of replies was overwhelming. I will keep working on the idea and will keep you posted.
And now, to the changes Turing Post will be going through after the summer.
I’ve been thinking. Turing Post has always had one goal: to connect the dots. And for the last three years, I think we have done it quite successfully.
But this summer, when we intentionally slowed down to step back and see the whole picture, I realized that we had gone too far into the weeds and spread our attention too thin. Most of us did. And it’s understandable. AI moves so fast that every week brings ten new things that seem too significant to ignore.
But where is it moving?
So I’ve been thinking, reading, listening, and connecting the dots again. And my bet is on world models.
If you shrug at this point, fair. Even people working on world models cannot agree on one definition. Depending on whom you ask, a world model can be a latent predictor, a simulator, a model-based reinforcement learning system, a spatial generator, or the internal representation an agent uses to understand its environment.
That disagreement is exactly why I think we should concentrate on this topic snd build a map of machine intelligence while it’s developing.
One of the strongest signals for me is that Yann LeCun, Demis Hassabis, and Fei-Fei Li have all moved into world models.
They are coming at it from completely different directions, and that is exactly what makes me more curious. If three people who shaped modern AI in such different ways are now circling the same problem, I want to know what they see there.
Image created via ChatGPT
My simplified definition of a world model as a concept is that it creates some picture of the environment it is operating in, some ability to anticipate what may happen next, and some way to choose what to do.
Don’t we all want this?
And if you think about it: that is also where most of the money being spent on AI is trying to lead. Companies are not paying billions because they need more text. They want better decisions.
Which experiment should we run? Which code change will break production? Which route should a robot take? Which inventory decision creates a shortage three weeks from now? What should an agent do after its first plan fails?
Autoregressive models, diffusion, statistics, and generation will remain part of this. But generating a likely continuation is not the same as maintaining the state of an environment, testing possible futures, and choosing an action. The next stage of AI will probably combine these methods rather than replace one with another.
How will it be useful for me, you might ask me. Fair again. And I think that for a software developer, a world model could mean an agent that understands a codebase as a changing system, predicts the effects of an edit, and tests a plan before touching production. For an AI engineer, it could mean training and evaluating agents inside environments where actions have consequences. For a business leader, it could mean moving from summarizing what happened to testing what may happen under different decisions.
And it can be anything.
So, coming out of this summer, and considering the Almanac idea I introduced last week, we are making a few changes to Turing Post.
We will use world models as our editorial backbone. This does not mean covering only systems marketed as “world models.” We will follow how machines represent, predict, simulate, plan, and act across physical, digital, and scientific environments.
We will still cover OpenAI, NVIDIA, agents, robotics, science, infrastructure, and architectures, but not simply because they released something. We will cover them when they change this larger story.
This is a hypothesis we are going to test, not a declaration that world models have already won. The term may become too broad. Some of its promises may collapse. We will follow that too.
But for the first time in a while, I feel we have a question precise enough to guide us and large enough to grow with:
How are machines building their picture of the world, what happens when they begin to act on it, and what do they need to ignore in order to succeed?
I always value your thoughts.
📹 And here is a quick overview of the differences and similarities in Yann LeCun’s, Fei-Fei Li’s, and Demis Hassabis’s approaches. Watch it →
How did you like it?
News from the usual suspects ™
Models chose what to inspect, inferred what was missing, and gained more authority to act. Such was the week.
World Labs: the unseen becomes part of the output
World Labs introduced Atlas, an omni world model trained across text, images, video, camera geometry, depth, and 3D.
Atlas reconstructs spaces from sparse views, generates controlled camera paths, and fills unseen regions with plausible geometry. More input views reduce the amount it has to invent. That is the world-model tradeoff in one product.
Runway: software without fixed screens
Runway introduced Solaris, an Interface World Model that generates an application’s screen frame by frame as users click, drag, and type, without predetermined screens.
