Bessent declines the shield, and Z.ai opens the harness

Bessent draws a responsibility line, and Washington's holdings come into view

Treasury Secretary Scott Bessent told CNBC on Monday that AI developers "need to take responsibility for themselves" rather than expect a federal "liability shield" — "It is humans who are responsible, not the AI" — and the signal is careful that this is Bessent's own stance, spoken from Treasury, with it unclear how far he speaks for the administration; hours later Trump posted that the Department of Justice "will rein things in if we have to" while continuing to downplay AI risk, a pairing the signal reads as enforcement discretion rather than rules, and harder to price than either. A day earlier the Washington Post had reported that the Trump family holds large AI-related investments — Dell, Micron and GE Vernova, per government filings cited in the story, which the signal notes it could not retrieve directly — a basket that sits at the physical layer of the stack, and one that hands the datacentre-approval fights a new argument unrelated to water or power bills. Bessent's other disclosure on CNBC was twelve hours on Sunday with Vice Premier He Lifeng ahead of this week's Trump–Xi summit, including the prospect of a channel for AI-related incidents — uncontrollable agents, non-state cyber actors, non-state bio actors — with the trade truce expiring on 10 November and the signal warning that incident channels between rivals have historically preceded controls rather than replaced them.

The capital stack thins at the top and thickens underneath

Bloomberg reports SoftBank, a BB+ issuer, seeking more than US$11 billion in a high-yield bond sale to refinance bridge loans taken on for part of its OpenAI commitment, alongside retail investors facing lockups in vehicles marketed as access to Anthropic and OpenAI upside; the signal is explicit that a BB+ issuer's natural market is high yield and that a refinancing does not by itself prove the late-cycle sequence is under way — what it shows is the size of the commitment relative to the balance sheet carrying it, with the retail vehicles the more telling half. John's counter, in conversation with Mark, is that unlike 1999 there are real revenues underneath; Mark's is that growth flattens as Chinese and open alternatives compete; John's position is that competition at the model layer does not touch demand for compute, and the signal's practical point is to separate lab-layer and infrastructure-layer exposure rather than treat "AI" as one factor. Monday's session supplied the infrastructure half of that picture: AMD closed above US$1 trillion for the first time after a 10% day capped a five-day, 25% run, with Lisa Su having guided to a doubling of data-centre sales in 2027 and Jensen Huang saying last week he expects to double chip shipments next year — two volume commitments made into a market that spent a month debating a capex slowdown, which the signal reads as a second-source premium now substantially priced, the asymmetry shifted to packaging, HBM, foundry capacity and power. In Australia, Telstra's pitch at Dreamforce was that the country has both renewable energy and water where US developers must often choose, set against a build of 162 operating data centres, about 130 more at planning, a Treasury estimate of $150 billion by 2030, Anthropic's lease at a proposed $32 billion Western Downs development whose first stage could be gas-powered under a state exemption, and organised opposition at Spotswood — Telstra being, as the signal notes, an interested connectivity-first party ("It is not our job to be building sovereign models"). The founders, meanwhile, keep leaving: David Swan's report anchors on Monarch's Sam Senior, told by executives at a large Australian utility that he seemed to have "a time machine for what's two or three years ahead" — one utility's anecdote, not a measured national lag, but if representative a diffusion gap that compounds, and one the signal turns into a hypothesis rather than a conclusion about "AI-exposed" companies in lagging markets.

Agents meet the platforms

Amazon has blocked Meta's Muse agent from shopping on amazon.com after asking Meta privately to keep it off and being refused, citing three objections — no disclosure, no self-identification when browsing, and an allegation that the agent captures and stores customer logins — while the same day Shopify's Tobias Lütke announced a Muse partnership for agentic checkout in Shopify stores; only one of Amazon's objections is a security claim, and the signal's point is that where a retailer lands on agentic commerce follows where its margin sits, ad-funded marketplaces on one side and transaction-funded stacks on the other, with no standard yet for agent identity or delegated authority. What running those agents looks like as infrastructure is the pitch of AX, a declarative control plane that exposes Task and Workspace as Kubernetes-style resources, sandboxes untrusted agent code under CPU and memory limits, and suspends and resumes tasks with filesystem state intact — "Agents are a new kind of workload. They are neither microservices nor batch jobs" — a rationale that is the vendor's own framing from its product page, though the signal finds it matches what practitioners describe. Note where MCP sits in that model: one declarable ingredient alongside Git repos and skills, not the platform, which is also the counter-case Maharshi Patel's essay sits against, in arguing that the Model Context Protocol solved a limitation later training runs removed — Mark's gloss is "a pre-watershed technology", outgrown rather than outcompeted, with the signal's warning that adoption, a Linux Foundation home and a vendor ecosystem can all expand at the moment the underlying need evaporates.

The harness is the moat, and Z.ai just gave one away

SpaceXAI's Grok 4.7, released Monday, lands at 46 on the Artificial Analysis Intelligence Index, 16th of 202 models, inching back past GLM-5.3 after index version 4.3.2 had dropped it behind GLM-5.3 and Qwen3.8; its US$2 input and US$6 output pricing sits at or below the Artificial Analysis median on both legs but remains a premium to GLM-5.3 at roughly $1.20/$4, and it consumed 240 million output tokens against a median of 94 million to complete the index — US$3.74 per index task, which the signal argues is the unit that moves gross margin rather than the sticker price. The release notes emphasise that the model was trained to "natively understand the Grok Bot harness", and on the same day Z.ai open-sourced ZCode, its coding-agent harness, already at 5,500 stars and 1,600 forks; Mark flagged it alongside DAPO from ByteDance and Tsinghua's SIA group, which is not new — its open release dates from March to May 2025 — but which the signal argues describes, together with ZCode, a different move from open weights: commoditising the production process, from RL infrastructure to the scaffolding that turns raw capability into completed work. The tell it names is whether a US lab responds by opening its own harness.

Who is left to check the work

The AI Evaluator Forum's letter of 18 September, with more than 100 signatories including Geoffrey Hinton, sets minimum conditions for third-party evaluators embedded inside frontier labs — meaningful independence, full editorial control, no payment contingent on findings, unfiltered access to boards, limits on NDAs, protection from retaliation — every clause of which, the signal observes, maps to a known failure of issuer-pays auditing; it lands the same week Bessent ruled out a liability shield, and the signal argues that makes third-party evaluation the most plausible route to a statutory safe harbour, converting advocacy into a draft standard the labs have a commercial reason to adopt. Po-Shen Loh, in a guest post on Terence Tao's blog, inverts the displacement story: the more advanced AI becomes, the more consequential untrusted decisions need watching, and the supply of trusted human verifiers runs short first — an argument, not evidence — and OpenAI's independent mathematics advisory group, announced two days later, "will not be responsible for advising us on how to pace our internal progress on mathematics", which the signal reads as a well-governed communications function rather than an answer to Loh's problem. At the coding desk, Iouri Khramtsov's practitioner post takes on the objection that agents trade quality for output by locating the fix in process: have the agent review the spec for gaps before code is written, hold unit-test coverage above 95%, and — on his self-reported account, not a controlled result — defects fall while output rises; the post does not engage with the studies reporting regressions, which leaves open whether those studies measured teams that added agents without adding layers.

Read all 15 signals →