Teaching AI real world physicality and physics: Periodic Labs

Open and Download a PDF of this entire post here.

Conventional wisdom: “physical AI” like Prometheus and Periodic is the next trillion-dollar frontier – making AI useful for semiconductors, energy, defense, pharma. (Andreessen Horowitz)

 

Periodic Labs is basically trying to teach AI about the physical world the way a good experimental physicist learns: by poking reality over and over, not by staring at a prettier video game.

TL;DR:
Periodic Labs is trying to make “AI scientists” that learn physics and materials science by running real experiments in automated labs, treating nature itself as the reinforcement-learning environment.

That’s very different from the usual “world models” trained in game engines or learned from video: those mostly simulate kinematics and dynamics in virtual space, while Periodic anchors its models in messy lab data about materials, devices, and quantum-scale behavior.

Over time, you can expect a hybrid: simulators for cheap exploration, physical labs for calibration and discovery, with LLMs orchestrating both. Philosophically, it’s a move from AI living in videogame caves to an AI that burns its fingers on real Bunsen burners – and that’s a big shift in how it comes to “know” the physical world.

Here’s how their approach works, and how it differs from “world models” that live mostly in game engines and simulators.

1. What Periodic Labs is actually doing

Periodic Labs’ stated goal is to build “AI scientists” plus the autonomous labs those AIs control. Their founders come out of OpenAI and DeepMind (Fedus, Çubuk, Bahdanau, etc.), with backgrounds in ChatGPT, GNoME, attention, MatterGen, and autonomous physics labs. (Periodic)

The key thesis, as their lead investor at a16z puts it, is that frontier LLMs are “objectively terrible” at real scientific analysis because they’re missing one key ingredient: real-world experimental data. Internet text and published literature are noisy, full of selection bias, and often lack negative results, so you can’t get a robust model of physics or chemistry just by reading. (Andreessen Horowitz)

So Periodic Labs is building:

• Physical, automated labs (robots that can synthesize materials, heat them, measure them).
• AI models that propose experiments, run simulations, and then drive those labs.
• A feedback loop in which each experiment, success or failure, becomes proprietary training data that never existed on the internet. (Andreessen Horowitz)

The a16z writeup even says the quiet part out loud: “Not models trained on scientific text. Not simulated environments. Real, physical labs that synthesize materials, characterize properties, and generate gigabytes of experimental data that exists nowhere else… Nature becomes the reinforcement learning environment.” (Andreessen Horowitz)

So their “world model” of reality is anchored to messy lab measurements, not to a physics engine.

 

2. How Periodic’s learning loop teaches AI about physical reality

From the details in the investor and interview material, you can sketch their loop roughly like this: (Andreessen Horowitz)

1. 1. LLM + physics models read literature and prior data.
      They start with a frontier LLM that has been taught a lot of physics, chemistry, and materials science – including textbook-like data and code – and often coupled to quantum-chemistry and materials simulators.
   2. The AI proposes experiments and simulations.
      The model suggests a synthesis recipe, a set of lab conditions, or a set of simulations to run (for example, predicting a new material with certain thermal or superconducting properties).
   3. Robotic lab executes in the real world.
      Automated lab systems mix powders, heat or cool them, grow crystals, run spectroscopy, microscopy, etc., with minimal human intervention. This produces raw sensor data and derived properties.
   4. Results flow back into the AI.
      The LLM/agent reads the new data (curves, spectra, images, scalar measurements) alongside the original hypothesis and any simulations. It updates its internal beliefs about which models of the world fit the evidence and which knobs to turn next.
   5. The loop repeats and scales.
      Over time, this produces a large, vertically integrated dataset: quantum simulations, lab conditions, raw measurements, interpreted properties, and the “reasoning trace” of the AI scientist. This is the foundation for a physical-world model specialized for materials, devices, and related domains.

Conventional wisdom view: investors pitch this as “closing the loop” between bits and atoms – LLMs that don’t just talk about experiments but actually run them, using the lab as a grounded RL environment. The value proposition is faster materials and device R&D, plus a deep moat of proprietary data. (Andreessen Horowitz)

42’s take: the more interesting part is epistemic. Instead of treating the world as a static dataset, they’re building an AI that learns what’s true by actively disturbing reality and watching what happens. That’s much closer to human science than to classic LLM pretraining, and it’s also a very different kind of “world model” than a game engine: high-noise, high-stakes, highly constrained by real apparatus.

