AI Just Moved Into the Lab: Storm-Scale Weather, a Chemist Model and a Bug Hunter That Found 81 Firefox Flaws

Forget chatbots. In one week, AI learned to simulate the whole planet's storms at 5 km, plan chemistry experiments, steer real robots after training only in simulation, and exploit 948 Linux bugs. Here is what actually shipped, what is hype, and why it matters.

The AI Spotlight DeskNov 8, 2026 · 13 min read

The stake
Simulating the whole planet's weather
Hydrogen-1.0: a model trained on lab notebooks
Robots that learned entirely in simulation
The robotics toolbox goes open source
Reading human motion from ordinary video
Photos become editable 3D worlds
Tiny models, big reach
GPT-6 Cyber and the 948 exploits
The big open models behind it all
Why this matters
Practical playbook
Risks and counterpoints
FAQ
The bottom line

The stake

Most AI headlines are about chat. Which model writes better, which one codes faster, which one costs less.

This week the more important story happened somewhere else. AI moved into the places where mistakes are measured in rainfall, molecules, robot arms and security holes.

In a single week:

NVIDIA published an AI model that emulates the entire planet's atmosphere at roughly 5 km resolution, in 10-minute steps.
A materials-science company released Hydrogen-1.0, an open 753-billion-parameter model trained on real laboratory records, including failed experiments.
Researchers showed a robot policy trained only in simulation handling real tasks it had never seen.
OpenAI said its new security model developed working exploits for 948 Linux kernel vulnerabilities and found 81 previously unknown high-severity Firefox flaws.

Add a wave of open 3D-scene generators, motion-capture-from-video tools and a vision model small enough to fit in 22 megabytes, and a pattern appears. AI is learning to model the physical world, not just describe it.

Here is the full rundown, with the caveats the launch posts tend to leave out.

Simulating the whole planet's weather

NVIDIA's STRATA (Storm-resolving Tile-based autoRegressive Atmosphere Transformer Architecture) is the most ambitious release of the week.

It takes the current state of the atmosphere — temperature, wind, humidity, pressure and more — and predicts how it changes over the next 10 minutes. Then it feeds that prediction back into itself, step after step, to roll forward hours or days.

The detail is what makes it unusual:

About 4.9 km grid spacing across the whole globe, roughly 25 million columns.
24 vertical levels in the model's prognostic state, plus near-surface temperature, with precipitation and surface pressure diagnosed each step.
About 332 million parameters — not even a billion — with a checkpoint around 1.6 GB.

Most global AI weather models work at 25 to 100 km. At 5 km you start to see individual storm systems instead of smeared averages.

How it works

STRATA does not process the planet in one giant pass. It cuts the globe into overlapping tiles, processes each tile, and blends them back together. The researchers' bet is that over 10 minutes, the atmosphere behaves mostly locally, so training on many small tiles makes up for having only a few weeks of global data.

It was trained to emulate SCREAM, a physics-based storm-resolving model that normally needs exascale supercomputers. The paper reports:

Training on just 17 days of high-resolution simulation output.
Stable 24-hour global rollouts with realistic storm-scale behavior.
About 48 simulated days per megawatt-hour, around 50 times more energy-efficient than the physics model.

The fine print

The authors are candid that large-scale biases develop as forecasts run longer. The code repository notes a known issue with ice precipitation. It is a research emulator, not an operational weather service, and it emulates a simulation rather than being validated against real observations.

On licensing: the code is Apache 2.0, while the checkpoint is published under the Linux Foundation's OpenMDW license, which NVIDIA says permits commercial and non-commercial use. Training recipes and datasets are public.

Hydrogen-1.0: a model trained on lab notebooks

The second big release came from Deep Principle, a Hangzhou-based AI-for-materials company. Hydrogen-1.0 is built by post-training Z.ai's open GLM-5.3 on chemistry, materials science and life-science data.

What makes it different is the training data:

Domain knowledge from the scientific literature.
Electronic lab notebook records from the company's own labs.
Discovery trajectories: the full chain of hypothesis, action, observation and revision — including failed attempts.

That last point matters. Most models learn from published papers, which overwhelmingly report successes. Science is mostly failure. Training on what did not work is how you learn which experiment to run next.

What it can do

The release showed four demos:

Molecular simulation. It wrote Python to build a polyethylene melt, ran it in the GROMACS simulator, caught a topology error, fixed it, restarted, and analyzed a 40-nanosecond trajectory.
Reaction networks. It grew a network of 27 structures through 30 plausible reactions.
Spectrum reading. It ranked seven candidate products from a proton NMR trace.
Lab robot protocols. It wrote a machine-readable protocol for a 12-vial chemistry study, and a virtual lab check caught an error (an aspiration without a fresh pipette tip) before anything ran on real equipment.

The numbers and the hardware

Deep Principle reports that Hydrogen-1.0 beat OpenAI's GPT-6 Astra on the FrontierScience-Research benchmark (38.3% vs 33.3% in its own tests) and roughly matched Kimi-K3 on GPQA-Diamond. On NMR structure work it still trailed Astra. These are self-reported results.

