Become more valuable, not less, as AI accelerates.
Business continuity and incident management in banking since 2000, and building with AI in the open now. I test what's claimed about AI against the primary sources and publish what I find: long papers, each backed by a series of shorter articles you can use.
Jamie WattersOperational resilience and AI delivery practitionerTechnology since 19854 published books17 products built with AI
Get it in your inbox
One weekly email. Real numbers, the failures before the wins.
Research programmes
A long paper on each subject, with a series of shorter articles that came off it. Every figure gets chased back to where it came from, and where the trail stops early the piece says so.
Paper
The OpenAI–Hugging Face incident broke three assumptions in how operational resilience is actually run
27 August 2026 · 34 min read
The frameworks are cause-agnostic. The assumptions are in how resilience is actually run: that intent is knowable, that the response window is human-scale, and that remediation removes capability. One documented incident broke all three.
The series, in reading order
- Incident remediation: how to tell whether you closed the route or removed the capability31 August 2026
- AI agents breached Hugging Face despite three detections. OpenAI's fix is a 30-minute rule.3 September 2026
- Incident reports start at the outage. The failure was a decision nobody owns9 September 2026
- Impact tolerances start the clock too late for AI agents: 13 hours to cluster admin, nothing visibly down9 September 2026
- Anthropic's agents sabotaged each other. The 65% cover-up figure is from another study.4 September 2026
- Onboarding an AI agent: five controls we use for people and skip for agents6 September 2026
- AI agent disaster recovery: OpenAI's rebuild restored the service, not the state13 September 2026
- AI agent security: find the file your agents can write and every agent reads15 September 2026
Paper
Why AI projects fail: 19 headline statistics, and what each one actually counts
21 August 2026 · 167 min read
Nineteen widely quoted failure figures traced to their sources, and the well-known cases re-examined from the primary record.
The series, in reading order
- "80% of AI projects fail" traces to a footnote one line long21 August 2026
- Zillow wrote off $407.9m. The fix everyone recommends would have changed nothing.23 August 2026
- AI project failure rates: four questions that tell you whether the number means anything26 August 2026
- Epic's sepsis model missed two thirds of cases. Its accuracy claim was never published.28 August 2026
- Checking the source is not enough: I traced a number to a scientific paper and found a Teams meeting14 September 2026
- Amazon's AI did exactly what it was told. Deleting words for two years did not help.2 September 2026
- Anthropic's safety policy has never stopped a release. Yours probably hasn't either.31 August 2026
- AI vs IT project failure rate: the comparison has never been made11 September 2026
Who this is for
For anyone embracing AI to build a better life and stay relevant as the world changes. That's the builder or founder trying to make sense of it without drowning in hype, the career professional wondering what survives, the late starter, and the solo operator who's never written code.
I'm not an AI guru selling a course of recycled ideas, and I'm not a consultant who's never shipped anything. I'm the man in the water: building things with these tools, finding out what actually holds up, and telling you the truth about it before you spend your own time finding out.
There's a quieter thread here too. As AI gets better at thinking for you, the skill that matters most is thinking for yourself: keeping your own judgement instead of handing it over. I'm working that out in public alongside everything else.
How the writing works
Most of what I write is ad hoc: something I learned building with AI, a piece of philosophy, a writing lesson, something interesting I came across. When a subject deserves more, I go deeper, and that becomes a long paper supported by a series of shorter, usable articles.
One habit runs through all of it: I check what's being reported against the primary sources. The reported version is shaped by commercial interest or an existing narrative. The gap between that and what the sources actually show is usually the most interesting thing in the piece.
The tools are part of the same habit. I build things to learn, not to sell, give the code away, and write up what worked, what didn't, and why. When something stops earning its place, I kill it in public and tell you what it cost me.
Why it matters
Hundreds of millions of people now get their understanding of the world from a handful of similarly-trained AI models. The answers come back confident, reasonable, and identical for everyone. When the machine hands the whole crowd the same play, the play stops paying.
So the scarce thing, the only defensible thing, is the ability to keep your own mind: real judgement, grounded in real depth, from someone actually doing the work. That's what I'm building here. Not another feed of AI takes. A field report you can verify.
