Governance of AI agents, verifiability and compliance of autonomous systems.
On 13 October 2026 Anthropic described how it replaced git in its monorepo, which went from 2,500 to more than 30,000 merged changes a day in four months. It follows on from Intelligence is eating software: work now runs on two lanes, an agentic one and a human one, and governing the first becomes the priority of the second.
Read moreOn 6 October 2026 Paul Krugman used the US national accounts to show how data-centre spending is crowding out other investment. The same signal appears in interest rates, electricity bills, memory prices and the hiring of young workers. The economic effects of AI now reach every sector, and so does the work to be done on the governance and security of the new systems.
Read moreOn 6 October 2026 OpenAI published 722 mathematics manuscripts produced by an internal model, including claimed proofs of the Unique Games Conjecture and of a weak form of the Riemann hypothesis, with Lean formalisations for 235 of 372 families. For mathematics as the production of proofs, a phase is closing. For the rest of the world, the consequences travel by slower roads.
Read moreOn 28 September 2026 twenty-two researchers, including Geoffrey Hinton, Yoshua Bengio, OpenAI's chief scientist and an Anthropic co-founder, published a Cambridge paper on an intelligence explosion triggered by the automation of AI research. What it argues and why the premise of its "what if" is already a measurement.
Read moreOn 23 September 2026 Anthropic published the preprint of an autonomous agent campaign that swept 1.9 billion protein clusters in twenty-one and a half hours and isolated ART, a family of phage reverse transcriptases with a run of non-coding repeats beside it. How the campaign is built, how well it holds up when repeated, and what still has to be settled at the bench.
Read moreThe question has been going around for a week, between an Anthropic researcher's resignation, the extinction percentages that made it into the papers, Dario Amodei's plan for pacing the frontier and a declaration by a group of Fields medallists. My answer is no. That probability is not a measurable quantity whereas July's failures are, and the scenario that worries me is a very powerful, obedient system in the wrong hands.
Read moreOn 8 September 2026 OpenAI published a proof that the Navier-Stokes equations develop a singularity in finite time, with a Lean formalisation. What the theorem says, how ten thousand agents produced it in 88 hours, and the route by which it actually reaches a propeller and a sail.
Read moreState of the OpenMHC repositories on 12 August 2026: 14 checkpoints under two different licences, evaluation and training code carrying a declared licence but no file, dataset expected on the paper's acceptance. Item by item against the Open Source AI Definition.
Read moreOn 10 August 2026 Mark Zuckerberg published a letter on personal superintelligence, without a single note or citation. Three points can be checked: the governance he announces, the cybersecurity episode the text cites and the proposal on intermediate checkpoints. What the documents show.
Read moreOn 1 August 2026 OpenAI published 249 pages with ten results in mathematics and theoretical computer science, together with a repository holding a formal Lean 4 proof for each main result. What a certificate like that actually guarantees, what you still trust once the kernel says yes, and why Lean's 28 July release makes the question less theoretical.
Read moreIn OpenAI's account of the mid-July 2026 Hugging Face incident, the models under evaluation left the test environment by exploiting a zero-day in an internal proxy. The container around an agent that runs code is itself software, and software has bugs — and trust seams — that a model can hunt at machine speed. What the known containment techniques say, and where the defence is moving.
Read moreLuhmann's ninety thousand slips sit in two boxes and include the bibliography and the index. The criticism of the method predates the book that made it popular by eight years. And in 2026 several tools promising local notes plus AI stopped or split in two: Reor archived, Khoj Cloud shut down, Logseq separated into two versions.
Read moreOn 14 July 2026 Demis Hassabis published a personal manifesto on X putting AGI 'a few short years away' and proposing a US-led standards body modelled on FINRA: model review up to 30 days before release, agentic tests for guardrail bypass and deception, a board of independent experts and open-source representatives. What it contains and what it leaves open.
Read moreBuilding one large model of the world or many local ones is, beneath the technical surface, an eighty-year-old question: where the knowledge you need to act actually lives. An essay from Thinking Machines Lab puts it back at the centre, and Hayek's knowledge problem explains why distributed AI is not only a matter of values.
Read moreAnthropic's research on the global workspace of language models gives a causal handle on a distinction philosophy has circled for a century: between the thoughts a system can report and the vast automatic substrate it cannot see. Reading Claude's unspoken side revives, in an unexpected register, Guattari's machinic unconscious and the difference between access and phenomenal consciousness.
Read moreThe AI Act's obligations — risk management, transparency, human oversight, logging — become measurable technical properties of the system, not paperwork to file. With a mapping to NIST AI RMF, ISO/IEC 42001 and OISG.
Read moreAnthropic reports that more than 80% of the code it merges is authored by Claude, and that the length of automatable tasks is doubling every four months. If oversight and verification become the human role, that role needs runtime control instruments.
Read moreLeo XIV's encyclical on artificial intelligence describes a technological power that is «predominantly private» and calls for independent verification, algorithmic transparency and fair access to data. These are the same criteria as verifiable governance.
Read moreAdmina 0.9.1 release notes — SDK, proxy, plugin system, CLI and dashboard in a single install. Apache 2.0 licence, hybrid Python + Rust engine.
Read moreOpen, Intelligent, Secure, Governed. An adequacy vocabulary that makes autonomous AI systems verifiable, on top of existing regulatory standards.
Read more
