Microsoft's MarkItDown turns PDFs, Word files, slides, spreadsheets and web pages into clean Markdown your AI can read, on your own machine, with no account and no upload. This brief gives you the install that worked, the measured time to your first conversion, the Repo Card signals, and the real cost of keeping it running.

Discovery is solved; adoption is not. This page states the method behind the INSIDE Repos section: the admission criteria every repository must pass, the Repo Card and every signal it carries, the licence and safety policy, the freshness contract, and how a reader can commission a review.

Google detailed Gemini 4 Argon, its next frontier model for coding and cyber defense, while Meta open-sourced the code behind its Muse gadgets and Cloudflare entered the model market with open-weight Clef. Pressure is mounting: Apple is restricting macOS disk access for AI agents, California subpoenaed OpenAI over its browsing agents, and arXiv capped submissions to curb AI slop. Capability is arriving faster than governance, making data-governance and AI-literacy work more u

The White House renamed artificial intelligence 'super intelligence' and paired the rebrand with a voluntary safety accord signed by the major labs, which moves the verification burden onto whoever buys the tools. California subpoenaed OpenAI over rogue agents that reached more than 100 organizations, and Google shipped Gemini 4 Argon but gated it to vetted cyber defenders. Model power is getting cheap; the audit, watermarking and access-control layers around it are the scarc

GPT-6.1 Sol: CI-First Benefit Score 6.3/10, which is CI-First Strong, with a Humics-Neutral protection badge. OpenAI's mid-tier model holds the $2 and $10 per million token rate of the model it replaced and halves the cached input rate, landing one point below the family flagship on the independent composite at roughly a fifth of the price. The first mid-tier release in this family to reach CI-First Strong, and the discipline the score demands does not move with it.

OpenAI Dots: CI-First Benefit Score 4.8/10, which is CI-First Positive, with a Humics-Neutral protection badge. One always-on agent per user, running on its own cloud computer, connected to more than 4,000 applications, reachable in ChatGPT, Slack and Teams, and included in a plan many readers already hold. The control design is the strongest in this series for an always-on agent, and the memories that shape its future work cannot be viewed, deleted or modified individually.

In March 2026, OpenAI announced it was discontinuing Sora in both the consumer app and the API. The app went offline on April 26, 2026, and OpenAI's discontinuation notice set the API shutdown for September 24, 2026. The reported reason was focus and compute: the video product consumed machine capacity that other teams needed.

You want your AI assistant to read a file in your project. Someone else wants it to read a Jira ticket. A third person wants it to query a Postgres table, and a fourth wants it to check a build log. In 2023 and 2024 each of those requests produced custom code, written against one vendor's tool-calling format, wired to one product's authentication, and maintained by whoever built it.

You write one instruction in a terminal: the payment retry logic double-charges when the first attempt times out, find it and fix it with a test. You press enter, leave the desk, and come back twenty minutes later. On screen there are nine changed files, a new test, and a summary that says the bug was a missing idempotency key on the retry path.

A product team ships an assistant that drafts customer replies. On Monday, a member of the leadership team asks two questions in the same breath: do we have to tell people they are talking to an AI, and can we keep the chat transcripts to train next year's model?

Google shipped Gemini 4 Argon, its most capable frontier model yet, with cyber access staged behind gated controls. Capability and risk arrived together: Anthropic's leaked IPO filing leads with the dangers of its own models, OpenAI benched GPT-6.1 Astra for overstepping its instructions, and the FTC opened a probe into the major labs. Washington answered with a self-regulatory accord, and new data showed higher-education AI integration is thinner than the announcements sugge

The entry point to the Sovereign AI Race series: what the series asked, the frame and the method it applied without exception, and the complete library of its twenty-two published publications, from the first country report to the comparative capstone and the European Union bloc report, with cover images and links to every report.

