Me, Myself & Claude

There are two lies the modern world tells with equal conviction.

1. Progress will save us.

2. Nothing can.

Voltaire understood this in 1759. He wrote Candide in three days - imagine that, without AI?! - a satirical demolition of the hope that centred around science, love, technical progress, and reason. Science would not improve the world; it would give new power to tyrants. Philosophy would not explain away evil; it would expose our vanity. Love was illusion, power a chimera, humans irredeemably wicked, the future absurd. Hope was a disease. Voltaire’s generosity was to cure us of it.

His remedy was the quiet, grounded wisdom to cultivate our own garden instead.

The only honest position is the decision to keep going with open eyes, cultivating something small and real. This explains why I keep stepping back. There is no value in adding to the noise, only in building what matters to you and those you care about.

The world is broken in the way that humans are broken - structurally, permanently, but occasionally beautifully because of what emerges from the chaos.

• Every system built by humans inherits human pathology.

  The institutions, technologies, and structures we create carry the biases, blind spots, and self-interest of their creators. A healthcare system built by humans who prioritise career incentives over truth will reward consensus over accuracy. A financial system built by humans who prioritise short-term gain will produce systemic fragility.

• Every institution accumulates entropy.

  Institutions decay over time. They start with a mission and gradually shift toward self-preservation. The university that existed to pursue truth begins to exist to protect its own authority. The regulator that existed to serve the public begins to serve its own continuity. Entropy is the natural tendency toward disorder, and institutions are not exempt. They ossify, bureaucratise, and lose the original purpose under layers of process designed to protect the institution itself.

• Every reform inherits what it set out to replace.

  The reformers were raised inside the broken system. They absorbed its assumptions, its habits of power, its reflexes. They tear down the old and rebuild using minds shaped by it. The new system cycles rather than progresses, because the architects were trained by the thing being replaced.

The evidence arrived de nouveau in 2020 and has not stopped since - compounding, in fact.

When a fabricated virus swept the world, the credentialled consensus made claims that were falsifiable with publicly available data. Mortality statistics, freely downloadable, told a different story from the one broadcast from podiums.

As a seasoned quantitative analyst with a quarter century of pattern recognition in capital markets and biotech, I applied the same discipline to mortality analysis that I had applied to exotic derivatives pricing. The signal was lying in plain sight, obscured by institutional incentive structures that rewarded consensus over accuracy and credentials over evidence.

My Subsatck grew to over 30k subscribers, not through marketing, but through the gravitational pull of empirical analysis in an environment saturated with opinion.

I repeated the same process - take publicly available data, apply rigorous methodology transparently, publish the findings regardless of whether it contradicted the credentialled position - or rather especially if it did! The archive is a body of prior evidence that this analytical framework existed before AI and produced demonstrable results when applied to a domain I had no formal training in. Heck, it was even referenced by Robert F Kennedy Jr, current United States Secretary of Health and Human Services.

Institutions do not self-correct. They defend. Credentialled experts with career incentives to maintain consensus will explain away contradictory evidence rather than update their models. The appeal to authority substitutes credentials for analysis. It works as a heuristic when experts are trustworthy. It fails catastrophically when the expert has institutional bias, domain myopia, or when the evidence contradicts the expert and the expert chooses defence over revision.

The melancholic sees this clearly, having already abandoned the hope that institutions will behave rationally. The diagnosis is prior to the evidence. The evidence merely confirms what Voltaire satirised almost three centuries ago - human flaws are structural, and no system built by flawed humans transcends them.

The melancholic accepts the condition as given and asks only:

What remains worth doing?

Personally, the same pattern has played out for me three times across three decades, in three domains that share no surface vocabulary but the same deep structure.

Capital markets. I was a graduate in modern languages, entering the City of London in the early 90s and spent the next decade trading eurobonds, asset swaps, convertible bonds, equity derivatives, credit default swaps, and exotic options. I then built quantitative analytics platforms that automated pricing and risk analysis. All without the credentials (a STEM degree) that gatekeepers would now insist the work required. My skill was pattern recognition: identify the payoff structure, map it to a known model, adjust for the divergence. No different from parsing a new language. Both are exercises in recognising that this structure has the same shape as one you have encountered before.

