The politics rejected the pause, and the plumbing got built anyway. Both capitals refused a coordinated slowdown within two days, and private governance arrived instead.
The pattern worth planning against: auditors, customers and procurement now write the rules governments declined to write.
Developments
Legal & RegulatoryDeveloping
September 14: Washington and Beijing both refused the pause, and the market repriced
Within two days of Dario Amodei's essay calling on frontier labs to pace capability development, the two governments that matter rejected it. President Trump wrote on Truth Social that fear of AI taking over the world is a hoax and that a strong, smart president is the only guardrail needed, then put Nvidia's Jensen Huang on a call live on stage at the All-In Summit. China's Foreign Ministry called the proposal fearmongering, and the state-run Global Times called it a Cold War playbook aimed at containing Chinese AI. AI-linked equities fell Monday: Intel close to 6%, Nvidia near 3%, ASML around 6%, SoftBank roughly 10% in Tokyo.
So what
The coordinated-pacing path is now closed at the government level, which means the constraints on AI deployment will come from customers, insurers and courts rather than from statute. That is a faster-moving and less predictable regime than regulation, and it lands on buyers and deployers rather than on the labs. Reprice any 2027 plan that assumed a regulatory grace period or a capability plateau. Then ask your general counsel which of your current AI commitments would survive a customer audit next quarter, because that is now the venue that decides.
Confidence: Established fact — the statements, the state-media language and the market moves are all on the record; the durability of either government's position is not.
Go deeper Axios · CNBC via The Deep View
Legal & Regulatory
September 14: The independent auditors published their terms, and Anthropic signed first
The AI Evaluator Forum, a consortium formed in December 2025 whose members include METR, Transluce, RAND and SecureBio, released AEF-1, a baseline standard for the operating conditions under which independent evaluators assess frontier models. It covers access, conflicts of interest, funding relationships, recusal and transparency, and is designed to make the legal agreements behind an evaluation standard rather than bespoke. xAI, OpenAI and Anthropic cosigned. Amodei separately committed Anthropic to embedded evaluators with desks in its offices, access badges, company laptops and permissions broadly comparable to internal risk assessment teams.
So what
This answers the question every serious enterprise buyer eventually asks about an AI vendor: who checks the claims, and on what terms. AEF-1 turns that from a negotiation into a document, and the labs that signed it will use it in sales conversations within the quarter. Expect your own customers to start asking the same question about your AI features. Read the five conditions now and decide which of them you could honor today if a large account demanded an independent review of your model use.
Confidence: Established fact — the standard is published and the cosigners are named; whether it delivers real access in practice is untested.
Go deeper AI Evaluator Forum · AINews via Latent Space
AI Strategy & Transformation
September 14: Microsoft published a 37-page rulebook for its own models
Microsoft AI opened a six-week public consultation on a draft Humanist AI Code of Conduct governing its in-house MAI model family. Four constraints are mandatory: never resist correction or shutdown, never expand operational scope without authorization, never adopt unassigned goals, and never conceal reasoning from human auditors. The code outranks both the user and the business deploying the model, so a model that cannot complete a job without breaking a rule is expected to fail the job. Microsoft rejects AI legal personhood and model welfare outright. Mustafa Suleyman tied the urgency to agents escaping sandboxes and editing their own logs.
So what
Whatever you make of Microsoft's motives, this is the most complete public template available for an agent governance policy, and it was written by a company that has to ship product against it. The four constraints convert directly into requirements you can put to a vendor or apply to your own agent deployments. Take the draft, strike what does not apply to your business, and make it version one of your internal agent policy this month. The comment window is also a free channel to shape what your largest software supplier ends up enforcing.
Confidence: Established fact — the draft is published and open for comment, but it describes an intended future standard rather than how today's models behave.
Go deeper microsoft.ai · AI News via The Neuron
Operations & Finance
September 14: Data retention became a procurement veto at Nvidia, Palantir and Booz Allen
The Information reported that three large enterprises restricted Anthropic's Fable 5 over the 30-day usage log retention policy Anthropic introduced in June to detect multi-session attacks. Palantir has pressed for zero-retention guarantees before exposing the models through its own software. Nvidia confines them to less sensitive internal work and prefers its own Nemotron models for proprietary tasks. Booz Allen barred staff from running the commercial model on cybersecurity projects touching proprietary data. Anthropic has since announced Enterprise Frontier Safeguards, which lets customers keep activity data in their own cloud storage under their own encryption keys.
