I sat with that for a while (longer than the conversation deserved, if I’m honest, because by then somebody else had moved on to arguing about whether AI is going to give us all cancer). And what I kept coming back to is: that is a sentence only a person who could have had a perfectly good life without the internet would ever say.

Because I couldn’t have. In 1996 I ran a search, on a machine that took four minutes to load a page and sounded like it was gargling gravel, and it is the reason I ended up applying to Cambridge. Take that search away and you get a smaller, considerably less interesting Vanessa: still clever, still furious, just never handed the door. The internet didn’t ruin my life. It’s structurally one of the two or three reasons I have the one I’ve got.

So no, I don’t think the problem is the internet. And when I look properly instead of just reflexively bitching, which I do plenty of, I don’t actually think the problem is AI either.

The objections people are throwing at AI right now are, mostly, completely correct. They’re just aimed at the wrong noun. Nobody’s angry at matrix multiplication. I spent a decade in finance, in exactly the kind of institutions that teach you the market self-corrects if you just get out of its way, before I ever touched a needle, so I know the tune when I hear it: what people are actually angry at is what happens when you hand the technology to people who believe regulation is theft and the public is a resource to be extracted from, then let them build the thing with nobody’s hand on the wheel.

Take data centres. Nobody’s objecting to “computers use electricity,” we made our peace with that around the time of the toaster. The objection is that they’re being dropped into places already short of water and power, and the shortage lands on whoever was already living there, never on whoever’s paying for the compute. You’ll notice nobody is proposing to put a hyperscale data centre two streets from Trump Tower. Funny, that. Scarcity, it turns out, is something we do to other people’s postcodes.

Or take Meta, whose own internal research reportedly found that when a teenage girl deleted a selfie, the system could read it as a flag of insecurity and serve her ads for makeup and cosmetic surgery within minutes. That’s not a bug in the algorithm. That’s the algorithm working exactly as commissioned, on a fifteen-year-old’s nervous system, for margin.

Or take Grok, currently being sued by the families of actual children because it has been used to generate sexualised images of minors, and whose owners, the absolute fuckers, reached for freedom of speech as their defence for a machine that manufactures child abuse material. I have run out of even sarcastic things to say about that one. It just sits in the chest like a stone.

None of that is a complaint about neural networks so much as about a philosophy, the one that got dressed up as innovation and let itself into government, insisting the market will sort it and if it doesn’t, well, that’s the cost of progress, a cost that never seems to land on whoever’s making the argument. We built an economy that treats a teenage girl’s self-loathing as a monetisable signal and a poor town’s water table as an acceptable write-off, and then act shocked when the technology built inside that economy behaves exactly like it.

The technology isn’t the disease. It’s the needle that finally shows you where the blockage always was: a society that decided profit outranks people, run by governments that quietly stopped representing anyone who wasn’t in the room when the deal got made.

Fix that, and I’ll happily sit down and argue with you about the actual merits of the algorithm. Until then, don’t blame the internet that got me into Cambridge. Blame the men currently deciding whose water gets to run dry.

yours in rage, Vanessa

Watch and thank me later.

https://www.instagram.com/reel/DchPqwovZfJ/

The first batch worked. Genuinely beautiful, clean, teachable images, done in an afternoon, and I had that specific smugness of someone who thinks they’ve cracked a system that other people are still struggling with. Then I asked for hands. And the system fell apart in the exact way every account of AI image generation warns you it will, except knowing the warning in advance does nothing for the feeling of watching it happen to you personally, at 11pm, over a slide about the Large Intestine channel.

I used to code for a living. JP Morgan, Goldman, the whole grim cathedral of it. I spent years of my actual life debugging: staring at a screen until a semicolon revealed itself as the culprit, feeling that particular ache behind the eyes that comes from holding a logic tree in your head for six hours straight. I thought I’d left that behind when I picked up needles instead of a keyboard. Turns out I just moved the debugging into a different room.

Now I debug prompts. I add “anatomically correct hands, five fingers” like an incantation. I try “hands resting naturally, not gesturing.” I try describing the hand in isolation as though the model has never seen one and I am teaching it from scratch, the way I’d teach a first-year student the difference between LI-4 and LI-3. None of it sticks reliably. The model drifts back to its own instincts within two or three regenerations, the way a patient’s shoulder drifts back to the same brace pattern no matter how many times you release it. You’re not fixing the thing. You’re negotiating with it, badly, on a short lease.

