A proposal for a permanent national institution that would train and certify the clinical AI competence
of Kenya’s entire health workforce — doctors, clinical officers, nurses, pharmacists and
laboratory staff alike.
The argument in one line: the variable that decides whether clinical AI helps or harms a patient is not
the model, it is the training of the clinician in front of it. A 2025 randomised trial found that
physicians who had already completed twenty hours of AI-literacy training still deferred to deliberately
erroneous model output — so the obvious curriculum is the wrong curriculum.
Five professional tracks, five gated levels, assessment by simulation and workplace observation rather
than attendance, and outcomes published whether or not they flatter us. The blueprint, competency
standards, five-year financial model and evaluation design are published in full.
Read the proposal →
AI & Medicine
Advanced — technical and clinical readers
⏱ 70 min
2026-10-12
The Institute has written at length about how it would teach and examine clinicians at every level. It has not yet said how it would examine the model those clinicians will consult. In the last essay I argued that the weights can be vouched for only by their behaviour, like a drug. That is half right. A model does not behave like a drug; it behaves like a candidate — it takes a history, reasons, advises, and can be wrong in all the ways a junior clinician can be wrong, plus several that no human can. This essay sets out the examination I would require before the Institute lets a model near a clinician, and the revalidation I would require afterwards. It examines the model through three windows. From the outside: a secure Kenyan item bank, standards set before any model is scored, a station battery borrowed from the OSCE, metamorphic tests borrowed from software engineering, reliability measured over repeated attempts rather than one, and graders whose own error is corrected statistically. From the inside, which is the one thing no examiner can do to a human candidate: linear probes, sparse dictionaries, attribution graphs and model diffing, used to ask whether the model knows it is being examined, whether its expressed confidence matches its internal state, whether its citations are load-bearing, and what a fine-tune or a quantisation actually changed. And around it: Syft and Grype for the software bill of materials and vulnerability scanning the Pharmacy and Poisons Board now expects, and a self-hosted Langfuse as the record of practice. It is also explicit about what interpretability cannot yet certify, which is most things.
Read article
In June I wrote a synthesis of what Cathie Wood and Raoul Pal said on Diary of a CEO about AI, crypto and the decade ahead. Since then three things have happened. Pal gave a long new interview on When Shift Happens in which he says we have about four years before AI changes how the economy works; his first book, The Everything Code, came out — I bought it and have already read more than a third, and I find it highly enlightening; and my own scorecards on the AI-boom stocks showed that earnings mattered far less than liquidity and the cost of money, which is very close to Pal's central claim. This is my attempt to set out, in depth, how I understand the framework underneath both of them: Pal's Everything Code — growth as population plus productivity plus debt, a refinancing cycle that drives liquidity, and debasement as policy rather than accident — Wood's deflationary convergence of five technology platforms, where the two agree and where they quietly contradict each other, what Pal actually said in October about his own trades, and where I now part company with the version of this argument I published in June.
⏱ 30 min
Read article
Three weeks ago I said the earnings season would tell us more than the previous three months had. It did, about the companies. Seven of the nine names from the July piece reported, and all seven beat on the quarter. All seven fell — six on their next trading day, Apple in after-hours trading on Monday evening. SanDisk printed $11 billion of revenue at an 85% gross margin, guided higher again and announced an eight-for-one split, then lost 13.5% in a session. Western Digital cleared the gross-margin signpost I set for it by more than two points and still fell. Intel had its worst day in decades because the memory shortage I wrote about in July is now cutting into its own customers' demand, and Apple — record revenue, guidance above consensus — fell because memory costs will take its gross margin down to 45–46%. Since 9 October the equal-weighted basket is down another 15%; since 1 July it is down 19%. This is the second scorecard: what each company reported, why the stocks fell anyway, how the signposts I set three weeks ago actually scored, and what I now think the market is pricing.
⏱ 26 min
Read article
AI & Medicine
Advanced — arithmetic only, but every step is shown
bitcoinLightning Network
2026-11-02
In September I published a working note about Malipo Rail, the payment system I am designing for the Kenya Institute for Clinical AI, and called it a draft 'while it is still wet'. It asserted that Lightning would carry the Institute's money and never explained how Lightning works. This is the full piece. It builds the Lightning Network from nothing, with numbers you can check: a payment channel between a clinic and the Institute, worked through five states, including what happens when one side cheats; the hash time-locked contract, with a real SHA-256 hash, that lets a payment hop across strangers without trusting any of them; a three-hop route with every fee and every timelock computed backwards from the destination; onion routing; why a payment can fail when every channel on the route is big enough, and the probability model that explains it; and liquidity — the operational problem that decides whether any of this works. It then crosses into shillings: how M-Pesa actually moves money, what the Daraja API exposes, and how a hold invoice can be wired to an M-Pesa callback so that a bitcoin payment and a shilling payout either both happen or neither does. Finally it re-specifies Malipo Rail in the light of all of it — including the places where working it through by hand showed that my own draft was wrong — and brings the Kenyan regulatory picture up to date, two days before the licensing deadline.