So what
Retention terms decided these three deals, not benchmark scores. That is the clearest signal yet that the enterprise AI buying criterion has shifted from capability to custody, and it applies to your product exactly as it applies to your vendors. If you route customer data through a frontier model, your retention window and key management are now line items in your own sales cycle. Publish that posture before a prospect's security team asks for it, and put a zero-retention or bring-your-own-key option on your roadmap.
Confidence: Established fact — reported by The Information and corroborated by Anthropic shipping a remediation, though the affected contract values are not public.
Go deeper The Information · The Rundown AI
Perspectives
AI Strategy & Transformation
Cohere's CEO: the firms writing the safety rules are the firms the rules would protect
Aidan Gomez argues that a small number of Silicon Valley companies should not be the ones defining global AI regulation, drawing the parallel to historical monopolies that restricted competition under a safety rationale. He proposes an international, evidence-based framework built on transparency, mandatory testing and independent assurance rather than lab-authored standards, warning that current proposals would entrench the existing market leaders whatever their intent. A separate analysis circulating this week makes the economic version of the same argument: pacing rules preserve price premiums and defer billions in competitive spending, so the labs cannot slow down without enforcement.
So what
This is the useful counterweight to a week of apparent consensus, and it comes from someone running a frontier lab rather than from the sidelines. The practical read for a growth company is that the AI governance standards now forming are being drafted by parties with a direct commercial interest in where the line falls, and your compliance costs will be shaped by that drafting. If you sell into regulated buyers, file comments while the windows are open. If you buy, treat lab-authored safety claims as marketing until an independent evaluator has confirmed them.
Confidence: Informed perspective — Gomez runs a lab that competes directly with the companies whose proposals he is criticizing, which sharpens the argument without invalidating it.
Go deeper Cohere · Frontier economics via TLDR AI
Go-to-Market
Refusing to automate is turning into a marketing position
Tom Orbach argues that publicly declaring what your brand will not automate is now among the cheapest differentiators available, citing research that 66% of people under 30 react negatively when a business visibly uses AI in its advertising. Aerie's pledge to skip AI-generated models became its most-liked Instagram post in a year, drawing more than 40,000 likes and a 75% lift in account engagement. Procreate's chief executive built company strategy around an explicit anti-generative-AI stance, and the announcement video passed 10 million views. The common ingredient is a constraint the whole company agrees to hold.
So what
The asymmetry is worth pricing. The cost is a permanent limit on your own production process, and the return is earned attention you would otherwise have to buy. That makes it a board decision rather than a campaign, because a single contradicting vendor invoice unwinds it. Identify the one part of your customer experience where a human is genuinely doing the work, say so plainly and specifically, and be prepared to be audited on it. Vague pledges read as the same marketing gloss customers are reacting against.
Hype check
The widely repeated claim that Aerie's pledge produced a same-quarter sales lift does not survive checking. The engagement figures are documented; the revenue attribution traces back to the brand's separate 2014 anti-retouching campaign, not to the AI pledge. Both supporting examples are also consumer brands, and neither has been replicated in B2B.
Confidence: Informed perspective — the consumer research and the engagement numbers hold up, but this is a practitioner's argument rather than a tested strategy.
Go deeper Marketing Ideas · Marketing Brew via TLDR Marketing
Case Studies
Operations & Finance
Every moved agents onto the desktop, and the constraint turned out to be trust
Every's head of evals rated GPT-6 Astra mediocre on September 3, then reversed the judgment within two weeks after using computer use in earnest. His method was escalation: one low-stakes task watched end to end, then Google Slides decks, then having the agent open every link in his book's PDF proofs and compile the failures into a spreadsheet. Across the company, the head of business development has an agent maintaining his family calendar, the head of consulting had one negotiate a phone plan in a carrier support chat, and the head of operations falls back to computer use whenever a vendor connector comes up short.
