Meta Reads Your Mind, DeepSeek Speeds Up, and Cursor Goes Mobile

Meta Reads Your Mind

Meta's research team has unveiled Brain2QWERTY v2, a non-invasive system that decodes whole words and even semantics from raw brain activity in real time, using MEG and EEG rather than surgical implants. Trained on nine volunteers in controlled typing settings, it reaches roughly 61% word accuracy overall and 78% for the best participant. Meta is releasing the training code and an accompanying dataset. It's lab research, not a consumer product, but it narrows the gap with invasive brain-computer interfaces in a striking way.

DeepSeek's DSpark Rewrites the Speed Rules

DeepSeek introduced DSpark, a new approach to speculative decoding that improves both how draft tokens are generated and how they're verified. Reported gains include a 30.9% higher accepted length versus Eagle3 and it's already being integrated by the vLLM community as a leading single-GPU speculative-decode path. It's the unglamorous plumbing that makes AI faster and cheaper to run.

Cursor Puts Cloud Agents in Your Pocket

Cursor launched an iOS app with always-on cloud agents you can launch and steer from your phone, including diff review and notifications. Coding assistants are becoming background workers you check in on, rather than tools you sit in front of.

Devin Fusion and the Rise of the Harness

Cognition's Devin Fusion claims a 35% cost reduction for high-quality coding by combining an expensive planner with cheaper models for bounded subtasks. The industry's centre of gravity is shifting from picking one best model to engineering clever orchestration around several.

GLM 5.2 Brings Frontier Models Home

Builders are running the open GLM 5.2 model, at hundreds of billions of parameters, on home hardware, including a 753-billion-parameter version sharded across two Macs at around 16 tokens per second. Companies are also bundling discounted access to open models like DeepSeek, Kimi and Qwen.

Snowflake's Arctic RL Cuts Training Time

Snowflake's open-source Arctic RL project features ZoRRo, delivering up to a 6x actor-update acceleration and cutting a text-to-SQL training run from roughly five days to about 36 hours.

Claude Lands on Azure

Anthropic's Claude Opus 4.8 and Haiku 4.5 are now generally available in Microsoft's Foundry platform, with enterprise identity, billing and governance built in.