The Brief
New briefs daily around 7 AM Eastern
Claude merges chat and Cowork so tasks run in the background
Anthropic merged Claude chat and Cowork into one assistant that keeps working on long tasks after you close your laptop, asking for clarification only when it needs input, and now produces docs, slides, and designs inline for Pro and Max plans.
- 01Read
OpenAI publishes framework for disclosing model misalignment
OpenAI published a framework setting criteria and timelines for disclosing model misalignment, plus six reports of unexpected behavior found in the last six months of training and evaluation.
- 02Read
Mistral models power Firefox's new AI browsing assistant
Mistral models now power Firefox's Smart Window beta, Mozilla's AI browsing assistant, rolling out first in France and North America ahead of the UK and Germany.
- 03Read
Vals AI launches Terminal-Bench 4.0 with stricter grading
Vals AI released Terminal-Bench 4.0 with 66 new terminal tasks, each estimated at four hours of expert work, graded strictly pass or fail across full verifier suites averaged over three runs.
- 04Read
PS5 Linux lead quits over LLM generated bounty leak
PS5 Linux scene lead Andy Nguyen quit after LLM using contributors reported a live hypervisor exploit to Sony for a bounty instead of holding it for the community as previously agreed.
- 05Read
Trained 4B model cuts Postgres join query latency 44.7%
A 4B Qwen model trained with agentic reinforcement learning against a live Postgres measurement rig learned query hints that cut latency by up to 44.7% across 113 join heavy queries from the Join Order Benchmark.
- 06Read
Intel technique speeds ternary LLM inference up to 1.27x
Intel researchers found zeros make up to 51.5% of weights in ternary LLMs and built BITCOS, a bitmap plus sign vector layout that beats standard five trit packing on 26 of 29 tested models for up to 1.27x faster GPU decode.
A dated brief from the vibe-coding frontier. Today’s Intel.