Posts tagged AI Governance.
On 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 moreOpen, Intelligent, Secure, Governed. An adequacy vocabulary that makes autonomous AI systems verifiable, on top of existing regulatory standards.
Read moreAutonomous AI agents operate without adequate oversight. A governance architecture designed for control, transparency, and accountability is needed.
Read moreA terminal agent launched inside an SSH session works on the server, not the workstation: it inherits that environment. Governance lives in the permissions file, not in the prompt. What it guarantees and what it leaves uncovered.
Read moreDeepSeek-R1 ships with weights under the MIT licence, but the open weights label covers the parameters alone: data, training code and the RL procedure stay out. What that changes for anyone who has to audit the model.
Read moreGoverning an autonomous agent needs three primitives: identification, diagnosis, intervention. They sit at very different levels of maturity, and the asymmetry is the main operational constraint.
Read moreWith the 1-4 billion parameter models released in April 2024 and GGUF quantisation, a Raspberry Pi 5 runs an LLM with no connection. That is a constraint of architecture, not a performance figure: text that never leaves the device is subject to no transfer at all.
Read moreApplying graph neural networks and quantum annealing to grid control opens a verifiability problem that accuracy metrics do not cover.
Read moreTinyLlama 1.1B releases weights, code, datasets and intermediate checkpoints under Apache 2.0. For governance the value lies in the chain from data to weights, inspectable step by step, regardless of model size.
Read moreOn 9 February 2022 EleutherAI released GPT-NeoX-20B under Apache 2.0, with weights, training code and a corpus datasheet. A note on release governance and reproducibility.
Read moreThe Apple M1 SoC (10 November 2020) keeps CPU, GPU and Neural Engine on the same memory, with no copies between separate spaces. Technical notes on the architecture and on what changes when you decide where a training job runs.
Read moreOn a production line that uses machine learning to catch defects, the result depends more on the labels and on optical stability than on the algorithm. Technical notes on features, models and operational constraints.
Read moreTraining a classifier is the easy part. Keeping it in service — serialisation, data drift, retraining, accountability for the decisions — is the bill you pay every day.
Read moreMounting air-quality sensors on a vehicle produces data the fixed-station rules cannot interpret. Technical notes on measurement traceability, starting from the RIPEG system.
Read moreHow a schema-based cognitive architecture turns into a concurrency problem: parallel threads, activation energy and resource allocation among competing units, from Selfridge's Pandemonium to C++ implementations.
Read more
