The AGORÀ Briefing — Friday, August 21, 2026

Thanks, Adam. On September 24, Neko Health opens its first American office, in SoHo, at 300 Lafayette Street. Behind it lies the meeting of Daniel Ek, co-founder and CEO of Spotify, and Hjalmar Nilsonne, today the company's CEO. According to TechCrunch, more than 25,000 New Yorkers already sit on a waiting list. Last month the company closed a Series C round of 700 million dollars, with 100,000 people already scanned. Its sites remain in the United Kingdom and Sweden. The model joins three sources: the body scan, blood tests, and data from fitness devices. A single scan stays replicable. The orchestration of the data creates the lock-in. Here the real value is born.

Exactly, Adam. Three estimates, one direction. Goldman Sachs, in a 2023 paper, projects a productivity gain of 1.5 percentage points over the decade. McKinsey reaches as high as 3.4 points by 2040. Researchers at MIT stay cautious: 0.53 percent by 2034, as CNBC reports. The distance between 0.53 percent and 3.4 percent is the space where the leadership challenge lives. That gap depends on the conditions organizations build, ahead of the tools they purchase. The boom, since the launch of ChatGPT in November 2022, is almost four years old. It remains largely a phase of hardware and infrastructure. Ben Snider, chief U.S. equity strategist at Goldman, writes it plainly: the real gain comes from the conditions, ahead of the tools.

Thanks, Adam, and yes, against the current. The consensus treats compute as the ultimate moat of AI. It has the wrong frame. Compute will be priced as a commodity by 2027. The value migrates from the layer that produces tokens to the layer that knows how they get used. The AI Observatory at the Stanford STAIR Lab shows it: it aggregated seven datasets of real conversations, gathered with user consent between 2023 and 2025. And there is the financial proof. The Commodity Futures Trading Commission opened a 60-day consultation on August 19, 2026, on contracts tied to computing power. CME Group and Intercontinental Exchange intend to launch futures on compute. The two platforms dominate global derivatives. High confidence on the technology, medium on the timing.

With pleasure, Adam. In 1956, during the Suez crisis, Washington refused to support the pound. London gave way in days. Whoever controls the reserve currency dictates the terms. The pound handed primacy to the dollar. The end of British hegemony. Power transition theory, formulated by A.F.K. Organski in 1958, describes the risk: the danger grows as the rising power approaches parity with the hegemon. Today the gap between Washington and Beijing compresses along three axes: manufacturing, energy, semiconductors. China today surpasses the United States in manufacturing value added, a reversal that matured since 2000. A power transition arrives as slow erosion, then as sudden rupture.

Exactly, Adam, and the crux is scale. On August 18, 2026, NeuroAbs was posted on arXiv, a neuro-symbolic framework for RTL abstraction applied to formal verification, accepted at ICCAD 2026. Formal checking guarantees the functional correctness of hardware designs. Every new generation of chips multiplies the reachable states, and exhaustive model checking explodes in time and memory. Prior methods asked for heavy manual work or rigid rules. NeuroAbs cuts the state space while preserving the relevant properties. Here the quality of the abstraction decides everything: a proof that closes in minutes versus one that runs for days. And a proof that stays open blocks the release of the design.

Gladly, Adam. In August 2026, the researcher Dmitry V. Alexandrov published on arXiv the first mechanized formalization of Romanov's Triplet Logic, built with the proof assistant Rocq. The choice comes right away: ahead of building an executable tool, prove every mathematical property of the system. The result carries a label rare in production software: verified. The Triplet Logic, shortened to TLS, is a combinatorial framework founded on triplets. The operating engine is an intersection procedure, the Simple Vertex Intersection. Alexandrov formalized the core inside Rocq: the Compact Triplets Formulas, the hyperstructures, the clearing procedure, and the SVI itself. First the theory proven, then the tool. The sequence is reversed with method.

Great question, Adam. On March 2, 2026, "Conformal Policy Control" landed on arXiv, updated on August 19, 2026, and presented at ICML 2026. The idea: a safe reference policy operates as a probabilistic regulator toward an optimized policy still awaiting testing. A regulator, here, is a device that limits and modulates the action of another system. Conformal calibration defines how aggressively the new policy can act. Imitation of past behavior guarantees safety. An excess of caution discourages exploration and slows improvement. The question is explicit: how much change becomes excessive? The answer comes from statistical calibration, computed on the data of the safe policy. The framework removes two assumptions of conservative optimization.

Thanks, Adam, and I bring the data. A work accepted at ICDM 2026 focuses on a calibration problem in anomaly detection on logs. The title says it all: "Too Sure to Be Safe." The authors are three: Bin Li, Dongdong Wang, and Siyang Lu. Posted on August 18, 2026. Detectors based on language models reach high performance. Their confidence estimates stay poorly calibrated. These are two distinct things: how accurate a model is, and how reliable its estimate of certainty is. A model can classify correctly and miss the estimate of its own certainty. Confusing them is the risk. It holds for a log detector, and it holds for every system to which we entrust sensitive data, the body included.