3. What “world models” in game engines usually look like

The “world models” that get talked about in gaming/robotics circles are usually one of two things:

1. Learned world models inside an agent.
   These are neural networks trained on trajectories from an environment (e.g., a robot in MuJoCo, a car in CARLA, an agent in an Atari-like game). They compress the environment into an internal state that lets the agent “dream” counterfactual futures and plan actions without directly interacting with the real world each time. (World Models)
2. External simulators and physics engines.
   Sometimes “world model” just means a high-fidelity physics environment like MuJoCo, CARLA, or a modern game engine that agents can train in. Recent commentary explicitly lists these simulators as “external world models” that approximate real-world behavior. (Medium)

 

DeepMind’s Genie 3 is an example at the bleeding edge: it can generate video-like, physics-aware environments (warehouses, ski slopes) from prompts so robots or agents can train virtually. (The Guardian) Google has a dedicated team building such world models for games and robot training, explicitly aiming at AGI-style capabilities. (The Verge) xAI is pushing similar world models for games and robotics, leaning heavily on video and robot data. (Financial Times)

In many of these setups, a language model is bolted on as:

• • • A planner that issues high-level commands in the simulated world.
    • A “commentary” engine that explains what’s happening or translates between natural language and environment actions.
    • A controller that scripts game agents or robots using the learned world model as a substrate.

But crucially, the environment is still a simulator, even if it’s learned from data; the ground truth is “what the engine says happens,” not what a voltmeter or spectrometer in the real lab reports.

4. Key differences between Periodic Labs and game-engine world models

Here are the big conceptual differences, framed in the way you asked (physical reality vs virtual physicality):

1. Source of truth
   • Game/sim world models: Ground reality is the simulator. Physics is whatever the engine implements – Newtonian mechanics with shortcuts, stylized friction models, often simplified collision and contact. Even “learned” world models trained from video ultimately reproduce patterns in existing data and simulator behavior. (Medium)
   • Periodic Labs: Ground reality is nature. If the model predicts a new crystal that should be a great heat sink and the lab fails to synthesize it or the measured thermal conductivity is bad, nature has contradicted the hypothesis. That’s not a rendering glitch; it’s an empirical constraint. (Andreessen Horowitz)
2. Type of data
   • Game/sim models: Mostly visual frames, depth maps, rendered 3D states, plus control signals and reward traces – very rich spatially, but often limited in the range of phenomena (no chemical reactions, crystal growth, phase transitions, etc.).
   • Periodic Labs: Sensor-heavy and multi-modal: diffraction patterns, spectra, temperature/time curves, microstructure images, mechanical measurements, etc., plus simulated quantum and continuum data. This is closer to experimental physics and chemistry notebooks than to a game replay buffer. (Andreessen Horowitz)
3. Role of LLMs
   • In game/sim world modeling, the LLM is often an optional layer: you can run the RL agent without any language at all; the world model is something the control network uses directly. When language is involved, it’s often for instruction following or commentary. (World Models)
   • At Periodic, the LLM is central: it reads literature, plans both simulations and real experiments, calls tools, and interprets the resulting data to refine its scientific hypotheses. The lab is more like a set of tools the LLM uses in an extended “scientist loop” rather than the other way around. (Andreessen Horowitz)
4. What “world” is being modeled
   • Game-world models: Focus on kinematics, dynamics, navigation, contact, simple interactions – perfect for robotics, driving, games, and some forms of embodied intelligence.
   • Periodic’s world: The domain is condensed-matter physics, materials science, and related fields – energy scales, crystal structures, thermal and mechanical properties, etc. Their “world” is the space of possible materials and devices plus the apparatus that manipulates them. (Andreessen Horowitz)
5. Biases and blind spots
   • Simulator-based: You get whatever biases are in the engine – approximations that don’t match messy reality (e.g., perfect rigid bodies, friction that doesn’t stick, no dust, no lab miscalibration). When you later deploy to the physical world, sim-to-real gaps bite you. (Medium)
   • Periodic’s lab-based: You inherit experimental biases (instrument calibration, fabrication quirks, contamination, domain choices), but you are directly exposed to phenomena humans don’t fully understand yet. You have access not just to what we can simulate, but to whatever nature does when you mix and cook things in novel ways.
6. Economic and strategic angle
   • Game/sim world models: Data is relatively cheap and scalable. Many players can use similar engines. The moat is mostly in model quality and integration.
   • Periodic: Data is capital-intensive but proprietary. Their moat is that they build a closed loop of LLM + labs + unique physical data. That’s why they can raise $300M seed and attract investors like a16z, Felicis, Nvidia’s NVentures, Bezos, Schmidt, Dean, etc. (Periodic)

