Meta shakes up the frontiers, open and closed
Interconnects analyzes how Meta's latest moves reshape both open and closed frontier AI. The newsletter's verdict is that US open-weight models are back.
Interconnects analyzes how Meta's latest moves reshape both open and closed frontier AI. The newsletter's verdict is that US open-weight models are back.
A Guardian column argues that Chinese open-weight models from developers such as DeepSeek and Xiaomi, released monthly and usable without costly licences, pose risks for Western businesses and governments. It contrasts them with Western AI, which it says comes with at least some security effort.
StepFun has published Step 5 Preview as an open-weight model, with FP8 and BF16 checkpoints on Hugging Face and ModelScope. The sparse mixture-of-experts model is built for long agentic tasks, with 600 billion parameters in total, 27 billion active per token and a context of up to one million tokens.
Perplexity has released SPACE, the sandbox that runs every task in Perplexity Computer, under the Apache 2.0 license. Each agent task gets its own Firecracker microVM, with credentials and outbound traffic controlled from outside it, which could help teams building coding or browsing agents.
Google researchers have open-sourced EnvHarness, a framework that adjusts agent-training environments to an agent's weak spots without changing the agent, the environment or the evaluation setup. The researchers say it can make learning both more effective and more efficient.
China intends to restrict exports of 23 core AI technologies, including model weights, source code and training methods for systems above 10^26 FLOPs or 2 trillion parameters. Open-source releases would still be cleared quickly once registered, but the move could push Western firms toward American open-weight models.