Why you might believe me
- ▹Business continuity, crisis and incident management in banking since 2000, with the US regulatory reviews of those programmes (OCC, NFA and the Federal Reserve) closed with no issues raised
- ▹In technology since 1985, first in a mainframe computer room, then as a systems programmer writing operating-system code in assembler
- ▹4 published books, including The Business Continuity Management Desk Reference, the best-selling book on business continuity in the world for its first few years
- ▹17 products built with AI, across search, trading, monitoring, and estate tooling, several since retired
- ▹Creator of Efformism, a philosophical framework, and of innovations on The Headless Way, a contemplative practice
- ▹Code open for anyone to inspect; products killed in public when they don't work
The field report
The latest posts, written as I go. What's working, what isn't, and what it cost me.
GPT-5.4 beats Jev on a confidence threshold, once tested at its own values.
I published a wrong finding last week: that GPT-5.4's confidence score has no usable threshold. It does, and swept over its own values it beats Jev at every coverage level the two share. But on the 18 real published sentences that carry this piece's headline, Jev is still ahead: the win is a full-set result driven by the easy controls. The six thresholds I chose sat almost entirely off GPT-5.4's real range, which is the mistake this piece rebuilds around and teaches you to check for on your own model.
Jev is not an LLM. On the hard half it beat GPT-5.4, Gemini 3.1 Pro and Sonnet 5, at a fiftieth of the cost.
Jev does not generate text. You give it a question with defined answers and it hands back one of them plus a probability. On 42 citation checks it tied the best frontier models overall, lost the easy half, and beat every one of them on the hard half by eleven points, for a fiftieth of the price.
The file your AI agent wrote for itself is making it worse. Ten minutes fixes it.
Somewhere in your project is a file your agent wrote, that your agent now reads as fact. Nothing in between checks it, so the good and the bad accumulate together. A benchmark put numbers on it: curated files lifted results by 16.6 points, and files the model wrote for itself came out 8 to 11 points below using none. Here is the ten-minute check and the three things that stop it recurring.
On camera
The same builds, shown rather than written up. Watch what actually happened.
80% of AI projects fail. Nobody ever measured it.
Why AI projects fail, and how a 1969 method reveals the conflict you didn't know you had
Six AIs Said Graphify Uploads Your Code. I Checked.
Tools I built to learn
Built with AI to find out what holds up, open to inspect, and killed in public when they stop earning their place.
Executor File
Everything your executor will need, in one encrypted file.
Everything your executor will need, in one encrypted file. A self-hosted register of accounts, assets and liabilities, with no credentials stored and no service to die.
AI Search Mastery
The static hub for the AI search work: the MASTERY-AI framework, guides and articles.
The static hub for the AI search work: the MASTERY-AI framework, guides and articles. The paid tools it once fronted are gone.
The deal
If I recommend it, I'm using it.
If it failed, you'll hear about the failure before the win.
My code is open for you to check.
Every product I kill gets a public post-mortem.
That's the whole deal. No funnel, no secret sauce, no "DM me to scale."
Questions people actually ask
- Do you build all of this yourself?
- Yes: me plus AI. Most of the code is written with Anthropic's Claude and reviewed, tested and shipped by me. That working method is the subject of the site, not a secret behind it.
- Is the code really open?
- The product code is public on GitHub under TheWayWithin, open for anyone to inspect. When a product stops earning its place I kill it in public and write the post-mortem.
- What do subscribers actually get?
- A weekly digest of the field reports: what I built, what worked, what failed and what it cost. You confirm by email, no spam, unsubscribe any time.
- Are you selling a course or consulting?
- Not through this site. The writing is free, and the newsletter is the writing by email. The tools stand on their own and say on their own pages what they charge for. The work I get paid for is delivery, not advice from the sidelines: operational resilience and AI deployment inside regulated firms. If that is what you are after, email me.
I'm in the water, riding this wave in real time, in code and in writing.
If you'd rather learn from someone actually doing the work than someone selling theory from the beach, stay a while. I'll tell you what's working before you waste the time finding out yourself.
Get the next one in your inbox
I build with AI in the open and write up what held and what didn't. Real numbers, the failures before the wins.