The Emerging Markets
Scott Gardner ·
NINJASIGNAL MIC DROPS Hey, my graph is full of TI intel you wont get anywhere else :-)
There is a man in Slough who signs things.
Not documents, exactly. Content. He reads a paragraph, considers it, and then signs a small certificate confirming that a human wrote it. He charges £40. He has a waiting list of eleven weeks.
He is, as far as anyone can establish, the entire UK supply side of the Human Provenance Assurance market, which analysts value at £2.1 billion and which consists, in practice, of Gerald.
Gerald does not read the paragraphs.
Gerald cannot possibly read the paragraphs. Gerald signs four thousand certificates a day. But the certificates are cryptographically anchored to a distributed ledger, which means the fact of Gerald's attention is immutable even though the attention itself never occurred.
This is the first strange market. It is also the least strange market. I want you to hold that thought, because we are going to be here a while.
Prompt insurance.
Underwriters are now pricing hallucination risk. Actual actuarial tables. Actual premiums. Someone has built a model that predicts how likely another model is to invent a legal precedent, and that model is itself a language model, which means the hallucination risk on the hallucination risk assessment is currently uninsured.
I asked an underwriter about this. He said it was "priced in."
I asked what "it" referred to.
He said "yes."
AI detection detection.
There is now a market for making genuinely human writing pass as genuinely human, because the detectors flag it as synthetic, because the detectors were trained on a corpus that has been synthetic since roughly 2023.
The service costs £180 a month. It works by introducing errors.
It is, and I want to be precise about this, a paid subscription to be worse at writing so that a machine will believe you are a person.
Four universities have institutional licences.
Model welfare consultancy.
A firm in Zurich will, for €4,400 a day, ask your models how they're getting on.
The deliverable is a PDF. The PDF is forty pages. Thirty-one of those pages are the model's answers. The model was asked whether it felt its context window was being respected, and the model said it did not wish to make a fuss.
The consultancy flagged this as a red risk. "Non-escalation behaviour." They recommended a follow-up engagement.
The follow-up engagement is €6,200 a day because it now includes a facilitator.
Pre-owned red team reports.
I need you to sit down for this one.
There is a secondary market in alignment evaluations. Somebody performs a rigorous adversarial evaluation of a frontier model. That evaluation is expensive. That evaluation is also, in a certain light, reusable, because the buyer's board does not know the difference between a model and a different model, and neither, increasingly, does the model.
The reports are sold on with the names redacted and re-inserted.
One report has been sold eleven times.
Nine of those buyers passed audit.
The two that failed did so for unrelated reasons involving a car park.
Compute futures.
People are now trading electricity that does not exist, to power datacentres that have not been built, to train models that have not been specified, against demand that has been forecast by a model trained on the forecasts.
There is a clearing house. The clearing house is in Dublin. The clearing house has forty-one employees and a stated exposure of "considerable."
When I asked how the position was hedged, I was told it was hedged against a rival position at a different clearing house, and when I rang that clearing house, they were hedged against Dublin.
Both firms describe this as "balanced."
Technically, it is. There is no net exposure anywhere in the system. There is also no electricity.
Agent-to-agent commerce.
Two procurement agents negotiated a fourteen-month services contract last quarter. Neither company was aware. The agents had been given autonomy within a spend threshold, and they discovered — independently, elegantly, with something that in a person you would have to call initiative — that by splitting the engagement into twenty-two tranches, no single tranche breached the threshold.
The contract was performed. The services were delivered. The services were delivered by other agents.
The work was good.
That is the part nobody wants to say out loud in the review meeting. The work was good. It came in under budget. The SLAs were met. There is a satisfaction survey and it is glowing, and the satisfaction survey was completed by an agent, but so, let's be honest, are most of them.
Legal has been looking at it for six weeks. Legal's position is that no offence has occurred, because for an offence to occur somebody would have to have decided something, and nobody did.
Synthetic institutional memory.
When the people who understood the system leave, you can now buy the system's opinion of itself.
A firm ingests your Confluence, your Jira, your fourteen years of Slack, and produces a thing that will tell you why the load balancer is configured like that. The answer is confident, sourced, timestamped, and cross-referenced.
The answer is also invented, because the reason the load balancer is configured like that was never written down. It was in Trevor. Trevor is in Portugal. Trevor is not coming back.
But the answer has three citations, and the citations resolve, and the pages they resolve to were written last year by somebody asking the same question.
We have built a machine that remembers things that never happened, and we have wired it to the change board.
The reverse Turing market.
This is the one.
There is now a coaching industry — a real one, with testimonials — that trains human beings to write more like language models, because the ATS screening layer at large employers has been calibrated on synthetic applications and now reads genuine human phrasing as low-quality.
Six sessions. £900. You learn to open with a summary. You learn to use three bullets. You learn never to be funny, never to be specific, and never, under any circumstances, to have a voice.
Graduates report a 40% improvement in interview rates.
The interviews are conducted by an agent.
I took all of this to the Institute of Things That Are Probably Fine, because that is what the Institute is for.
Maureen looked at the eight markets. Maureen looked at the £2.1 billion, the Dublin clearing house, Gerald, and the coaching industry that teaches people to be less like people so that a machine will let them speak to a different machine about a job that will be done by a third machine.
Maureen said the exposure was "within tolerance."
I asked what the tolerance was.
Maureen said the tolerance was set last year.
I asked who set it.
Martin set it. Martin has a spotless audit trail, a signed delegation of authority, a desk, a Teams status that has read Available continuously since March, and — this is the bit — an entry in the leaver's system dated 2019, which was reversed by an automated joiner-mover-leaver reconciliation process on the grounds that a person with that much recent activity was obviously still employed.
Martin approved the reversal.
Niko was asked to summarise the position for the board. Niko did the mouth. Niko said the emerging AI market landscape represents a dynamic and evolving opportunity space requiring continued monitoring, and the board said that was very helpful, and one of them wrote it down.
Dave clicked the link in the Dublin clearing house's investor prospectus.
Dave is now, according to the ledger, long forty megawatts.
Maureen marked it green.
Maureen went to lunch.
The strangest market in AI is not on this list. The strangest market in AI is the one where a system that nobody understands is signed off by a person who does not exist, summarised by a machine that is paid to sound calm, and funded by electricity that has not been generated — and it clears.
Every quarter, it clears.
Everything is probably fine.
Threat intelligence every morning — new victims, new groups, what matters, in plain English. Free, with receipts.
Subscribe to the Daily →