The knowledge hub: AI and engineering from inside China. Build logs of real systems, including what broke.https://www.francisokafor.comen-ushttps://www.francisokafor.com/insights/china-university-major-cuts-ai-policy-nigeriahttps://www.francisokafor.com/insights/china-university-major-cuts-ai-policy-nigeriaChina's education ministry revoked or suspended 12,200 undergraduate programme slots in five years and added 10,200. The number everyone repeated is real. What it means has been reported wrong.Tue, 01 Sep 2026 09:00:00 GMThttps://www.francisokafor.com/insights/what-happens-when-ai-gets-cheaphttps://www.francisokafor.com/insights/what-happens-when-ai-gets-cheapGeneric reasoning costs $0.05 per million tokens. The premium moved to what nobody wrote down: how one mould behaves, what a Lagos trader really charges, what a supplier means by 差不多.Tue, 01 Sep 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/keycloak-single-sign-on-distributed-siteshttps://www.francisokafor.com/field-notes/keycloak-single-sign-on-distributed-sitesRealms, PKCE, token clocks, brokering and the database nobody budgeted for. What actually breaks when single sign-on spans a dozen sites, and what to decide before it does.Mon, 31 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/llm-api-pricing-dollars-revenue-naira-token-taxhttps://www.francisokafor.com/insights/llm-api-pricing-dollars-revenue-naira-token-taxToken prices are set in dollars while African revenue is set in naira, and tokenizers charge some African languages nine times more tokens for the same sentence. That maths decides viability, not modeSat, 29 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/remote-work-brain-drain-nigeria-engineershttps://www.francisokafor.com/insights/remote-work-brain-drain-nigeria-engineersAn engineer who stays in Lagos and works for a foreign employer is present but economically absent. The country keeps the person and loses the tax base, the mentorship and the firm never built.Wed, 26 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/is-ai-a-bubble-engineer-assessment-capex-depreciationhttps://www.francisokafor.com/insights/is-ai-a-bubble-engineer-assessment-capex-depreciationHyperscaler capex reached roughly $745bn for 2026 while the two largest labs run at about $105bn combined. The bubble is real. The compute will outlive it. An engineer's read from Shenzhen.Wed, 26 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/chinese-open-source-ai-models-two-storieshttps://www.francisokafor.com/insights/chinese-open-source-ai-models-two-storiesThe weights are real and the licences are real, and they say different things. Read the Chinese announcement next to the English one and the gift starts to look like a distribution strategy.Wed, 26 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/knowledge-graph-vs-vector-databasehttps://www.francisokafor.com/field-notes/knowledge-graph-vs-vector-databaseVector search retrieves what looks similar. It cannot tell you which shipped assemblies a bad plating bath touches two hops out. That is traversal, and traversal is a different operation.Fri, 21 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/harness-engineering-ai-agent-guardrailshttps://www.francisokafor.com/field-notes/harness-engineering-ai-agent-guardrailsThe model is a component you swap. The harness around it, tool schemas, validation, sandboxes, budgets, permissions and audit logs, is where nearly every production AI incident actually starts.Thu, 20 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/vertical-saas-foundation-models-generic-layerhttps://www.francisokafor.com/insights/vertical-saas-foundation-models-generic-layerThe wrapper layer is gone. What survives is proprietary data, workflow depth, regulatory position and distribution, and the proof sits in Harvey's post-training run and a 180 million euro EU tender.Thu, 20 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/made-in-china-quality-2026https://www.francisokafor.com/insights/made-in-china-quality-2026The label used to mean cheap and derivative. In batteries, solar, drones, robots and open-weight models it now marks the frontier. The perception lag is worth money to anyone testing on evidence.Wed, 19 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/ai-agents-for-non-technical-teamshttps://www.francisokafor.com/field-notes/ai-agents-for-non-technical-teamsAgentic tools have moved out of the private chat window and into the folders a team depends on. What that relocation actually changes, checked against vendor documentation as of August 2026.Wed, 19 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/china-humanoid-robots-shipping-supply-chainhttps://www.francisokafor.com/insights/china-humanoid-robots-shipping-supply-chainAn ageing workforce and an