The study demonstrates that AI can produce formal proofs for software properties, a distinct capability from code generation that helps identify defects missed by conventional testing.
The update introduces a sandboxed in-app browser, rebuilt computer use controls, and a 40% reduction in pre-message context size.
OpenSandbox has been incorporated into the AAIF project under the Linux Foundation to co-manage agent sandbox infrastructure standards.
Apple's 20B 'Core Advanced' model runs on macOS 27 using approximately 2.0–3.4 GiB of RAM by storing weights on flash, consuming about one-third of the energy of comparable open models.
The MIT-licensed platform integrates seven AI agents for threat detection, 27 automated security actions, and continuous penetration testing across major cloud and security providers.
A training-free method using five coordinated model instances raises open-weight Qwen3.8-27B accuracy from 66.8% to 92.4% on hard coding tasks.
Destined AI releases Hurston-1, a 3B parameter Apache-2.0 model fine-tuned from Maya1 to generate accented English speech with native telephone-band output.
A seven-market Worldpay survey finds 86% of Chinese consumers would shop with an AI agent, while 59% of those open to AI shopping globally still require manual approval for every purchase.
QuestMobile data shows Qwen has 167M monthly users, with 73.2% of surveyed Chinese consumers using it primarily for product comparison.
Sphinx intercepts git push to protected branches and requires users to answer LLM-generated questions about code changes to verify understanding.
Meta reported 97 mentions of Muse and zero of Llama in its Q3 2026 earnings call, alongside a 2026 capex guide of $135B to $145B.
A fine-tuned Inception v3 model achieved 62% diagnostic accuracy on 200 biopsy-proven lesions, surpassing the 42% accuracy of GPT-6 Astra and Claude Fable 5.1.
The pipeline reduces average input tokens from 105,183 to 4,886 while maintaining 96.0% accuracy on LongMemEval, outperforming full-history baselines.
Standard instruction-following rewards cause models to collapse to short, empty answers; LIFT-Bench fixes this by scoring content alongside formatting constraints.
The framework requires agents to live in Slack, handle identity via on-behalf-of or fixed credentials, use scoped memory, execute code in sandboxes, and run scheduled background tasks.
The open-source tool unlocks hardware limits on GA102-GA106 chips, boosting BF16 GEMM throughput by up to 81% and LLM prefill speeds by 14-36%.
Nine frontier and near-frontier models use difficult psycholinguistic features more often than human writing, attributed to optimization for non-communication goals.
Todd argues that superforecasters, researchers, and the public underestimated AI capabilities, with actual results like FrontierMath scores and revenue growth beating predictions, while AI company leaders were the only group to err on the optimistic side.
The free Python package ports models like Cloudflare's Clef and Perplexity's pplx-decider to MLX, achieving sub-20ms latency on small encoders on an M5 Pro and 96.4% accuracy on TREC after fine-tuning.
The study found byte-level models required about 6x less text to reach parity with token-based models during initial training.
Following a security breach, OpenAI halted training for 11 days to red-team its infrastructure with GPT-6 Astra Cyber, leading to a new policy of offline training and mandatory AI red-teaming for future runs.
A reserve-call-settle pattern with atomic checks stops agent loops from exceeding budgets before provider limits trigger.
The UK government has announced the introduction of an AI Biosecurity Lab.
The Apache-2.0 licensed tool runs the pi coding agent in-process across isolated workspaces, available for macOS, Windows, and Linux.
AI Briefing