Notes
Type
Topic
35 notes
035Claude Mythos 5.1 and Gemini 3.8 Flash Cyber: the safeguard became an access tierAnthropic and Google each shipped a frontier model at two safeguard levels within two days in September 2026, with the permissive version reserved for vetted organisations. The control moved from the prompt to the customer, and that changes what a security team has to prove.034Claude Opus 5.5 and GPT-6 Sol: flagship capability at mid-tier prices, three weeks laterOn 22 September 2026 Anthropic shipped Claude Opus 5.5 and OpenAI shipped GPT-6 Sol and Luna. The capability that was the flagship at the start of the month now costs a mid-tier price, which makes every model routing table written in August out of date.033Decision models: what Jev, Solar Decide and Span-01 mean for enterprise workflowsIn September 2026 three vendors shipped models that return a typed answer and a probability instead of text. For routing, triage and compliance flags that is a better interface than prose, provided you demand the calibration evidence and build a branch for "unsure".032Claude formalised Fermat's Last Theorem in Lean: what 13 million checked lines teach about verifying agent outputAnthropic's agents produced the first end-to-end, computer-checked proof of Fermat's Last Theorem in 11 days. The mathematics was not new. The verification was, and so is what it shows about where trust has to sit when output outgrows anyone's ability to read it.031Meta Muse Spark 1.3 and the Contributor tier: the discount is paid in training dataMeta released Muse Spark 1.3 on 2 September 2026 with a model trained to ask before it acts, and kept a Contributor tier that costs a fraction of Standard in exchange for training on your prompts and completions. The behaviour is the better story. The pricing is the one that needs a policy.030OpenAI Astra for Law: the model is not the product, the index isOpenAI launched Astra for Law on 17 September 2026. It is GPT-6 Astra plus legal instructions, settings and a US-only Legal Search Index, not a new model. The benchmark gain came from retrieval, and that is the lesson for anyone building legal AI outside the US.029OpenAI DevDay 2026: dots agents get their own computer, a day after OpenAI held back GPT-6.1 AstraOpenAI launched dots, always-on agents with their own cloud computer, browser and rules, on 29 September 2026. The day before, it said the next Astra model would not ship because it did not meet OpenAI's bar for staying within scope and authorization. Read together, those two announcements are the most useful thing DevDay said to anyone building agents.028OpenAI switched off the Sora API with no replacement: what a deprecation policy actually promisesOpenAI removed the Videos API and every Sora 2 model on 24 September 2026, exactly six months after notice, and named no successor. The policy was honoured to the letter. That is the part worth reading if you build on someone else's model.027Salesforce's 16 September outage: a stalled login service and the dependency nobody listsFor about seven and a half hours on 16 September 2026, Salesforce customers across every region hit severe delays and errors because requests were stalling on an internal login service. The failure pattern is old and well understood. What it asks of companies that run their operations on SaaS is less often written down.026Ternary Bonsai 2 27B: Qwen3.8-27B in 5.9GB, and where the 98% stopsPrismML released Ternary Bonsai 2 27B on 17 September 2026, a ternary-weight build of Qwen3.8-27B that fits in 5.9GB and keeps 98.2% of the full model's benchmark average. On long-horizon agent tasks it keeps about three quarters. Here is how to read both numbers before you replace a 4-bit quant.025A successful pilot is not a decisionThe most common outcome of an enterprise pilot is not failure. It is a pilot that works, is agreed to have worked, and is followed by nothing.024The code got cheaper. The reading didn'tAn agent can produce more code in an afternoon than I can properly read in a week. The constraint moved, and most of the workflows built around it have not.023The EU AI Act's high-risk deadline moved. The transparency rules did not.The 2 August 2026 date a lot of teams were building towards was deferred to December 2027 by the Digital Omnibus. But Article 50 transparency did start applying on 2 August, and it catches ordinary AI products, not just high-risk ones.022MCP went stateless: what the 2026-07-28 spec changes if you run agentsThe Model Context Protocol dropped sessions at the protocol layer. No initialize handshake, no Mcp-Session-Id header, every request self-describing. Here is what actually breaks, what gets easier, and what I would change first.021A 27B model under Apache 2.0: what Qwen3.8-27B changes for local AIAlibaba published Qwen3.8-27B weights on 14 August 2026 under Apache 2.0, dense rather than mixture-of-experts, with a 262k native context. The licence matters more than the benchmark scores, and here is why.020Why Arabic-first software is more than translating an English UIDirection, typography, terminology and information hierarchy all change. Doing it at the end guarantees an interface that reads as a translation, and users notice.019Build logs are more valuable than founder motivation postsNobody has ever hired anyone because of a post about resilience. Showing what you built, what broke and what it cost is a different kind of signal.018Saudi Arabia is an interesting place to build enterprise AINot for the reasons usually given. The interesting parts are the constraints - data that will not move, Arabic as a first-class requirement, and buyers who ask hard questions early.017A CMS is only good if the person using it doesn't need a developerThe complaint was specific - uploading galleries and video was too complicated. That is a better brief than most product requirements.016cXML isn't REST: lessons from building a SAP Ariba integrationA protocol from a different era of enterprise software, with different assumptions - and REST habits actively work against you.015Why enterprise integrations are harder than the API documentation suggestsThe documentation describes the happy path against a system you do not control, cannot test against, and will be debugging weeks after the failure.014What I actually use AI for as a technical founderA practical list, including the parts where it does not help and the places I have decided not to let it near.013The best architecture is sometimes hybrid: local + cloud AIHybrid is usually the right answer and usually implemented badly - arrived at by accident rather than designed, which gives you the costs of both and the clarity of neither.012Why internal tools deserve better UXEmployees use internal software for eight hours a day and never get to choose it. That is an argument for more design attention, not less.011Local AI vs cloud AI: what actually makes sense for Saudi companies?The comparison usually gets made on price, which is the least interesting axis. Here is how the decision actually breaks down, including when the honest answer is cloud.010Why I'm interested in Mac minis for local AI deploymentsA small machine on a shelf is a genuinely viable AI deployment target now. Here is what that is good for, and the parts people leave out.009Building software for receptionists, guards, hosts and admins at the same timeMulti-role UX is a data-modelling problem before it is an interface problem. Get the shared object right and the screens follow; get it wrong and no design rescues it.008"No source, no claim" should be the default for high-stakes AIIn legal, compliance and finance, a confident wrong answer costs more than a refusal - and unlike a refusal, it does not announce itself.007AI agents are easy to demo. Production agents are different.The demo works because you chose the input. Production is where an agent meets a permissions model, a flaky API and someone who will not read the output carefully.006What I learned rebuilding a Framer site in codeA faithful rebuild is not a pixel copy. The judgement is deciding which details are the design and which are artefacts of the tool.005How I think about SEO for AI search enginesThe honest version: most of it is old advice done properly, plus being an entity a machine can identify with confidence.004The architecture behind a source-grounded legal AI assistantRetrieval as a gate rather than a hint, citations carried as data rather than text, and why corpus preparation matters more than model choice.003Why every technical founder should understand their production stackNot so you can do everything yourself. So you can tell the difference between a two-hour problem and a two-week one while somebody is telling you it is fine.002SEO is not just titles and keywords anymoreThe work that matters now is making a machine understand what an entity is - connected structured data, consistent identity, and content that can be cited.001What building visitor management software taught me about physical operationsSoftware that touches a physical space has a constraint most products never meet - there is a queue, and the queue does not wait for your loading state.