explicit industrial policy point the same direction. Beneath the humanoid demos sits a component supply chain that quietly changed hands: reducers, sensors, actuators.Wed, 19 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/how-to-stay-relevant-as-a-developer-ai-shifthttps://www.francisokafor.com/insights/how-to-stay-relevant-as-a-developer-ai-shiftTwo randomized trials of AI-assisted developers reached opposite results. The difference was not the tool. It was how much of the task depended on context nobody had written down.Tue, 18 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/why-chinese-ai-labs-release-open-weightshttps://www.francisokafor.com/insights/why-chinese-ai-labs-release-open-weightsChinese labs release open weights because a 2017 State Council plan told them to export standards rather than products, and because downloaded weights cannot be switched off by anyone.Sun, 16 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/chinese-open-weight-ai-africa-infrastructure-influencehttps://www.francisokafor.com/insights/chinese-open-weight-ai-africa-infrastructure-influenceChinese firms built much of Africa's power and telecom. Now the AI models are Chinese and free, and free is a more durable form of influence than any loan.Sat, 15 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/ai-detectors-flag-non-native-english-writershttps://www.francisokafor.com/insights/ai-detectors-flag-non-native-english-writersAI writing detectors measure how predictable your prose is, which is a proxy for native fluency rather than for cheating. A watermark can prove a model was involved, never who wrote what.Fri, 14 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/signals/bottleneck-moved-to-review-not-product-managementhttps://www.francisokafor.com/signals/bottleneck-moved-to-review-not-product-managementAndrew Ng is right that cheap code exposed a decision problem. He is describing an old constraint with the lights on. The new one is verification, and it does not scale.Fri, 14 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/software-engineering-skills-ai-cannot-replacehttps://www.francisokafor.com/insights/software-engineering-skills-ai-cannot-replaceProducing a function was never the scarce skill. Deciding what should exist, what breaks it and what the system costs in three years is the work, and none of that got cheaper when generation did.Wed, 12 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/how-ai-text-watermarking-workshttps://www.francisokafor.com/field-notes/how-ai-text-watermarking-worksThe mark is not added to finished text. It is made of the model's word choices, and it can show a model was involved without ever showing how much of the writing was yours.Tue, 11 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/four-levels-of-ai-fluency-prompter-automator-builder-orchestratorhttps://www.francisokafor.com/insights/four-levels-of-ai-fluency-prompter-automator-builder-orchestratorMost organisations are staffed entirely at level one and are surprised when level three does not happen by itself. The four levels of AI fluency are separated by what you own when it breaks.Thu, 06 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/signals/taalas-hc1-amd-bought-the-memory-trickhttps://www.francisokafor.com/signals/taalas-hc1-amd-bought-the-memory-trickThe HC1 burns Llama 3.1 8B into silicon and quotes 17,000 tokens a second. Taalas' own product page calls it a technology demonstrator. That phrase is the whole story.Thu, 06 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/how-to-test-llm-applications-without-calling-the-modelhttps://www.francisokafor.com/field-notes/how-to-test-llm-applications-without-calling-the-modelThe model is the only part of an LLM application that cannot be tested deterministically. Everything around it can be, and that surrounding code is where most production incidents actually start.Wed, 05 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/chinese-ai-models-daily-engineering-work-from-shenzhenhttps://www.francisokafor.com/field-notes/chinese-ai-models-daily-engineering-work-from-shenzhenBenchmark parity is not workflow parity. The four Chinese frontier models are genuinely competitive on price and quality, and the reasons a Western team still cannot use them are not technical.Mon, 03 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/ai-regulation-comparison-eu-us-china-african-unionhttps://www.francisokafor.com/insights/ai-regulation-comparison-eu-us-china-african-unionFour regimes, four instruments: a product safety file in Brussels, a court docket in Washington, a filing number in Beijing and a review date in Addis. The artefacts converge. The triggers do not.Sun, 02 