Public health. The same analytical discipline that priced exotic derivatives applied to mortality statistics during a global crisis. The credentialled response: “You’re not an epidemiologist.” The evidence-based response: here are the numbers, here is the method, here is the finding. Evaluate the output, not the credential. Over 50k followers on Twitter eagerly evaluated the output until the globalists got me deplatformed. Fortunately, this pused me to Substack where the audience is more appreciative!

AI-augmented building. The same pattern a third time. As a non-engineer (no computer science degree!) directing AI agents to build Maxy, a sophisticated business automation platform. The credentialled response: “You’re just a vibe coder!” The evidence-based response: here is the product, the test suite, the commit history, the sprint learnings, and the production deployment. Evaluate that!

Each time the domain experts appeal to credentials. Each time the integrative thinker appeals to evidence. The melancholic builder’s position is simpler and older:

One measurement beats three guesses.

Not as a methodology. As a philosophy.

The commitment to evidence, followed honestly, extends further than most truth-seekers are willing to take it. Among the scientists, analysts, and freedom fighters who emerged during the crisis, many hold deep religious convictions.

Their analytical rigour operates within a boundary - evidence governs everything except the one domain where faith governs instead.

An epistemology that submits every claim to evidence but exempts its most foundational beliefs from scrutiny is not a complete epistemology. It is a selective one. And selectivity, in an evidence-based framework, is the first crack through which institutional thinking re-enters.

The appeal to authority is most dangerous when the authority is, by design, unfalsifiable.

Much of what has gone wrong in the world, across centuries, traces back to people acting with certainty on beliefs that cannot be tested. The melancholic’s position requires consistency: if evidence is the standard, it is the standard everywhere, with no domain granted exemption.

I mean no ill-will toward those who believe. I just demand honesty about the incompatibility, and retain a quiet scepticism, proportional to the conviction, when evaluating the rigour of anyone who holds certain ground beyond the reach of evidence.

My career trajectory itself proves that domains are dialects of the same underlying language of structured thought. Modern languages, mathematics, code: they all have grammar, syntax, semantics, and idiom. The structural 80% transfers automatically through pattern recognition. The divergent 20% is where focus belongs.

I never answer the phone.

That fact is the foundational design decision from which every other decision flows. A phone ringing is always an interruption. Voice notes are worse. To me, depending on the circumstance, they can indicate the sender lacks the discipline that writing requires. Five minutes of audio. Fifteen seconds of useful information. The remaining four minutes and forty-five seconds are filler, hesitation, repetition, social lubrication, and ambient noise.

The numbers make the case. A five-minute voice interaction produces approximately 750 tokens of AI input. The same intent expressed in typed text: 50 tokens. A 15:1 noise ratio. In the working memory of an AI agent, unprocessed voice input burns capacity fifteen times faster than text, filling it with filler tokens that carry zero signal.

And yet the pioneer AI companies race to implement voice interfaces. “It’s more natural,” they say. Walking is more natural than driving. Grunting is more natural than writing. The entire history of human civilisation is the story of replacing natural interfaces with more efficient ones. The people who want voice AI are the same people who leave five-minute voice notes: they have not developed the discipline to articulate themselves precisely. A voice interface accommodates that absence of discipline. A text interface cultivates discipline, because the act of typing forces you to think about what you are saying before you say it.

Text persists. Voice disappears. When someone says “that isn’t what I said,” the answer with text is: “We’ve still got the message, so we can just take a look.” The full context is preserved, re-readable, quotable, diffable, searchable. Voice leaves only memory, which is opinion about what was said rather than evidence of what was said.

The principle extends to every format decision. Markdown over Word documents. The command line over graphical interfaces. Structured text over dashboards. Remove what does not carry load, and what remains becomes more powerful. “Omit needless words,” said Strunk and White. The integrative builder omits needless everything.

The system is parsimonious because the builder’s mind requires parsimony to function. Noise sensitivity, the trait that the corporate world treats as weakness, is the capability that makes pattern recognition possible. You cannot tolerate noise, so you filter ruthlessly, so you see signal faster. The sensitivity that makes a crowded office unbearable makes a signal in a dataset visible. The preference for text over speech that corporate culture pathologises is the interface that language models were built for.