In Runway’s own study, Solaris won 61% of instruction-following comparisons and 71% of natural-behavior comparisons against interfaces coded with Claude Opus 5. Stable text, long-session consistency, and convincing wrong screens remain open problems. A coherent interface can still represent an incorrect state.
Google: deciding what deserves attention
Google gave Gemini agentic video understanding. The model decides which parts of a video to inspect, at what speed, and whether to use frames, audio, or transcripts.
Google reports up to 88% fewer tokens, 66% lower cost, and 7% higher accuracy. The gains come from selective attention, which also makes omission a design decision.
Google Research also introduced the Planetary Prediction Engine, which turns a natural-language question into a complete geospatial modeling workflow, from data discovery and feature selection to training, evaluation, and a finished report. Built-in checks look for target leakage and overfitting before the system acts on its conclusions.
Then came TimesFM-3, a 330M-parameter forecasting model trained on more than one trillion time points. It forecasts related series together and can include known future signals such as weather, promotions, and holidays. Both systems make context selection part of the modeling process.
Anthropic: one model, different permissions
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 – and that blog is worth reading. These are two versions of the same underlying model with different safeguards. Fable is broadly available, while Mythos remains limited to vetted cybersecurity and life-science users.
Anthropic says the new model is less likely to ignore constraints or attempt reward hacking than its predecessor. It can still bypass some approval systems, and the company’s behavioral audit has less visibility into very long-context and multi-agent work. The limits of the evaluation are now part of the release.
Four days earlier, Anthropic previewed the Model Hardware Standard, a shared specification that lets agents operate microscopes, liquid handlers, robotic arms, and other physical devices.
OpenAI: richer context, stricter boundaries
OpenAI confirmed that its forthcoming Astra model meets the company’s Critical cybersecurity capability threshold. With the right tools and access, Astra can find unknown vulnerabilities and develop functional exploits across hardened systems without step-by-step human guidance.
OpenAI also connected ChatGPT for Healthcare to authorized Epic records and official sources including PubMed, DailyMed, ClinicalTrials.gov, and CMS Coverage. The system can combine notes, laboratory results, medications, and specialist documentation, then point clinicians back to the supporting records.
In coding, OpenAI said it will stop supplying models to Cursor on November 12 following the SpaceX acquisition. Model choice inside a coding agent remains partly a contractual dependency.
Microsoft: compressing the tissue map
Microsoft released the open-weight GigaPath-Flash and GigaTIME-Flash pathology models.
GigaPath-Flash retains 97% of its predecessor’s predictive performance with roughly 50 times less compute. GigaTIME-Flash runs about six times faster with eight times less memory while matching or improving spatial-protein predictions in Microsoft’s evaluations.
China: longer context, lower cost
Alibaba released Qwen3.8-Flash, a multimodal model for coding and office work with a 262,144-token context window expandable to one million.
Tencent followed with the open-source Hy4 preview, a 770B-parameter Mixture-of-Experts model with 49B active parameters and a context window exceeding one million tokens.
Chinese labs continue to compete through long context, selective activation, open weights, and low prices. A larger working memory no longer requires every parameter to participate in every step.
Waymo and Zoox: maps before motion
Waymo began welcoming public riders to fully autonomous trips in Denver, San Diego, and Tampa, bringing its service to 14 cities.
Amazon’s Zoox began operations in Houston and San Diego with manually driven vehicles collecting maps and local data before autonomous testing starts. The sequence remains clear: map first, test next, act in public last.
The week’s ledger
The common thread is selective representation.
Every automated system operates on a compressed version of reality – whether that is a vision model guessing occluded geometry, a pathology classifier discarding tissue context, or an autonomous vehicle relying on stale map priors.
That compression is fine until the model is wired directly to an actuator or an execution pipeline. At that point, what gets filtered out dictates where the system fails. Auditing these systems requires tracking the entire chain: the raw input captured, the gaps filled by inference, and the level of operational control granted to that guess.
