Researchers built an open-source simulation where models pilot a 2,000-kilogram vehicle using cold-gas thrusters to execute stunts in winds reaching 33 meters per second. Frontier models completely failed real-time kickflip landings over a Cybertruck because their decision speeds of one to six seconds missed the required 0.2-second control window. After receiving a thirteen-dollar reinforcement learning post-training run, Qwen3-4B reached a 0.09-second reaction time. This allowed the compact model to achieve a 90% real-time landing rate on a 14-meter road.
Key points
Massive general-purpose models cannot handle dynamic aerial navigation when their inferen…: Massive general-purpose models cannot handle dynamic aerial navigation when their inference latency exceeds the 0.2-second operational threshold.
Small four-billion-parameter models can be specialized cheaply to dramatically surpass fr…: Small four-billion-parameter models can be specialized cheaply to dramatically surpass frontier architectures in real-time physical control.
Removing latency constraints allowed paused systems like DeepSeek to land 87% of stunts,…: Removing latency constraints allowed paused systems like DeepSeek to land 87% of stunts, proving computational speed was the primary failure point.
Sources checked by our analysis pipeline: 1
Last Thursday (October 15), Tesla's new Roadster drove off a ramp that curled over until the car was upside down, rolled upright in the air and landed. It then used thrusters to hop over a Cybertruck. Turns out the future does have flying cars.
The demo had already been pushed back two weeks, "due to high winds" (Elon Musk on X). That made us wonder: with a language model in the pilot's seat, how would the car do in that wind?
We built an RL environment around a physics simulator of a flying roadster to find out, handed the controls to LLMs, and turned the wind up. We let some of the latest best models fly it: DeepSeek V4.1 Flash, Meta's Muse Spark Open, Reflection's Beam 1.0, NVIDIA's Nemotron 3.5 Super VL, TypeSafe's Jev 1.14, and even OpenAI's GPT-6 Astra.
Asked for decisions in real time, these models crashed almost every flight: most take 1 to 6 s to answer, and the car needs a new action every 0.2 s. When we paused the world so the models could think…
Enterprise software giant Salesforce is bolstering its presence in the national security sector by absorbing specialized defense technology. The purchase integrates targeted artificial intelligence solutions designed specifically to manage complex military logistics operations. This strategy directly fuels the development of Salesforce's dedicated defense arm, which was established just twelve months ago.
The retail and cloud computing giant is committing massive capital resources to secure critical components for its next-generation computational infrastructure. This initiative directs substantial capital toward supply chain partners to support expanding artificial intelligence hardware requirements. The dedicated funding effort demonstrates the immense financial scale necessary to maintain advanced manufacturing pipelines in the tech sector.
Developers have compiled an operational 6502 central processing unit directly into a language model framework, bypassing the standard machine learning training pipeline entirely. By converting architectural logic into a neural network's internal states, the system faithfully replicates original gaming hardware through token prediction. This technical achievement demonstrates an unconventional method for repurposing transformer parameters to execute deterministic computing workloads.