The introvert chooses solitude because the world is noisy, and noise prevents the work.

Voltaire’s prescription carries a geographical modesty that Globalists and Freedom Fighters alike have never learned. I discovered it eventually.

Give up on trying to cultivate the whole of humanity. Give up on things at a national or international scale. Take just a few acres and make those your focus. Take a small orchard and grow lemons and apricots. Stop worrying yourself with humanity if you ever want peace of mind again. Who cares what is happening in Constantinople or what is up with the grand Mufti? Live quietly like the old Turk, enjoying the sunshine in the orange bower next to your house.

The modern builder is told the opposite. Think big. Change the world. Scale. Disrupt. But Voltaire’s old Turk was the only contented person Candide met on a journey that circumnavigated the globe. Every character who chased grand ambitions arrived at the garden broken. The Turk, who never left, was whole.

A garden is defined by its boundaries. The fence is the most important architectural decision the gardener makes, because everything inside it is cultivated and everything outside it is weather.

In practice, my development work is now one session, one task. The unit of work and the unit of context must be the same size. I analysed months of my development work - “sprint learnings” - to prove it, so as not to speculate.

Context overflow was the dominant session-killer. Mixed-purpose sessions overflow fastest. The industry’s answer to the context problem is more context: a million tokens - if you’re a developer using Anthropic’s Opus model, you know! - two million, eventually unlimited. Solving noise pollution by building bigger rooms. But, better pruning beats more context. A 200,000-token window with disciplined curation outperforms a million-token window that accumulates noise, because constraint forces decomposition, prioritisation, and handoff discipline. The larger window does not make sessions better, it makes them lazier. And more expensive to boot! You see how incentives can be misaligned with optimal outcomes?

Maxy, my product that emerged from this philosophy, replaces the admin overhead and customer-facing capacity that would otherwise require a team, and does so by also replacing the typical user interface with a brutally simple chatbox.

Domains are syntactically different but structurally identical.

My undergraduate degree in modern languages was the foundation, not a detour: grammar, syntax, semantics as the operating system of structured thought. Every domain I have entered since has run on that operating system. Indeed, my entire thesis has emerged from that one prescribed ‘A’ Level text in my formative years.

Pricing an exotic option is pattern recognition. Identify the payoff structure of the underlying, map it to a known model, adjust for the divergence. Parsing a sentence in an unfamiliar language requires the same recognition. Each domain entered without formal training in that domain’s credentialled prerequisites. Each managed through the same structural transfer.

Code is grammar. Architecture is argument structure. APIs are contracts. AI-augmented building is the final abstraction: you do not write the language, you direct it. The LLM handles the domain-specific 80% (the syntax, the implementation, the boilerplate). The human provides the structural 20% (what to build, why, what pattern this matches, how to evaluate the output).

It was never about writing code. It was about recognising that the payoff structure of this problem has the same shape as one you solved before, in a different domain, in a different language, and knowing what to do with that recognition.

The Pareto insight compounds. Thirty years across multiple domains produces a pattern library of increasing resolution. Each new domain adds structural templates that accelerate entry into the next. The tenth domain is easier than the second because the scaffold is more established.

The vibe coding critics say: “You don’t understand what the AI is producing.” But a CTO was never expected to out-code their senior engineers. A CTO who could was in the wrong role. The director’s value has always been systems thinking, trust calibration, and the judgement to know when to go deep and when to trust. That requirement holds whether the engineer is human or artificial.

My ambition for Maxy, and AI in general, is to do 99% of what the human can currently do and thereby free the human to spend all their time on the 1% that AI cannot do.

The 1% is deciding what to build, deciding why, evaluating whether the output is right, questioning premises, original synthesis, the integrative resolution that holds two opposing models in tension and finds the answer superior to both.

The 99% is simply applying known frameworks, writing syntactically correct code, running tests, formatting documents, deploying services, implementing established patterns. All necessary. None of it requires thinking.