⏱ 50 min
Read article
AI & Medicine
All readers — arithmetic only, no prior maths assumed
Learning With ErrorsLWE
2026-10-19
My patient record is called MedLattice, and across four essays I have spent something like twenty-five thousand words on its cryptography. The mathematics it is actually named after — the lattice problem that ML-KEM rests on — got one paragraph. I took Shor's algorithm apart by hand and then asked you to take its replacement on trust. This post repays that debt. It builds Learning With Errors from nothing: clock arithmetic, two simultaneous equations you can solve on a napkin, the same two equations with a whisper of noise added, and the moment the napkin method collapses. It shows why the problem is secretly a question about finding the nearest point on a grid in hundreds of dimensions, locks and unlocks a single bit by hand with numbers you can check, rebuilds the whole thing in a spreadsheet and in forty lines of Python, and then walks from that toy to the real ML-KEM-1024 that protects every MedLattice record — rings, modules, the number-theoretic transform, the decryption-failure budget and the transform that makes a fragile scheme safe to use. It then puts the assumption on trial against everything 2026 has thrown at it: a claimed quantum break of Module-LWE posted in May and dismantled within weeks, a provable new analysis that shaves up to 68 bits off one classical attack estimate, quantum resource estimates for RSA falling towards a hundred thousand qubits, and a side-channel paper presented eight days ago. The mathematics held. The honest conclusion is about where it did not need to.
⏱ 45 min
Read article
On 2 July I argued that the sharp pullback in memory and storage stocks was an entry point, not an exit, and walked through nine companies across the AI hardware stack. Three months later the fundamentals I was relying on came in even stronger than I expected — SanDisk's revenue rose 372% in a single quarter at an 85% gross margin, Micron printed a $54 billion quarter — and the stocks I was most enthusiastic about are among the worst performers in the group. An equal-weighted basket of the nine is down 5.6% while the semiconductor ETF is up 1.9%; the two names I was coolest on, NVIDIA and Apple, are the two best. This is the honest scorecard: what happened company by company, why record earnings met falling share prices, which parts of the bear case I wrote and then underweighted turned out to be the story, and what I am now watching.
⏱ 22 min
Read article
AI & Medicine
Advanced — technical and clinical readers
local LLMopen-weight models
2026-09-28
Commercial chatbots hedge when asked for medical advice, and the obvious conclusion is that the Kenya Institute for Clinical AI should run its own model so that its clinicians get straight answers. I think the conclusion is right and the reason is wrong. Refusal is the weakest argument for local inference; the strong ones are that an API call carrying patient data is a cross-border transfer under section 47 of the Digital Health Act, that a hosted model can change underneath a validation study without notice, and that the Pharmacy and Poisons Board's 2026 guideline on medical device software now expects change control, rollback and traceability that only an owner of the artefact can provide. This essay sets out what running the Institute's own model actually involves: which open-weight models are serious candidates and what each is built from, what 'controlling the source code' can and cannot mean when the thing that matters is a trillion numbers nobody can read, the hardware in three tiers with landed Kenyan costs, the power and backup design for a grid that went dark nationwide on 29 July, the full software stack, the team, a five-year cost of ownership, whether DeepSeek V4 is an answer, and — at length — what none of this does.
⏱ 55 min
Read article
AI & Medicine
All readers — no technical background assumed
quantum computingShor's algorithm
2026-09-22
Every conference I go to lately, somebody wants to talk to me about whether AI is about to take over the world. Nobody has ever once asked me whether the same era of computing might quietly read every patient record I have spent this year designing MedLattice to protect. It should be the other way around. AI takeover is speculative — disputed timelines, disputed probabilities, no working example. Shor's algorithm is not speculative. It has existed, proven, since 1994; the only open question is how many years of engineering stand between now and a machine large enough to run it. This post is the explanation I wish existed before I had to go and build cryptography around the problem myself: what encryption actually is and how a record like MedLattice's gets locked in the first place, why a classical computer cannot pick that lock no matter how large it gets, what a quantum computer is actually doing differently — with superposition explained properly, not just named and left — Euclid's algorithm worked by hand because it turns out to be the quiet backbone of the whole attack, Shor's algorithm itself walked through on numbers small enough to check on paper, and exactly how a quantum computer would derive a private key from a public one. It closes with why AES-256 is not the part that's actually in danger, and why the lattice mathematics behind ML-KEM resists this attack in a structurally different way than RSA or elliptic curves ever could.
⏱ 38 min
Read article
AI & Medicine
All readers — no technical background assumed
HyperledgerBesu
2026-09-18
MedLattice's own specification states its topology in one line — 'permissioned EVM consortium (Besu/QBFT)' — and moves on. This post is the argument behind that line, built from the ground up: what a blockchain is actually doing underneath the word, why a permissioned chain and not a public one, what the Byzantine Generals Problem is and why every consensus protocol in use today is answering a question posed in 1982, how Practical Byzantine Fault Tolerance became Istanbul BFT became QBFT, why Besu specifically rather than Fabric or Corda, and how MedLattice's own three contracts — IdentityRegistry, ConsentCapabilityManager, RecordAnchorRegistry — sit on top of all of it. It closes with what the choice does not solve.
⏱ 25 min
Read article