And here’s the bit that actually stings: I’m not sure this is saving me time. I costed it out in my head the way I used to cost out a trade, and the honest answer is that the hours spent wrestling a machine into producing a competent hand are not meaningfully fewer than the hours it would take an illustrator, except an illustrator costs money I don’t currently have earmarked for six-fingered problems, and the machine costs only my evening and my patience, both of which I apparently value at zero.

I can’t draw. I want to be clear that this isn’t false modesty, I have tried, my anatomical sketches look like something a Reiki practitioner produced during a fever. So the choice isn’t AI versus my own competent hand-drawn diagrams. It’s AI versus not having the slides at all, or AI versus quietly, resentfully paying someone I can’t afford. Which means I keep going back to the thing that keeps failing me, the way you keep going back to a stubborn point that isn’t releasing, because the alternative is admitting you don’t have another tool for the job.

There’s a version of this essay that ends with a tidy lesson about patience or about the limits of the technology, and I don’t have one. I have a folder of hands with six fingers and a teaching deadline and the low-grade despair of someone who used to get paid extremely well to solve exactly this kind of problem and now can’t make a computer draw a freaking thumb.

— Vanessa

Still one finger over budget.

A colleague of mine, an academic whose entire scholarly identity is built on refusing to let Chinese medicine get flattened into “science” for the comfort of white men in tweed, recently discovered he had written a book.

Read more

The doctor on her phone does not exist. Neither does the office behind him, and the framed diploma on the wall was generated in the same thirty seconds as his sympathetic face and his warm, unhurried voice. The moringa is real. So is the salmonella, which the FDA and the CDC have since traced.

The New York Times went looking and found hundreds of them. AI avatars dressed as physicians and, more to the point for us, as Eastern medicine gurus: fabricated credentials, staged offices, that specific soft-focus authority we have spent centuries cultivating and they have learned to print on demand. They sell unproven supplements in miracle-cure language, and they aim, deliberately, at older women with chronic conditions. The people medicine is worst at. The people most likely to have been dismissed, mislabelled, and sent home with an antidepressant and a shrug.

Someone always says the technology is the problem.

The technology is the accelerant, not the arsonist, and the grift is older than the transistor. Wellness has always sold counterfeit authority to frightened people, and it has always known that a woman in perimenopause with joints on fire and a rheumatologist who has given up is the easiest mark there is. What changed is the cost. It used to take a charlatan with charisma, a rented room and a laminator. Now it takes a prompt. When the marginal cost of a fake expert drops to zero you get an infinite supply of them, operated by anonymous LLCs against which there is currently no law, because it is not yet illegal to invent a person and have them practise medicine at you.

And here is the part we would rather skip. We are soft to this. Not because we are stupid, but because some of us already trade in the currency they are counterfeiting. We explain cupping marks as “releasing toxins.” We let tone stand in for evidence, because tone is what a frightened nervous system reaches for first. We taught the audience to trust the soothing voice over the true one, and then acted surprised when someone built a machine that does soothing at scale and means none of it. One audit found a hundred and five of these products sold with less interaction and overdose information than a packet of fucking paracetamol. Suggestive names. Discount codes. The full apparatus of care with the care removed.

The woman with the autoimmune condition got worse. That is not an abstraction and it is not her fault. She did the thing all of us do when we are frightened and unwell: she believed someone who sounded like they knew.

The only defence any of us has left is the boring one. A real name. A real registration number. A body in a real room that can be complained about, struck off, sued, made to answer. The unglamorous, auditable, embodied kind of authority that a generated guru will never have to carry, because you cannot subpoena a prompt.

Keep your credentials real. It is, increasingly, the whole job.

— Vanessa

Real hands, real number, verifiably salmonella-free

“We don’t want it to be known that we are working on this.”

The “this” is Project Panama. That is the actual internal name Anthropic gave the operation, which is already funnier and grimmer than anything I could invent. From early 2024 until the filings were unsealed this …

Read more

The model is called Inkling. It comes from Thinking Machines, the lab run by Mira Murati, who used to be the Chief Technology Officer at OpenAI and who raised two billion dollars before shipping anything, which tells you roughly where we are as a civilisation. The model underperforms. Nobody disputes this. Sukhareva’s argument is that the model was never the product.

