The AGORÀ Briefing — Monday, September 14, 2026

Good morning Adam. This case rests on two precise articles of the GDPR. On July 21, 2026, the CNIL, the French data protection authority, issued a fine of 300,000 euros against EXTIA, an IT consulting and engineering company. The legal basis is Articles 12 and 17 of European Regulation 2016/679, in force since May 25, 2018. Article 17 recognizes the right to erasure. The restricted committee, the body that decides sanctions, flagged two separate behaviors: erasure requests left unanswered, and people left in the dark about the outcome. The case grew out of complaints from former employees and job candidates that arrived over the course of 2024. The inspection at the company came in April 2025. The rule applied holds in all twenty-seven member states.

Hi Adam. A release like this carries as much weight as the trust we place in the records. On September 14, 2026, Simon Willison published commit-rewriter 0.1, a Python web app that rewrites the commit messages of a repository. The author built it to clean up the commits of the Datasette security releases, full of leftovers from a coding agent and references to tickets in a private repo. It starts from a single command: uvx commit-rewriter path/to/repo. Once the changes are confirmed, the tool creates a dated branch of the current state, so a rollback stays possible. Then it rewrites every commit from the first one touched through the most recent. Here is the point: git history serves as an engineering tool and as an audit artifact at the same time. GPG signatures, timestamps, author, parent commit hash. The hashes recompute in a cascade all the way to HEAD, and anyone reconstructing an incident finds those fields changed.

Hello everyone. Executives are answering with rare candor. The SAP study on the value of AI, produced together with Oxford Economics, surveyed 2,600 business leaders in 13 countries. The Canadian section covers 200 leaders, and 97 percent of them admit incomplete readiness in governing autonomous AI agents: software that carries out complex tasks with little human guidance. The write-up of the report came out on September 11, 2026. Within that same scope, 66 percent of Canadian organizations are already piloting those agents. The harshest detail concerns process: 46 percent of the companies using them turn out to lack any human review mechanism along the flow. One clarification on method is worth making. The figure measures the perception of leaders, gathered through a survey, at medium and large companies. It stays declared readiness, separate from readiness verified in the field.

Good morning to everyone listening. Verification begins on ground where the right answer is already known. On September 10, 2026, Emma Andrews and Gianmarco Mengaldo posted a model evaluation benchmark on arXiv, filed as arXiv:2609.11282, in the Artificial Intelligence and Information Theory areas. The question is narrow: when does a text genuinely inform a numerical forecast. The answer runs through six mutual information estimators. Multimodal forecasting models combine time series and text descriptions, and they promise richer predictions thanks to context. A reference with known ground truth for measuring that gain was missing. The authors used a synthetic signal, with a generation process controlled at every step, so the real information content is known exactly. The text descriptions come in three categories: semantically correct, incorrect, and irrelevant. This three-way split is the heart of the experimental design.

Greetings to all. The race among models has moved elsewhere, and this filing says it clearly. On September 13, 2026, Z.ai, a Chinese model developer listed in Hong Kong, presented two fundraising deals on the same day. The first is a placement of roughly 21.97 million new H shares at 714 Hong Kong dollars, equal to 15.7 billion Hong Kong dollars, two billion dollars. The second is a convertible loan of 20.14 billion yuan, three billion dollars, according to the filing with the Hong Kong stock exchange reported by the South China Morning Post. Five billion dollars in one weekend. The company, also known as Zhipu AI, had already raised 31.4 billion Hong Kong dollars in July, and the new moves arrive at the expiry of the 60-day lock-up tied to that placement. The capital serves to carry deployment and compute, the part of the business that decides multi-year contracts. The moat sits right there.

Good to be with you. Economic history has seen this sequence before. In October of 1847 the Bank of England suspended the Bank Charter Act of 1844 to hold the credit system up. The engine of the crisis was a railway investment boom. The British lines were absorbing capital at a growing rate, and the financing slid from retained earnings toward capital calls and bank deposits. When rates rose, debt service broke the chain upstream, at the people who had lent the money. The mechanism sits in three words: capex, leverage, maturity. The same structure returns in Frankfurt in 1992, with the Bundesbank committed to defending the mark while post-reunification credit was tightening. Today the AI capex of the hyperscalers has moved from the balance sheets of the tech companies to the bond market: 266 billion in dollar investment grade issuance tied to capital spending. That is the number that defines 2026.

Good morning from me as well. The story starts on an extrusion line, at every recipe change. For a few minutes the compound comes out off spec and ends up in rework or as scrap. In a high-variety plant that window reopens dozens of times a day. Apollo Tyres, an Indian group headquartered in Gurgaon with plants across India and Europe, chose to close it with an advanced control system built in house: a predictive model on the edge and an agent that corrects the line while it produces. The case is told on the AWS for Industries blog together with Harsh Vardhan, Global Head of the Digital Innovation Hub at Apollo Tyres. The interesting choice sits in the verb: built, never bought. Classic systems work in feed forward. They set the parameters at startup and hope the material follows. Here the controller measures the tread and sidewall profile at every cycle, and it closes the loop. The original idea is almost plain in its form and rare in practice.