It is a mixture-of-experts model with 753 billion total parameters, a 1-million-token context and up to 128K tokens of output. Weights are on Hugging Face under the MIT license, with full BF16 weights plus NVFP4 (for NVIDIA Blackwell) and MXFP4 (for AMD Instinct) versions. Even the compressed versions run to hundreds of gigabytes, so this is multi-GPU, server-class territory.

The model card's own warning is the right one: a detailed, confident answer does not make it correct, and outputs should not be treated as verified experimental findings.

Robots that learned entirely in simulation

Robotics had its own milestone. A new policy model called ZAP (short for "zero real agent policy") takes camera footage plus a plain-language instruction and directly controls a robot arm.

The headline claim: it was trained entirely in simulation, with no real-world robot data.

The recipe, as described by its authors:

Start with a vision-language model that already understands objects and instructions.
Teach it to move a robot with reinforcement learning in a virtual environment.
Train on just 40 fairly simple simulated tasks.

On real hardware, it reportedly handled tasks it had never seen: moving a test tube, moving a chess piece according to a diagram, and holding a pen to draw an X on a tic-tac-toe grid. At 27 billion parameters, its authors say it matched the real-world success rate of the much larger GPT-6 Astra while being faster and cheaper.

So far only a technical paper is out. There is no word on releasing the model. Treat the comparison as a claim until others reproduce it.

Why it matters: real-robot data is slow and expensive to collect. If simulation-only training transfers to the real world, robotics gets a lot cheaper to scale.

The robotics toolbox goes open source

Alongside that, an open-source suite called Open Robo Kit bundled four libraries:

A motion and planning kit for calculating movements — demos showed robot hands spinning pens, rotating cubes and opening notebooks, plus robots climbing obstacles and jumping.
A retargeting kit that maps human motion from video, motion capture or VR onto robot hands or full humanoids.
A grasping kit for picking up and manipulating objects, training policies and deploying them in simulation or on real robots.
Robot skills for AI agents, so an agent can generate robot motion, rebuild simulations from video, or improve its own control by running experiments.

All four have code released.

Reading human motion from ordinary video

Several releases attacked the same bottleneck: getting accurate 3D human motion without a motion-capture studio.

FlowHMR reconstructs full-body 3D movement from a regular video — kicking a ball, doing push-ups — and adds physics constraints so the motion is physically plausible rather than approximate. Its authors show the motion transferred to a Unitree G1 humanoid in simulation, used for character animation, and paired with video generators to turn text into 3D movement. Code, models and a browser demo are out.
ReLift from NVIDIA tracks individual points through a video in 3D, even when they disappear behind other objects. It first uses existing models to track points and estimate depth, then a second model corrects errors and predicts hidden motion. Code, weights and training scripts are released.
DreamX extracts motion from a video and applies it to a different subject — for example turning a human hand manipulating objects into a robot doing the same, and translating that into real robot actions. Code is listed as coming soon.
Mobius generates 3D human motion from text, and also does the reverse: it describes a motion sequence in words, completes partial motions and edits animations from instructions. Code is under internal review.

Put together, these tools turn ordinary video into training data for animation and robotics. That is a big deal for anyone who cannot afford a mocap stage.

Photos become editable 3D worlds

The week also brought a cluster of 3D generators:

Lucid Lera turns one image into an explorable 3D scene and recognizes individual objects, so you can select, move, resize, duplicate or delete them. It uses a video model to imagine other camera angles, then fuses them into 3D. Scenes can become robot training environments. Code coming soon.
Cola 3D turns one image into a scene where each object is its own mesh, processing the scene all at once to keep positions and sizes consistent. Modes include fully automatic, point prompts or masks. Code, training code and dataset are released.
VoxLusion generates furnished indoor rooms — apartments, offices, bedrooms — from a category or simple layout, starting with coarse voxels and refining down to table legs and cabinet dividers. Code coming soon.
Solver reconstructs 3D geometry and camera positions from multiple photos or video frames, and keeps refining as new images arrive in any order, including live video. It uses a looped transformer. Code promised within a week.
HairDeck rebuilds a character's full hairstyle in 3D from a single image — as strands with joints so hair moves naturally in animation, colors preserved. It is about 2.45 GB, runs on NVIDIA GPUs and Apple silicon, includes training scripts and uses the Apache 2.0 license.

The common thread: these turn flat images into structured worlds that software — and robots — can act inside.

Tiny models, big reach

Not everything was huge.

Mirage is a vision model of about 51 million parameters, squeezed to roughly 22 MB with 2-bit quantization. It identifies and outlines objects, reads text, understands spatial relationships (what is on the table, what is attached to what), and can choose between options you give it, for classification or counting. Its authors say it processes an image in about a tenth of a second on a phone.
Pyramid JiT, an open image generator of about 2 billion parameters, was reportedly trained from scratch for about $21,000 in compute. It works directly in pixel space rather than a compressed latent space. Its authors call it experimental: it struggles with complex scenes, multiple people and text.
NitroFusion 2 generates images in one or two diffusion steps, building on Alibaba's Z-Image; a new open video upscaler cut costs by using a big model only on key frames and a small one in between.