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/content-credentials-prove-capture-not-authorshiphttps://www.francisokafor.com/insights/content-credentials-prove-capture-not-authorshipContent Credentials prove a file's chain of custody. They do not prove a person made the decisions. Two very different products are being sold under one promise, at $10 and at $9,195.Sun, 02 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/ai-model-availability-by-regionhttps://www.francisokafor.com/insights/ai-model-availability-by-regionWhich AI models you may use is settled by export licences, regulatory filings, data residency rules and licence territory clauses. Capability is not a gate anywhere in that chain.Sun, 02 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/seniors-2034-entry-level-hiring-aihttps://www.francisokafor.com/insights/seniors-2034-entry-level-hiring-aiJudgment is manufactured in the boring ninety percent of junior work. Shut the intake now and the shortage lands in the mid-2030s, when nothing can fix it quickly.Sat, 01 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/ai-agent-security-controls-before-you-sign-offhttps://www.francisokafor.com/field-notes/ai-agent-security-controls-before-you-sign-offA language model cannot tell your instructions from the text it retrieves, because both arrive as the same token stream. Every real control follows from accepting that rather than trying to fix it.Sat, 01 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/designing-apis-for-ai-agentshttps://www.francisokafor.com/field-notes/designing-apis-for-ai-agentsAn API read by a model has different requirements from one read by a developer. Names become prompt, errors become instructions and idempotency stops being optional once the retry loop is autonomous.Sat, 01 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/ai-visual-inspection-defect-detection-lighting-false-rejectshttps://www.francisokafor.com/field-notes/ai-visual-inspection-defect-detection-lighting-false-rejectsClassical vision still wins geometric checks. Learned models earn their place on variable and rare defects. Accuracy is set by lighting, fixturing and the false-reject threshold you can afford.Sat, 01 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/what-ai-video-generation-is-actually-good-forhttps://www.francisokafor.com/field-notes/what-ai-video-generation-is-actually-good-forGenerative video in August 2026 is a shot generator, not a film generator. The documented limits explain why, and they explain exactly where it earns its place.Sat, 01 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/agent-control-loop-design-production-agentshttps://www.francisokafor.com/field-notes/agent-control-loop-design-production-agentsPrompt engineering was never the hard part. The loop is: what the agent sees, what it may touch, what stops it and what a step costs. Most agent failures are loop failures, not model failures.Sat, 01 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/agentic-ai-in-production-what-changed-what-did-nothttps://www.francisokafor.com/insights/agentic-ai-in-production-what-changed-what-did-notFrontier models resolve about 95 percent of SWE-bench Verified and 20.6 percent of OSWorld 2.0. The gap is not noise. Step count, not task difficulty, is what still breaks agents in production.Sat, 01 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/ai-legacy-code-modernization-comprehension-not-translationhttps://www.francisokafor.com/field-notes/ai-legacy-code-modernization-comprehension-not-translationLegacy modernization was never mainly a translation problem. The hard parts are comprehension and behaviour preservation, and a model's real value sits almost entirely in the first one.Sat, 01 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/ai-in-the-global-south-priced-in-dollars-written-in-englishhttps://www.francisokafor.com/insights/ai-in-the-global-south-priced-in-dollars-written-in-englishThe AI risks that dominate the global conversation are the ones that bite hardest in rich countries. From Lagos and Shenzhen the salient questions are language, price, labour and state capacity.Sat, 01 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/genome-language-models-engineer-read-evo-2https://www.francisokafor.com/field-notes/genome-language-models-engineer-read-evo-2DNA is a sequence and transformers are good at sequences. That analogy carries genome language models further than it should. What Evo 2 actually does, what it is used for and where the pitch breaks.Sat, 01 Aug 2026 09:00:00 GMThttps://www.francisokafor.com/insights/who-writes-robot-safety-standardshttps://www.francisokafor.com/insights/who-writes-robot-safety-standardsRobot governance is settled in ISO committees, insurer-funded biomechanics labs and tender documents, years before any parliament votes. In July 2026 the EU