I know. For most of you this already reads like a weather report in a language you don’t speak. Stay with me, because the actual idea underneath is more relevant to a solo practitioner than nine tenths of the AI noise you’re being asked to have opinions about. To get there I have to explain three things: what a model’s “weights” are, how these things learn, and why teaching them anything new tends to make them stupider. I’ll be quick.

Read more

Apple accused OpenAI of running a small internal spy ring to steal hardware secrets. The US military told its …

Read more

Here is the Neijing as a nursery. Pastel. Rounded. Little smiling ghosts drifting past the channel system. The Crossbow Mechanism, Ji, that subtle trigger moment the classics …

Read more

I want it to know me without selling me moustache removal cream for perimenopausal women.

That is apparently a difficult bar to clear.

What I mean is this: I want the version of AI that knows my paren…

Read more

The heatwave arrived with the particular smugness of a weather event that knows it doesn’t belong here. Thirty-five degrees i…

Read more

Good for them. Genuinely. Ethical frameworks are useful things. I have several myself.

What they also have, presumably, is a smartphone, a bank account, and a Netflix subscription. Which means this conversation is going to get uncomfortable quite fast.

Let’s do the bit…

Read more

Not the same words, exactly. But the same architecture. Recipient address. Covelo logo somewhere in the right corner. Date. Re: line. Dear So-and-So. A body that manages to sound warm without being gushing and clinical without being cold. Yours sincerely. Credentials. BAcC number. Footer with the address that I always have to look up even though I’ve had the same postcode since 2019.

I write letters to GPs, to consultants, to physios, to patients, to my landlady when the boiler is broken. Each one involves the same fifteen minutes of assembling pieces I already know into a format I’ve already established, while a small voice in my head says surely you could have automated this by now.

Reader, I could.

Read more

SHEN 🫀: heart era. this is your main character energy. your presence. when someone walks in a room and you’re like “oh they’re HERE” that’s good shen. bad shen is doom-scrolling at 2am not re…

Read more

Read more

AI has developed the same unfortunate quality as TCM jargon: people deploy the terms freely, assume everyone else is keeping up, and quietly terrify anyone who missed the memo. The difference is that “Liver Qi stagnation” took two thousand years to calcify. AI managed it in about eighteen months.

Here is the glossary. Ordered from “I basically know this” to “I have been nodding at this word for six months and something in me has died.”

LLM (Large Language Model)

What people mean when they say “AI.” ChatGPT, Claude, Gemini. All LLMs.

The crucial thing to understand: they do not look things up. They generate. Trained on vast quantities of text, they learned to predict what comes next. Word by word. At extraordinary speed and scale.

Think of it as autocomplete that went to university, read everything, and now has opinions. This is impressive. It is also the source of most of the problems.

Prompt

The thing you type. Your question, your instruction, your task.

Most people treat a prompt like a wish. One vague sentence, dispatched hopefully, followed by surprise when the output is also vague. Then they blame the AI. The AI, for its part, answered exactly what you asked. It just wasn’t what you meant.

A good prompt is closer to a clinical intake form: who is this for, what do I actually want, in what format, at what length, what should it definitely not do. The model will always try to answer something. If you don’t specify what, it decides. This is not always ideal.

Read more

The heart that makes you a good acupuncturist is the same heart that charges less than your physio colleague, stays thirty minutes late without mentioning it, waives the cancellation fee for the person who cried last session, and quietly swallows the cost of every postgraduate course you’ve taken since graduating. Facial acupuncture diploma. Obstetrics module. That Japanese needling weekend in Stroud that cost £600 and a Saturday. All of it folded into the same hourly rate you set in 2019 because you didn’t want to seem greedy.

You are not greedy. You are underpaid and calling it vocation.

In TCM we talk about Qi stagnation as the root of so much suffering. Constraint. Things not moving that should move. The textbook says: abundance is unstuck Qi. It says this about the body, but it applies just as neatly to a business run by someone who went into medicine to avoid being a businessperson.

The good news: you already have everything you need to unstick it. You just haven’t looked at it all in one place.

What I’m suggesting is uncomfortable and takes about an afternoon.

Read more

Also on the shelf: hundreds of foil containers, four first aid kits, and the quiet dignity of researchers who thought this …

Read more

There is a cheaper way.

Claude Cowork is a desktop tool that operates your computer on your behalf: opens files, writes code, creates things. I used it to build a multi-page clinic website on a Sunday morning. It cost nothing beyond the Claude subscription I was already paying for.

Here is exactly how.

Read more