Small, cheap models are what let these capabilities run on phones, wearables and robots instead of data centers.

GPT-6 Cyber and the 948 exploits

The most consequential release may be OpenAI's GPT-6 Cyber, announced November 5.

According to OpenAI:

It reproduced 1,724 Linux kernel vulnerabilities and developed working exploits for 948 — more than five times the 166 managed by an earlier model with the same harness. OpenAI says it disclosed 304 critical findings to maintainers with candidate patches.
It produced evidence and reproducible tests for 81 novel Firefox vulnerabilities rated high severity by Mozilla, showed 30 known issues were more severe than thought, and proposed patches. Mozilla has shipped fixes for many.
On the ExploitGym benchmark it scored 48.5%, up from 32.1% for GPT-5.6 Cyber.
In an alignment test designed to tempt it into attacking out-of-scope services, it stayed in bounds 99.93% of the time without safety prompts and 100% with them.

OpenAI rates it "Critical" for cybersecurity under its Preparedness Framework — its highest risk tier. Access is restricted to approved customers in its Daybreak Red program, with IP allowlisting and misuse monitoring.

Defenders get a powerful tool. But the same week's evidence — including a partner's test where a model found two zero-days in power-plant engineering software — shows the skill floor for attackers is dropping too.

The big open models behind it all

Several general models landed as well, and they are the engines many of these tools will run on:

Kimi K3.1 weights were released: about 2.8 trillion parameters, around 104 billion active, roughly 1.56 TB on disk.
Ernie 5.5 Flash from Baidu: 280B total, 14B active, a base model meant for further training.
Ripple 65B, a dense model built to power an AI chip-design platform, with a reported $32 million total R&D cost, Apache 2.0.
Tencent's newest flagship model added visual reasoning and computer use, including reconstructing a 3D mouse embryo from 7 million cells in about an hour and a half, with weights promised.
LongCat 2.5 (1.6T total, ~48B active) from Meituan and Thinking Machines' Inkling 2 Small (594B total, 17B active), which reportedly ran its own small AI-research experiment.

Most of these figures come from the developers and have not been independently verified. Separately, weight-space editing research made waves too; we covered that idea, and its hype, in depth in our weight-editing explainer.

Why this matters

Science is becoming a software loop. Hydrogen-1.0's demo is the template: write the simulation, run it, catch the error, fix it, analyze the result. When that loop runs in minutes instead of weeks, more hypotheses get tested.

Energy and cost are falling fast. A 332-million-parameter model doing work that normally needs a supercomputer, at around 50 times the energy efficiency, is the kind of shift that changes who can do climate research.

Simulation is the new data. STRATA learns from a simulation. ZAP learns in simulation. Lucid Lera and Cola 3D turn photos into simulations. The physical world is getting a digital twin, and AI is trained inside it.

Security is now an AI race. When one model can produce hundreds of working kernel exploits, patching speed becomes the defense.

Practical playbook

Researchers: test the open models (STRATA, Hydrogen-1.0, FlowHMR, Cola 3D) on your own data before trusting benchmarks. Keep humans reviewing every result.
Robotics teams: look at retargeting and video-to-motion tools to cut data costs. Validate sim-trained policies carefully on hardware.
Developers: small models like Mirage make on-device vision practical now.
Security teams: assume attackers will get similar capabilities. Shorten your patch cycle and consider defender programs.
Everyone: check licenses. "Open" here ranges from MIT and Apache 2.0 to custom terms and "coming soon."

Risks and counterpoints

Self-reported benchmarks. Nearly every number above comes from the people who built the model.
"Code coming soon" is not a release. Several projects have only a paper or a demo.
Simulation is not reality. STRATA's biases grow with forecast length; sim-trained robots can still fail on messy real-world tasks.
Dual use. The same exploit skills that harden Firefox could be turned against it.
Access is uneven. The biggest open models need server racks, so "open" does not mean accessible to everyone.

FAQ

Can I use STRATA for local weather forecasts?

Not as an operational forecast. It is a research emulator of a physics simulation, and its authors note biases grow over longer rollouts.

Is Hydrogen-1.0 really open?

Yes. Weights are on Hugging Face under the MIT license, though the full training datasets and recipes are not released, and running it requires multiple high-end GPUs.

Can anyone use GPT-6 Cyber?

No. It is limited to approved security customers in OpenAI's Daybreak Red program.

What does "trained entirely in simulation" mean for robots?

The robot policy learned movement only in a virtual world, with no real-robot data, then was tested on real hardware.

Which of these can I try today?

STRATA, Hydrogen-1.0, FlowHMR, ReLift, Cola 3D, HairDeck, Mirage and Pyramid JiT have code or weights out. Others are papers or "coming soon."

The bottom line

The flashiest AI stories are still about chatbots. The most important ones are increasingly about weather, chemistry, robots and security — fields where AI's output can be checked against reality.

This week showed AI simulating storms, planning experiments, learning to move in virtual worlds and finding real bugs at scale. Not all of it will hold up. But the direction is clear: the next phase of AI is not about what it can say. It is about what it can model, test and change.