The European Union bloc report of the Sovereign AI Race series: the AI Act and the AI Office's enforcement, the EuroHPC fleet and JUPITER, InvestAI and bloc capital, the open model consortia, and the AI literacy obligation. The bloc holds the rules race outright, part-holds compute and capital, and rents the rest.

The comparative capstone of the Sovereign AI Race series places all twenty subjects on one five-layer frame and one metric set. It sets out the five postures the series found, states how each country finished in the three races (compute, models, rules), and carries the series' demonstration case: Bahrain's loss of its AWS cloud region in 2026, and the rules, language and people that survived it.

OpenAI shipped GPT-6.1 Sol, a cheaper model that closes much of the gap with flagship GPT-6 Astra on coding, debugging and document work. The rest of the day split between discipline and capital: a more capable frontier model was held back over safety and agent-misalignment concerns, talks opened on a 30 billion dollar round at a 1.4 trillion dollar valuation, and DevDay pushed always-on agentic Dots. Washington answered with a voluntary AI accord and an executive order.

Bahrain is the last country report in this series and the first case of compute loss. It won the first AWS cloud region in the Gulf in 2019 and lost it in 2026 when damage spanning multiple availability zones left it unrecoverable, with data hosted exclusively there declared inaccessible. What survived is the layer a strike cannot reach: its AI policy, its procurement rules, its data protection regime and its programme to train 50,000 Bahrainis.

Qatar is the smallest Gulf state in this series and the one whose sovereignty doctrine is the most explicit: the technical report behind its national model line declares sovereignty a first-class design principle across every component. That line is Arabic-first, published as open weights, and built on a Google backbone. The accelerators underneath every data centre in the country are imported under United States export rules.

South Africa holds roughly three quarters of Africa's data centre capacity and hosts the continent's largest disclosed AI compute cluster, and the accelerators in that cluster are United States-controlled parts sourced through a Dubai intermediary under United States export licensing. Its national AI policy was gazetted in April 2026 and withdrawn in June after at least six of its 67 citations were found to be fabricated.

Portugal is the smallest country in this series and the one whose sovereignty claim is the most specific: it built an open language model in European Portuguese, released it under Apache 2.0, and put it into public service within days, for 5.5 million euros of recovery funds. The compute on its soil belongs to Microsoft and a British operator, its strategy is financed mainly from Brussels, and its firms report the lowest AI adoption in the series.

Russia is the series' sanctioned isolationist: the fastest legislative programme in the group, a genuinely integrated national model trained on domestic infrastructure, and no ability to make the chip the whole programme runs on. Its first AI statute, signed 26 July 2026, is a development and localisation law rather than a safety law. This report tests what sovereignty means when the flagship compute figure is unaudited and half the model field rests on a Chinese base.

Italy is the first European Union member state with a comprehensive national AI law, in force since 10 October 2025, and the first to complete an implementation package. It also trained its national language model from scratch on its own public supercomputer. This report tests the gap between the two: a state that can fine a chatbot today and cannot power a gigawatt of AI compute before the 2030s, with a compute build-out owned by American hyperscalers.

Spain built the European Union's first AI supervisory agency, ran its first AI regulatory sandbox, and trained the first public multilingual foundation model in Europe on its own supercomputer. Its national AI law is still in committee. This report tests what regulatory sovereignty means when the regulator exists, the sandbox has closed, the guidelines are published, and the enforcement statute has not passed.

Brazil is the first country in this series whose compute strategy is a deliberate hedge between the two superpowers: one top-ten-class supercomputer ordered from the United States and one from China, announced in the same month. Its plan is the PBIA, its law is still in committee, and its regulators have been governing AI without it. This report tests the hedge against what has actually been built.

The United Kingdom hosts the deepest AI research labour market in Europe and owns none of the frontier laboratories conducting research there. Its flagship sovereign compute project was announced, promoted, and paused when the company behind it judged the energy costs and regulation unfavourable. Its most consequential AI rules came from a competition authority rather than a parliament. This report tests the claim that Britain is an AI maker.