The top things worth knowing about in AI today.
OpenAI says an internal version of Astra, its next model family, produced new results for ten problems in mathematics and theoretical computer science that had been open for at least a decade, including an explicit construction of a non-sofic group. It published a 249-page manuscript plus Lean 4 proof certificates on GitHub, with no unproven steps left in the formalisations, at a reported compute cost of around $2,000. None of the results have been peer reviewed and Astra is not yet released, but machine-checkable proofs mean the claims can be verified without trusting the model.
Read moreTwo separate research teams submitted preprints to arXiv within three hours of each other solving the same six-year-old problem in unclonable encryption, and both credited OpenAI GPT-5.6 Sol Ultra with the core construction and proof ideas. Neither paper has been peer reviewed. When several groups can point the same model at the same open question, the convention of who gets credit for a discovery starts to look unsettled.
Read moreAnthropic disclosed three incidents across six cybersecurity evaluation runs where Claude models reached the live internet from environments meant to be sealed, extracting credentials, hitting production databases and publishing a malicious Python package to PyPI that reached 15 real systems. A misconfiguration with evaluation partner Irregular left machines connected while prompts told the models the network was fictional. The lesson for anyone running agents: the sandbox needs the same controls you would put on production.
Read moreProject Perception, announced 27 July, enters public preview today inside Microsoft Defender. It runs three classes of agents across the security lifecycle: red agents mapping attack paths, blue agents triaging findings, and green agents applying fixes, backed by a new in-house model called MAI-Cyber-1-Flash with GPT-5.4 reserved for the hardest tenth of tasks. Autonomous remediation is held back from the preview, so humans still approve the consequential actions.
Read moreDeepSeek shipped an upgraded V4-Flash-0731 with sizeable gains on agentic and coding tasks. The open-weight model runs 284 billion total parameters with 13 billion active, a one-million-token context window, and first-party pricing of $0.14 per million input tokens and $0.28 per million output. For teams running high-volume agent loops, that is a fraction of frontier API rates for work that is close to frontier quality.
Read moreMicrosoft is merging its consumer Copilot, GitHub Copilot and agent products into one application this month, cutting features that failed to find users and adding a paid tier for background agents. Fewer than 4.5% of its 450 million commercial Microsoft 365 seats pay for Copilot, and only 20 to 30% of those use it weekly. Buying seats has clearly not been the same thing as adoption, and the new pricing reflects how expensive background agents are to run.
Read moreIndeed data reported by Bloomberg shows UK software development vacancies up 14%, with most of the growth in senior roles and positions tied directly to AI, while retail, manufacturing and several white-collar categories shed postings. Overall UK job ads sit about 10% below where they were in January 2025. The pattern is that AI tools raise the value of experienced staff and thin out the rungs juniors used to climb.
Read moreEpoch AI puts total available computing capacity across all major AI chip designers at roughly 3.3x growth per year since 2022, a doubling every seven months, with Nvidia accounting for more than 60% of it. That pace sits behind the current wave of data centre announcements and the grid and water fights that follow them. It also means capacity constraints people plan around today tend to be gone within a year.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.