The AI Illusion: The Fire Exit

A year ago I published a series of fifty articles called The AI Illusion, covering artificial intelligence across finance, work, warfare, governance and truth, and I closed it with a short story. This is the second series, and I am beginning it the same way, with a story. A great deal has changed in twelve months. Not the pace of the technology, though that is real enough, but the widening gap between what the technology can now do and the near-total absence of anyone in a position to govern it. Most governments and regulators are not slow off the mark. They are not yet at the mark. We should ask whether we have already built something that sits beyond democratic reach. The story is in four parts. It is called The Fire Exit, and it describes a danger that is in plain sight and that almost no one is discussing, which is the usual condition of such dangers until the moment they arrive. You might conclude from it that the sensible course is to sell everything and run for the hills. Very few people ever do. There is a particular comfort in staying with the crowd, and at present the crowd is dancing. The trouble is that it is all dancing on one side of the boat. Prologue MERIDIAN ASSET MANAGEMENT LLC RISK OVERSIGHT COMMITTEE MINUTES OF EXTRAORDINARY MEETING 14 April 2025, 09:00 Eastern Standard Time Board Room C, 47th Floor, 383 Madison Avenue, New York Present: Chairman, Risk Oversight Committee; Chief Risk Officer; Chief Investment Officer; Head of Equities; External Risk Adviser (non-voting); General Counsel (observer) Apologies: None The Chairman opened the meeting at 09:04 and noted that a quorum was present. 1. PURPOSE OF MEETING The Committee had been convened to consider a formal proposal from the Investment Committee to waive the Fund’s standard single-sector concentration limit of 15 percent for the purposes of artificial intelligence infrastructure exposure within the domestic equity portfolio. 2. PRESENTATION OF PROPOSAL The Chief Investment Officer presented the proposal. He noted that the Fund’s current AI infrastructure exposure stood at 14.7 percent of the domestic equity portfolio, approaching the limit established in the 2019 risk framework. He further noted that peer analysis indicated median competitor exposure in the sector at 18.3 percent, with the upper quartile at 24.1 percent. The proposal was to raise the permitted limit to 25 percent for a period of eighteen months, subject to quarterly review. The External Risk Adviser noted that the Fund’s largest individual position within the sector, held across multiple vehicles, represented 11.2 percent of the domestic equity portfolio. She asked whether a single-name limit should also be reviewed. The Chief Investment Officer noted that the existing single-name limit of 5 percent applied at the vehicle level rather than the consolidated fund level and that the current structure was therefore compliant. The General Counsel observed that the minutes should reflect that the waiver was being sought in response to competitive positioning analysis rather than a change in the underlying risk assessment of the sector. 3. DISCUSSION One member raised the observation that Berkshire Hathaway’s reported cash position of approximately 347 billion dollars might be considered a relevant external data point. The Chief Investment Officer noted that Berkshire operated under structural constraints not applicable to the Fund, and that the comparison was not directly useful. No further discussion was recorded on this point. The Committee considered the proposal. The risk-adjusted return profile of the sector over the preceding twenty-four months was noted. The competitive implications of underexposure relative to peers were discussed at length. The word structural appeared six times in the course of the discussion. The word risk appeared eleven times. 4. RESOLUTION The Committee resolved unanimously to approve the waiver as proposed, subject to quarterly review. The next review was scheduled for 14 July 2025. There being no further business, the Chairman closed the meeting at 09:47. These minutes were approved as a true and accurate record. [Signed] Chairman, Risk Oversight Committee, Meridian Asset Management LLC Document reference: ROM/2025/04/14/EOC RESTRICTED: For internal distribution only. Not for external release. I. The Paper Beijing, 10 September 2026 Dr Fang Wei arrived at the Ministry of Industry and Information Technology building on West Chang’an Avenue at ten minutes past two. He had arrived at two o’clock, as instructed, but the security desk had no record of his appointment and the responsible officer had needed to make three telephone calls before the matter was resolved. This was not unusual. He was shown to a conference room on the seventh floor and given tea in a glass. The room contained a projector screen, a table that could seat twenty, and four other people, none of whom Fang knew well. A deputy director whose name he did not immediately recall was arranging papers at the head of the table. On the screen was a logo: NovaMind, with the characters below it that Fang had always thought were slightly too formal for a commercial product. He had written three of the four foundational papers underlying NovaMind’s architecture between 2021 and 2024. The fourth had been written by a team of twelve and was, in his private assessment, considerably worse for it. He had not complained. This was not how things were done. He had not been told what this meeting was about. The deputy director began at two-fifteen. The North American commercial licensing programme, he explained, had been reviewed and approved at the relevant level. The pricing reflected NovaMind’s commitment to accessible, high-quality artificial intelligence services. The programme would be valid for an initial period of eighteen months. Fang noted the word introductory, turned it over once, and put it down. He thought about what eighteen months of introductory meant when it was backed by state subsidy and applied to a market whose entire pricing architecture assumed you could not compete. The deputy director paused in the way of people who expect acknowledgement rather than questions. Two of the four others nodded. One wrote something down. He had written a paper in 2023 on cascade failures in complex interconnected systems. It had been accepted by a journal, presented at a conference in Shenzhen, and subsequently ignored by everyone who mattered. The paper argued, using examples drawn from the 1987 equity market crash and the 2010 Flash Crash, that complex systems did not fail gradually: they absorbed pressure until a threshold was crossed, and then reorganised catastrophically. The threshold event, the paper noted, could not be identified by those inside it. It became visible only after the reorganisation was complete. The paper had included an appendix on external validation signals: observable data that suggested a system was under stress before that stress became visible in prices. The most recent example he had used, added in the final revision, was the cash position of Berkshire Hathaway. At the time of writing, Berkshire held approximately 280 billion dollars in cash and short-term Treasuries. It had been selling equities, including a significant position in Apple, for six consecutive quarters. Its chairman had declined, publicly and repeatedly, to find assets at prices he considered rational. Fang had used it as an illustration of how signals could be simultaneously visible and ignored. He had noted, with some care, that Berkshire’s cash position was not a secret. It appeared in every quarterly filing, was discussed in every financial publication of consequence, and had been attributed by several commentators to the chairman’s age or the fund’s unwieldy size. Nobody at the Ministry had read the paper, as far as Fang knew. The appendix on Berkshire had not therefore been discussed. He raised his hand slightly, not quite a full gesture. “The American AI infrastruc… Get full access to David McCreadie at davidmccreadie.substack.com/subscribe