Things to think about #13 - Is AI plateauing? and Monetising US hegemony
Adam Butler, head of ReSolve Asset Management, makes an interesting observation on AI in the wake of the publicised roll-out of ChatGPT 5. In effect, he argues that the AI cycle is over, for now.
The problem isn’t that the models stopped improving. It’s that the improvements we need are measured in orders of magnitude, not percentage points. Every step up the scaling laws now demands a city’s worth of electricity and a sovereign wealth fund’s worth of GPUs. You can still squeeze clever tricks out of mixture-of-experts or chain tiny specialists into something that looks like agency; that keeps the demo videos cinematic. It just doesn’t get us to super-intelligence. For that we need either an architectural miracle (unforecastable by definition) or a civil-engineering miracle (a decade-long sprint to build nuclear plants and 2-nanometer fabs). The first is luck. The second is politics. Both are scarce.
I am in no position to evaluate whether Adam is right that the laws of diminishing returns are now biting for AI, though the idea that AI demand for energy, in competition with demand in the wider economy, is now a binding constraint is a fascinating one. Adam’s point on the domain-specific risks of AI hallucinations also is on the money, I think.
Meanwhile the models we have remain, at their core, next-token roulette wheels. Chain enough spins together and tiny error probabilities compound into existential glitches. In domains where you can automatically verify an answer—unit tests pass, the protein binds—those glitches are an acceptable tax. You iterate until it works. In domains where judgment is qualitative, the tax becomes fatal.
I still think that the "natural language to code" element in LLMs is something that has huge potential if we consider the initial condition where everything "runs on excel and email". The disruption is here, but many domains won't be able to adopt current AI models given the still-significant risk of errors and hallucinations.
Monetising US hegemony
Many analysts have tried to come up with a coherent framework for Mr. Trump’s seemingly incoherent foreign and economic policy. Not surprisingly, Karthik Sankaran hits the bullseye with his idea that the US under president Trump is effectively trying to monetise its economic, political and military hegemony. He says:
“How should we understand this policy mix? Miran’s [the] goal might best be described as an attempt to monetize the US’s primacy. In his words, “The world can still have the American defense umbrella and trading system, but it’s got to start paying its fair share for them.” The desired outcome of his vision retains the United States’ position as global policeman and reserve provider, while making allied nations “participate in bearing the costs.”
But Washington’s current foreign policy trajectory does not seem to be in the business of making allies. Washington has imposed tariffs on a broad swathe of countries, and thrown the status of NATO into question—the preeminent commitment to a US-led international security order. Facing frank expressions of a starkly transactional and hierarchical view of diplomacy—as in the case of Trump’s statement that the US should sell allies aircrafts at a quality discount, in case they aren’t always allies—and extreme economic policy volatility, disincentives are mounting for countries tied financially to this security provider.”
Recent news that AMD and Nvidia, producers of the key semiconductors used to run everything from AI models to military equipment, have agreed to pay the US government 15% of their revenues to be granted export licenses to sale of their products to China is a vindication of Mr. Sankaran’s thesis. As Mike Bird, a journalist for The Economist, notes;
[This is] another fascinating instance in which Trump's China politics is actually much softer than the DC consensus. Paying a fairly modest markup on advanced chips is a great deal for Chinese buyers when most of the US political establishment wants to stop you buying them at all.
Perhaps it isn’t fascinating as much as it is entirely logical for Mr. Trump’s administration to settle on such an outcome.