The architecture divides labor across the RP2350 processor, dedicating one core to USB and display management while the other runs the generative and decision-making models. To operate within strict hardware limits, the system tokenizes visual frames into patch grids written in NumPy that encode vital state details like headings or minimaps. Users can let an autonomous model play, take manual control of the generated environment via a button press, or train a new reinforcement learning agent directly on the hardware.
Key points
Dual neural networks can simulate game mechanics and play them simultaneously without usi…: Dual neural networks can simulate game mechanics and play them simultaneously without using the original underlying game code.
Hardware efficiency is achieved by executing between 2 million and 8.2 million multiply-a…: Hardware efficiency is achieved by executing between 2 million and 8.2 million multiply-accumulate operations per frame using looped architectures.
The platform supports dynamic interactions, allowing human players to interrupt the simul…: The platform supports dynamic interactions, allowing human players to interrupt the simulation or training fresh reinforcement learning agents directly on the device.
Sources checked by our analysis pipeline: 1
Using 2 small neural networks to generate and play the game at the same time on a Raspberry Pi Pico 2 W. The first one creates the game: given the current frame and the action, it predicts the next frame and the reward. The second one plays that game: it uses the generated frame and picks an action.
Every frame above was sampled by the world model network, and the actions were taken by the actor network.
I got interested in looped transformers, where one transformer block is applied several times with the same weights instead of stacking different layers. Reusing weights is cheap on memory, so I wanted to see if a looped world model could be small enough to run on a microcontroller. I picked the Pico 2 W because it is cheap and has an FPU, and small games because I could write the real versions directly into numpy and check every generated frame against them. Later I added a streaming RL agent so the board could also learn to play inside its own dream.
The chipmaker's potential financial backing aims to support the debt required to construct a massive compute facility for an emerging AI cloud provider. This arrangement would directly expand the physical infrastructure available to power intensive model development. Ultimately, the partnership positions Anthropic as the primary tenant to take advantage of the newly developed capacity.
Researchers demonstrated that an automated system built on Z.ai's GLM 5.3 can target the vulnerable window between a fix's publication and its deployment to users. By analyzing publicly available code updates, the system crafted zero-interaction exploits against Mozilla Firefox. This breakthrough lowers the financial and technical barriers to executing sophisticated cyberattacks, posing a heightened risk to open-source software ecosystems.
Major search engine updates are disrupting discovery patterns, pushing traditional web results downward and significantly diminishing referral traffic across the internet. Platforms are adapting their product strategies to build standalone audiences that rely on internal feeds rather than external web indexing. This shift threatens long-standing search distribution models and prompts publishers to defend their proprietary content.