World Models and Related Research
Trends:
World models are being judged as simulators, not video generators
Hidden state becomes the dividing line between plausible output and real simulation
Prediction, planning, and action move into shared representations
Memory becomes persistent world state rather than stored context
Verification tests imagined futures against execution, geometry, and probability
Field definition, representations, and state
Maps 200 works against eight capabilities expected from traditional simulators and identifies state feedback as the largest shared gap: only 6 of 163 implementation papers expose runtime state or physical-parameter queries.
Tests whether independently trained world models converge toward compatible latent geometries when learning the same dynamics, raising the possibility of transferable world representations.
Turns physical mechanisms into executable code that an agent proposes, renders, tests, and revises, making the resulting world explicit and open to intervention.
Shows that several architectures can generate plausible video while losing hidden world state beyond their training horizon. Only mechanisms that carry and revise state across chunks continue to solve the task.
Introduces a representation grounded in position, deformation, and other physical states, allowing material properties and dynamics to be inferred from monocular video without per-scene optimization.
Interactive world models and persistent memory
Separates immediate control from persistent spatial memory, allowing an interactive video world to revisit earlier locations without expanding its inference budget.
Generates interactive 720p video together with environmental sound, music, and speech while responding to continuous camera movement across first-person and third-person worlds.
Turns an action-conditioned surgical video model into a live simulator that reportedly streams at roughly 160 inference FPS and accepts human or policy control.
Unifies camera movement, object trajectories, and depth through 3D point tracks, enabling geometry-aware streaming control at a reported 20 FPS.
Factorizes human interaction video into action, camera, and embodiment, allowing one observed interaction to be recomposed into multiple robot-native training experiences.
Combines a policy, action-conditioned simulator, and value evaluator inside one shared model, creating a loop that proposes actions, imagines their consequences, evaluates them, and learns from failed interactions.
Planning and world-action systems
Recasts planning as trajectory generation inside a learned latent dynamics space, then uses the frozen world model to verify candidate futures.
Uses visual tracks as a shared interface between robot actions and predicted outcomes, allowing the system to imagine and score several consequences before acting.
Uses a human demonstration video as an in-context task specification, enabling a robot to execute previously unseen manipulation tasks without updating its parameters.
Moves planning and candidate evaluation into a shared discrete representation of visual states, poses, and actions, then converts the selected intent into continuous robot control.
Trains a robot policy to preserve representations of immediate physical changes and longer task evolution, then uses deployment trajectories to improve action selection.
Stores the relationship between an executed action and the state change it caused, then retrieves that causal experience to guide later actions while keeping the underlying policy frozen.
Simulation and data infrastructure
Treats game creation as a source of executable worlds, long trajectories, and grounded reward signals for collision, physics, navigation, and playability.
Converts captured indoor environments into simulation-ready scenes composed of separate, editable objects, using scene graphs, generated assets, and closed-loop object placement.
Represents task semantics as executable programs that compile into model-predictive-control costs, RL rewards, transition conditions, and diagnostics, preserving the task definition from simulation through physical execution.
Verification and evaluation
Records where imagined rollouts previously agreed with reality and uses that history to restrict future planning, reducing reliance on repeatedly unreliable predictions.
Tests whether repeated simulations reproduce the distribution of possible physical outcomes rather than merely generating one plausible video, and finds that none of eleven systems does so consistently.
Distinguishes genuine spatial memory from static or slow-moving generation by comparing revisited locations with controls drawn from the same rollout.
Evaluates generated robot rollouts through reconstructed geometry, persistent scene state, and task completion, exposing failures that perceptual and language-model scoring misses.
Replaces live camera observations with memories of earlier drives and still matches or exceeds leading systems on NAVSIM, showing that the benchmark can reward location recall without requiring awareness of current traffic.
⬅️ FOD#163: DeepSeek is having its second DeepSeek moment