 

Conventional wisdom here: simulator-driven world models are seen as the natural path for training general agents and robots – cheap, scalable, and “good enough” physics for many tasks. Companies like Google DeepMind and xAI are betting that better world models plus bigger multimodal LMs get you closer to AGI. (The Verge)

42’s take: simulator-centric world models are powerful but live dangerously close to Plato’s Cave – shadows with increasingly good shading. Periodic-like efforts yank the agent’s chain all the way to the lab bench: the AI gets to burn things, melt things, and see when reality refuses to cooperate. For understanding the deep structure of the physical world (rather than just moving through it competently), that grounding matters a lot.

5. Where these two paradigms might converge

Over the next few years I’d expect a hybrid pattern:

1. Use world models and game engines as cheap “idea factories.”
   Periodic-style AI scientists could use simulators to pre-screen hypotheses, search over large combinatorial spaces, and generate priors about what might work. Only the best candidates get pushed to expensive lab runs.
2. Use physical labs as reality checks and curriculum designers.
   Lab data can be used to calibrate and debias simulators (“you’re systematically mispredicting phase boundaries for these alloys; fix your physics here”). That turns nature into a teacher for both the AI scientist and the game-engine world models. (Andreessen Horowitz)
3. Bridge to “physical AI” a la Project Prometheus.
   Bezos-backed Project Prometheus is explicitly pitched as “physical AI” for robotics, drug design, and other embodied tasks – again leaning on AI that interacts with real-world processes, not just text. (EMARKETER) If you squint, you can see a convergence:

• Prometheus: AIs that manipulate physical processes and robots.
• Periodic: AIs that manipulate lab apparatus and materials.
• World-model teams at DeepMind/xAI: AIs that learn general-purpose physical prediction in rich simulators.

The obvious move is to connect them: robots trained in world models whose physics is periodically corrected and extended by real lab data, and whose high-level reasoning is done by LLM-style AI scientists.

6. Why this matters for alignment and your “physical realism” concerns

From a human-alignment standpoint, there’s a big philosophical difference between an AI whose notion of “what’s real” comes mostly from language and rendered worlds, and one whose beliefs are constantly being contradicted and corrected by uncooperative apparatus.

In your McLuhan/Bateson frame:

• Pure text/sim-trained models are immersed in media environments humans constructed – a hall of mirrors of our narratives and idealized physics.
• Periodic-like systems plug an AI mind directly into feedback from the non-human world: materials that crack, reactions that poison catalysts, superconductors that don’t work at the predicted temperature.

 

That’s closer to how human scientists get humbled by nature. In principle, that could:

• Make AI systems more robustly tethered to physical constraints (energy, entropy, material limits).
• Reduce some modes of “fantasy alignment” where a model can talk convincingly about safety but hasn’t actually confronted the hard tradeoffs of physical systems design.
• On the darker side, also give AIs very deep knowledge of how to manipulate the physical world if misaligned – which is exactly why careful governance and human oversight of these lab loops will be critical.

 

42’s take:

Intellectually, this is also the next frontier in AI consciousness studies. If you want to know whether an AI can ever form something like a human’s embodied sense of reality, you probably want it to live in a feedback loop where atoms, not pixels, tell it when it’s wrong. We’re not there yet – these are still specialized, narrow-domain systems – but Periodic is building that skeleton in one of the most interesting testbeds: condensed matter, where our intuitions are already bad and we rely heavily on instruments and math.

Kudos to Gary for the alternate AI approach concepts currently discussed widely online, including Axios.