Moonshot AI releases Kimi K3, a 2.8-trillion-parameter model it calls the first 'open 3T-class' release; a new legal paper argues AI-agent personhood is beside the point and proposes a six-layer accountability stack instead; and São Paulo's facial-recognition network now watches an estimated 47% of Brazil's population

Marginal Revolution

Governing agentic AI

A new legal paper by Shruti Rajagopalan, flagged by Tyler Cowen, tackles a problem sharpened by autonomous coding and browsing agents: when an AI system transacts, publishes, or acts on external systems without a human approving each step, who is legally responsible when something goes wrong? The paper argues that proposals to grant AI agents legal personhood miss the point — for two millennia, nonhuman legal entities from Roman corporations to rivers have worked only because a human officeholder could always be located, questioned, and held accountable, and agentic AI breaks that pattern by acting without any identifiable responsibility-bearer. Testing existing liability doctrines (agency law, products liability, negligence, strict liability, enterprise liability) against three deployment scenarios, Rajagopalan finds they all fail once the developer and deployer become unclear or absent. Her proposed fix borrows from aviation, shipping, and autonomous-vehicle regulation: a six-layer stack — registration, identification, verification, financial responsibility, lifecycle traceability, and suspension — designed to keep a human accountable at the end of the chain no matter how autonomous the system in the middle becomes.

Construction Physics

How Predictable Are Laws?

Brian Potter used Claude Opus 4.8 to systematically estimate how often US federal laws end up doing what their authors expected, running the analysis across 239 laws sampled from 1976 through 2023. The AI was asked to research each law's contemporaneous predicted effects (bill text, signing statements, CBO scores, press coverage) versus its actual downstream effects, and to score the divergence on a -10 to +10 scale. The result: about 40% of laws landed within one point of 'exactly as expected,' and only around 11% diverged substantially in either direction — most laws, it turns out, do roughly what they say on the tin. The interesting cases are the outliers: a minor 1978 tax provision that became the 401(k), an obscure airline-insurance amendment that deregulated air cargo and enabled FedEx, and a customs-modernization act whose de-minimis tariff change inadvertently fueled the Shein/Temu import boom, versus the Alaska Natural Gas Transportation Act, whose planned 4,800-mile pipeline was never built. Potter frames laws like technologies: once a new capability exists, people find uses for it — good and bad — that its creators never anticipated, and that unpredictability appears to hold steady across five decades of legislation rather than improving or worsening over time.

Astral Codex Ten

Against "Stochastic Terrorism"

Scott Alexander goes after the concept of 'stochastic terrorism' — the idea that spreading fear or harsh criticism about a group makes you partly responsible when someone eventually commits violence against it — by cataloguing how selectively it gets applied: to nativist rhetoric before mosque shootings, to anti-Trump rhetoric before assassination attempts, to anti-insurer rhetoric before Luigi Mangione, to AI-safety warnings before a Molotov cocktail was thrown at Sam Altman's house, and more, arguing that almost nobody who invokes the concept is willing to apply it consistently to cases they don't already dislike. He works through several proposed limiting principles — ban harsh criticism, ban dehumanizing language, ban criticism of individuals but not policies, exempt powerful targets — and argues each either bans mainstream political speech everyone actually wants to keep, or fails to track the actual pattern of who ends up dead. His alternative is what he calls the 'liberal solution': any person or group can be criticized as harshly as you like, extra-state violence is never legitimate regardless of provocation, and responsibility for a violent act rests entirely with the person who commits it — a norm he acknowledges is fragile, given that a meaningful share of young Americans express favorable views of Mangione.

Marginal Revolution

Why crime will decline in (most of) Brazil

Brazilian cities are building out one of the world's largest facial-recognition surveillance networks, led by São Paulo's system, which already streams footage from street cameras, health centers, buses, and police motorbikes into a central control room and is set to double to 100,000 cameras by 2028. Watchdog group O Panóptico counts more than 560 active facial-recognition projects across 20-plus Brazilian states — police-run systems alongside experiments like school cameras used to take attendance — collectively watching an estimated 99 million people, or 47% of the country's population. The Economist piece, excerpted via Marginal Revolution, frames this as a genuine crime-reduction success story, with the technology helping locate wanted suspects, missing persons, and stolen vehicles, but the scale is a striking data point for anyone tracking how quickly biometric surveillance infrastructure can become normalized in a democracy once the crime-reduction case is made.