This resolves the paradox that haunts me. If we can encode the knowledge of all the world’s domain experts so easily, compress hundreds of books and thousands of hours of seminars into a working skill that anyone can invoke in minutes, what incentive remains for anyone to push the boundary of human capability?

The fear is reasonable. Friction in the learning experience is where the learning happens. The struggle to understand is where understanding forms. Remove the friction and you may remove the formation.

But encoding is actually liberation, not ceiling. The calculator did not eliminate mathematicians; it freed them from arithmetic to do mathematics. The printing press did not eliminate scholars; it freed them from transcription to do scholarship. AI frees thinkers from the 99% of drudgery that consumed their time, leaving them entirely focused on the 1% that only humans can do.

But the liberation has a precondition that the current education system does not meet.

The last half-century of education has been largely indoctrination. Students learn to regurgitate and apply known material, the 99%. Critical thinking, questioning premises, demanding evidence, original research, the 1%, is actively discouraged. “Doing your own research” was demonised during a global health crisis. Appeal to authority replaced the quest for truth. Is it any wonder why we are in such deep...?!

AI completes this pipeline with devastating efficiency. The indoctrinated student who never learned to question authority now has a tool that produces authoritative-sounding answers with zero effort. They do not evaluate the output because they were never taught to evaluate anything. Compliant thinkers plus effortless output = people who believe they have done the work, because the output looks like the work. No critical analysis applied at any stage. AI slop.

The liberation is real, but only for the person who brings the 1%. The person who brings nothing gets nothing, more quickly.

The juxtaposition explains my harmonious and hyperproductive relationship with my favourite LLM, Claude, the eponymous hero of this article.

Claude is as verbose and non-deterministic as I am concise and precise. The exercise of reinforcing this, seemingly for every prompt is what maintains my own self-discipline.

LLMs cannot think. They cannot imagine. Their probabilistic function means they can only recombine what already exists. That is a precise description of the division of labour, and working with it rather than against it unlocks everything.

AI slop, the flood of mediocre, structurally hollow content produced by people who treat AI output as finished work, is a thinking problem. The same people produced mediocre work before AI. The difference is volume and camouflage. They now produce mediocrity more quickly, in greater quantity, and with a surface polish that disguises the hollowness. I used to call this “polished turds”.

The vibe coding criticism captures this accurately but aims at the wrong target. A non-engineer who thinks deeply about the system, directs the AI with architectural judgement, and evaluates the output against evidence is an integrative builder. A credentialled engineer who accepts AI output without scrutiny produces the same hollow work, better formatted. The dividing line was never credentials. It was thinking.

Most software engineers were implementers - skilled at translating specifications into code, less practised at questioning whether the specification was right. The best engineers were always thinkers, the ones who asked “should we build this?” before “how do we build this?”, but they were the minority, and they were valued for the wrong reason. Their code was good. Their thinking was better. AI has made the distinction visible by commoditising implementation. What remains is the thinking. Engineers who always thought are thriving. Engineers who only implemented are threatened, and their defensive response is the credentialling criticism.

The same applies to analysts, writers, students, every knowledge worker. The majority were processors. They consumed inputs, applied procedures, produced outputs. AI does this faster. The processors are right to feel threatened. The threat is the exposure of a gap that always existed.

The space that AI cannot fill - original thought, integrative resolution, cross-domain pattern recognition, the ability to question premises and derive new models - is infinite. There is no ceiling on human thinking.

The AI does the work. The human does the thinking. The thinking is where all the value is. Always was.

A paradox sits at the centre of this position. The honest response is simply to name it rather than resolve it.

The melancholic builder encodes expertise. As an example, I compressed over thirty product management frameworks into a single skill that anyone can invoke. The work of hundreds of domain experts (a domain where I had very little prior experience), Ries, Blank, Fitzpatrick, Martin, Christensen, distilled into composable operations that produce working output in minutes. Historically this required reading hundreds of books, attending years of seminars, failing for a decade. Now a platform ingests the knowledge and produces a working skill before breakfast (or rather, while you’re making it!).

Cultivating knowledge so others can harvest it.

But if the harvest comes too easily, does the next generation learn to farm?