moved its own deadline to wait for the stanFri, 24 Jul 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/digital-twin-in-manufacturing-stalls-at-level-twohttps://www.francisokafor.com/field-notes/digital-twin-in-manufacturing-stalls-at-level-twoA digital twin is only useful when live telemetry feeds it and a decision comes back out. A 3D model that never updates is a rendering, and the data feed was always the expensive part.Fri, 17 Jul 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/ai-research-assistant-literature-review-grounded-synthesishttps://www.francisokafor.com/field-notes/ai-research-assistant-literature-review-grounded-synthesisGemini Notebook answers only from sources you supply. That fence moves the error rather than deleting it, and the part it cannot do, judging whether your corpus is any good, stays with you.Thu, 16 Jul 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/sample-efficiency-reinforcement-learning-foal-62-attemptshttps://www.francisokafor.com/field-notes/sample-efficiency-reinforcement-learning-foal-62-attemptsReinforcement learning brute-forces locomotion in simulation. A foal walks after about 62 attempts. The gap is not compute, it is the structure a random policy does not arrive with.Wed, 01 Jul 2026 09:00:00 GMThttps://www.francisokafor.com/insights/technology-as-soft-power-side-doorhttps://www.francisokafor.com/insights/technology-as-soft-power-side-doorInfluence is rarely won by broadcasting. It is won by becoming infrastructure: the handset default, the network the engineers trained on, the standards seat, the freely licensed model weights.Wed, 01 Jul 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/ai-data-privacy-checklist-before-shipping-ai-featureshttps://www.francisokafor.com/field-notes/ai-data-privacy-checklist-before-shipping-ai-featuresRetention windows, zero-retention gaps, personal data in vector stores and the erasure problem, written for the engineer who has to configure it before Friday.Wed, 01 Jul 2026 09:00:00 GMThttps://www.francisokafor.com/insights/how-bubble-tea-subsidies-manufactured-ai-adoption-in-chinahttps://www.francisokafor.com/insights/how-bubble-tea-subsidies-manufactured-ai-adoption-in-chinaChina's mass AI adoption was manufactured by firms owning payments, delivery and a super-app, not by consumer curiosity. Subsidy bought usage. Whether it bought habit is a separate question.Tue, 30 Jun 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/event-backbone-distributed-manufacturing-siteshttps://www.francisokafor.com/field-notes/event-backbone-distributed-manufacturing-sitesAn event log across distributed plants buys ordering per partition, replayable history and one shared dependency, and most factories would be better served by a database for another five years.Thu, 25 Jun 2026 09:00:00 GMThttps://www.francisokafor.com/insights/wheel-legged-robots-locomotion-tradeoffshttps://www.francisokafor.com/insights/wheel-legged-robots-locomotion-tradeoffsWheels are cheap per metre. Legs handle discontinuity. The hybrid tries to buy both and pays in actuators, distal mass and a switching problem. An engineer's account of when that trade is worth it.Wed, 17 Jun 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/llm-evals-regression-tests-probabilistic-softwarehttps://www.francisokafor.com/field-notes/llm-evals-regression-tests-probabilistic-softwareTeams that cannot measure their AI system cannot improve it, so every change becomes an argument. Evals are the regression tests of probabilistic software and the first thing to build.Thu, 11 Jun 2026 09:00:00 GMThttps://www.francisokafor.com/insights/ai-for-small-business-six-wins-three-ways-to-waste-moneyhttps://www.francisokafor.com/insights/ai-for-small-business-six-wins-three-ways-to-waste-moneyThe AI wins for a small business are boring: quoting, document extraction, translation for China trade. The waste is predictable: platforms bought before anyone defined the process.Thu, 11 Jun 2026 09:00:00 GMThttps://www.francisokafor.com/field-notes/five-prompt-engineering-best-practices-that-hold-uphttps://www.francisokafor.com/field-notes/five-prompt-engineering-best-practices-that-hold-upMost prompting advice is folklore repeated without a control. Five practices survive testing, four popular ones do not, and the research reversed direction as the models improved.Mon, 08 Jun 2026 09:00:00 GMThttps://www.francisokafor.com/insights/shenzhen-hardware-manufacturing-density-not-cheap-labourhttps://www.francisokafor.com/insights/shenzhen-hardware-manufacturing-density-not-cheap-labourShenzhen stopped being cheap years ago. What it still sells is density: an iteration loop that runs in four days because the mould shop, the fab and the assembler are all inside an hour's drive.Mon, 01 Jun 2026 09:00:00 GMT