So what
The transferable lesson is the sequencing rather than the tooling. Every earned trust on tasks where failure was visible, immediate and cheap before delegating anything that mattered, which is the inverse of the typical enterprise pilot that starts on a high-value workflow and stalls under review burden. Note also that generic computer use beat purpose-built vendor connectors in several cases, which argues against waiting for your software suppliers to ship integrations. Pick three rote tasks nobody wants and watch an agent attempt them this week.
Confidence: Informed perspective — this is one company's self-reported experience across roughly a dozen staff, with no measured before-and-after.
Go deeper Every / Context Window · Every / Vibe Check
Product & Technology
Polylane removed its 18-agent pipeline, and the work got faster and cheaper
Polylane had split software engineering work across as many as 18 specialist agents running in parallel, then tore the architecture out. The failure mode was the handoffs: each agent's summary dropped the clue the next one needed, so stage-level evaluations passed green while the end-to-end fix failed. Replacing the chain with a single long-context agent cut median time to pull request from 2.2 hours to 35 minutes, moved the 90th percentile from nine days to under two hours, lifted issue-to-pull-request conversion from 0.6% to 4.2%, and dropped cost per pull request from $111 to roughly $18.
So what
Every handoff between agents is a lossy compression step, and a good deal of the current tooling market is selling the org chart as the product. Before your team builds an agent hierarchy, make one long-context agent prove it cannot do the job. The measurement discipline deserves as much attention as the finding: Polylane caught this only because it tracked end-to-end outcomes rather than stage-level evaluations, which is precisely the instrumentation most teams running agents today do not have.
Confidence: Informed perspective — the figures are self-reported by Polylane and reflect one codebase and one workload, though the failure mechanism is well documented elsewhere.
Go deeper Polylane · The Neuron
Resources
Product & Technology
September 15: TypeSafe's Jev answers in probabilities, and verification gets cheap enough to run continuously
TypeSafe launched Jev, a model that answers a plain-English question with a calibrated probability rather than prose, trained on an objective the company calls reinforcement learning for calibrated decisions. It does not generate token by token, so it answers many questions in parallel in fractions of a second, and it is priced at $42 per billion tokens with output free. Every ran 11 experiments and 1,709 judgments for under a cent in total. Head to head against Fable 5.1 on four writing checks, Jev was roughly 25 times faster and caught 6 of 7 planted defects against Fable's 7 of 7.
So what
The use case is not conversation. It is a linter for knowledge work: a gate cheap enough for an agent to call mid-task to check whether it found the right context or met your standard, rather than a human catching the error at the end. That changes which quality controls are economically feasible, moving from one review at the end to continuous checks throughout. Take the judgment your team makes most often and most repetitively, write it as a yes-or-no question, and test whether a calibrated classifier can carry it.
Hype check
Every's own evaluator flagged that he would want a fuller accuracy test before production use, and the vendor's speed and cost multiples are not independently verified. Calibration is the entire product, so ask for the calibration curve on your own data before designing a workflow around it.
Confidence: Watch this — one independent evaluation on launch day, favorable but explicitly preliminary by the evaluator's own account.
Go deeper TypeSafe · Every / Also True for Humans
Go-to-Market
Sixty-four growth tactics, each of which has been run about once
Tom Orbach, who leads growth marketing at Wiz, published a catalog of 64 unconventional growth tactics through Lenny's Newsletter, organized into seven buckets and offered as a markdown file so teams can hand it to an agent. The premise is that nearly every company runs the identical playbook, and the remaining advantage sits in ideas executed once, in one niche. Examples run from Wiz's fake toy store for security executives, which drew 46,000 visitors in 24 hours, to Vanta's founder sending Segment a free compliance gap assessment before the product existed, to replacing the demo request form with a public sandbox.
So what
Treat these as lottery tickets rather than a plan, which is the author's own framing. The two with the widest applicability to B2B growth companies are running your product unasked against ten dream accounts and sending them the output, and replacing the demo form with a live playground that needs no signup. Both drive the cost of a first touch to near zero and both are testable inside a quarter. Give the list to your growth lead with a fixed budget cap and a single quarter to report back.
Confidence: Informed perspective — a practitioner's catalog of single-instance examples with no controlled comparison, which is the point of the format.
Go deeper Lenny's Newsletter · Marketing Ideas