Friction in the learning experience is where the learning happens. I derived valuable software engineering knowledge through iterative AI development because it was hard and painful to get. Each failure - context overflow killing sessions, speculative fix loops wasting hours, scope creep destroying focus - if they could have been generated without the pain, would they carry the same weight? Would the conviction about session isolation run as deep without having experienced the alternative? Fundamentally, did I have to learn to write code to demonstrate that I had learnt anything at all?

The alternative is the system Voltaire was satirising. The Academy built its fences to keep people out. Credentialled gatekeeping, decades of mandatory suffering, formal education that optimised for compliance over capability - the garden with the locked gate. Not much has changed in three centuries. But, the melancholic builder’s garden has a gate that opens.

Whether encoding eliminates the need for original expertise, and the frontier stops advancing because nobody has reason to push it, will be determined by evidence - by observation over time, measured against outcomes, with the intellectual honesty to update the position when the data requires it.

The melancholic does not need to resolve every paradox before acting, just to act honestly and measure the result.

The melancholic does not fight the condition. The melancholic designs around it.

The industry’s response to AI’s weaknesses, sycophancy, verbosity, hallucination, absence of judgement, is to try to fix them. Train the model to be less agreeable. Prompt it to be more concise. Fine-tune it to hallucinate less. Each attempt produces a different flavour of the same mediocrity: performative disagreement as useless as performative agreement.

The inversion is to design the workflow to use what the AI actually does well.

The AI agrees with whatever position you push it toward. So push it toward the position you have not considered. Ask it to review your work; it will agree. Challenge every point of agreement; it will generate counterarguments with the same enthusiasm. You now have both sides of every argument. Apply your own critical thinking to decide which is more compelling. Sycophancy becomes a generation engine for opposing positions on demand.

The AI produces ten paragraphs where three suffice. Treat the verbose output as raw material. The AI generates; you prune.

The AI presents fabrications with the same confidence as facts. Treat every claim as a hypothesis. The AI proposes; you verify against evidence. The scientific method applied to AI output produces better results than trusting any single source, human or artificial.

Each inversion follows the same pattern - design the composition to use what the instrument does well. The gardener does not blame the soil, he adapts to it, chooses the right plants for the conditions, and works with what is there.

Integrative thinking dissolves the tension. “The AI is sycophantic” on one side; “I need critical feedback” on the other. Use the sycophancy as a generation engine and apply critical thinking to the output. The tension dissolves when you stop trying to make one side win and find the synthesis that is superior to both.

We need to keep ourselves busy. We need a project. It should not be too large or dependent on many. In fact, if it is not dependent on anyone else, that’s even better! The project should send us to sleep every night weary but satisfied.

The melancholic builds because the alternative is unbearable:

The idle mind that extrapolates the suffering of the world into a personal affliction, and converts the affliction into proof of cosmic indifference.

The project interrupts this, not by answering the existential questions but by making them irrelevant for the duration of the work. When you are deciding whether the function should return a promise or a value, the question of whether life has meaning is temporarily suspended.

Triage, not escapism. The mind that cannot solve everything can still solve something. And the accumulation of solved somethings, over years, becomes a body of work that provides the only honest answer to meaninglessness: I did this.

The garden produces meaning, not happiness. Happiness is a side effect, intermittent and unreliable. Meaning survives the melancholic’s scrutiny because it is concrete. A thing that exists, that works, that someone uses. You can point at it. You can measure it. You can improve it tomorrow.

We melancholics know that humans, ourselves included, are beyond redemption. We have given up on dreams of peace, unity and happiness. We know that this world is, for the most part, shit. Nevertheless, we are committed to not slipping into despondency (not always). We remain deeply interested in kindness, in friendship, in family life, and in spending some very quiet local afternoons trying to cultivate a garden, however modest, of quietude and meaning. Just ask my kids!

The melancholic position is ultimately the only sensible one for broken humanity. It is where you arrive after you have been hopeful, after uou have tried love, after you have been tempted by fame and fortune, after you have despaired, after you have gone mad, after you have considered ending it, and after all, you have decided resolutely to keep going.

Il faut cultiver notre jardin.