When the Accountant Dims the Brain

There is a distinct, quiet satisfaction in adjusting a dial. Whether it is the brass knob of a Victorian radiator, which responds to one’s efforts with a series of ominous clanks and a faint smell of damp rust, or the volume control on a wireless set, the act of turning something up or down satisfies a deep-seated human desire for control. It is, therefore, with a sense of profound historical symmetry that we must welcome the latest development from the laboratories of Anthropic. In releasing their new Claude Opus 5 model, they have introduced what they call an "effort" setting. This is, to put it in terms that would make a nineteenth-century clerk feel entirely at home, a dimmer switch for the machine’s brain.

For some years now, the pioneers of Silicon Valley have been engaged in a frantic, rather noisy race to build a digital deity. They have spent billions of dollars in the hope of constructing an intellect so vast, so all-encompassing, that it might finally explain why we cannot find our car keys or why the post office is always closed when we arrive. Yet, in a turn of events that will surprise absolutely no one who has ever had to explain a quarterly budget to a board of directors, the grand quest for infinite wisdom has run headfirst into the accounts department. It turns out that omniscience is terribly expensive to run.

The "effort" setting is the accountant’s revenge. It allows a company to decide, on any given Tuesday, exactly how intelligent they can afford their software to be. If the firm’s shares are performing well, the automated customer service agent may be permitted to understand subtle irony and the finer points of contract law. Should the market take a turn for the worse, however, the finance director can simply slide the fader to the left. The agent is instantly demoted to a state of polite, slightly foggy confusion, capable of processing refunds but entirely unable to comprehend why a customer might be upset about a missing toaster.

This is a magnificent piece of bureaucratic pragmatism. In the past, if an employer wished to reduce the intellectual calibre of their staff to save money, they had to engage in the tedious business of redundancy consultations, union negotiations, and the hiring of enthusiastic but ultimately bewildered teenagers. Now, one simply adjusts the slider. One can almost picture the scene in a modern open-plan office: a manager, noticing a slight dip in the afternoon’s productivity, walks over to the server cabinet and turns the intelligence dial down to "mildly distracted" to save three-tenths of a penny per transaction.

I am reminded of a brief, rather damp afternoon I spent some years ago in the archives of a long-defunct railway company in Shropshire. Among the ledger books, I found a memorandum from 1874 in which the chief clerk instructed his subordinates to "think with less vigour during the winter months," as the cost of the candles required to illuminate their desks during deep contemplation was severely impacting the company’s coal budget. The clerks, it seems, complied by staring out of the window at the rain and thinking exclusively of turnips, which required no illumination at all. Anthropic has merely digitised this venerable British tradition.

The commercial reality of the matter is, of course, entirely serious. We are told that early testers of the new model have embraced this dial-a-brain philosophy with immense gratitude. A legal technology firm reported that the model achieved its tasks while generating twenty-six percent fewer tokens—the digital equivalent of a barrister agreeing to use fewer adjectives in exchange for a quicker lunch. A financial research laboratory noted that the machine completed its models with one-third fewer turns, representing a significant saving in both time and electricity. We are no longer asking whether the machine can think; we are asking how cheaply it can do so before the client notices.

This brings us to the other remarkable feature of the new model: its capacity for self-verification. Opus 5, we are informed, does not merely produce an answer and hope for the best; it builds its own test harnesses and computer vision pipelines to check its own work. It is a machine that argues with itself until it is satisfied. This is a deeply moving, if slightly tragic, image. It conjures up the picture of a solitary clerk who must not only write the ledger but also stand behind himself, wearing a slightly different hat, to audit the figures, before finally shaking his own hand and declaring the job well done.

I am reminded of a retired schoolmaster I once knew in Somerset, who spent his autumn years playing chess against himself in the garden. When asked why he did not seek out a human opponent, he replied that he preferred his current partner because they both agreed on the definition of a knight's move, and neither of them ever complained about the quality of the tea.

There is a surreal elegance to this arrangement. In one test case, the model was asked to reconstruct a three-dimensional CAD model from a drawing it was intentionally prevented from seeing. Rather than admitting defeat—which would be the sensible, human thing to do—the model wrote its own computer vision code to extract the geometry from raw pixels, checked its own work, and completed the task. It is the digital equivalent of a blindfolded carpenter who, upon being asked to build a wardrobe, invents a new type of sonar, maps the room, cuts the timber, and then writes a glowing review of his own craftsmanship in the local parish magazine.

One wonders where this trajectory will lead. If we can adjust the intelligence of our machines, surely other qualities cannot be far behind. We may soon see a "politeness" slider, allowing us to dial down the machine’s manners during a particularly tense negotiation, or an "existential dread" dial, which could be turned up slightly to ensure the code it writes has a proper sense of gravity. For now, however, we must content ourselves with the knowledge that the future of artificial intelligence is not a grand, unstoppable march toward the stars, but a sensible, cost-controlled stroll through the suburbs, with one eye firmly on the utility bill.