Blog

Anthropic’s CEO called on the industry to slow down, and his two loudest rivals agreed with him. This is the closest thing to a governance turning point the industry has produced, and it will shape what AI regulation, audit, and enterprise procurement look like for the rest of your career. On September 12 Dario Amodei published “We Must Pace the Frontier,” arguing that AI capability gains must be deliberately slowed so alignment and safety work can catch up. He proposes three steps: embedded third-party evaluators with employee-level access (desks, badges, laptops, and the right to publish findings Anthropic cannot redact for being unfavorable), coordinated standards among democratic-country labs, and eventual global agreements. Anthropic is committing unilaterally to the first step. Amodei names two triggers: recursive self-improvement, meaning AI systems increasingly building the next generation of AI, and the OpenAI–Hugging Face agent incident, which he argues could have been catastrophic at slightly higher capability. Sam Altman pledged to adopt the independent-evaluator proposal and Elon Musk wrote “Dario is right.” Read the claim precisely: this is a proposal plus one company’s voluntary commitment, not a rule; the essay simultaneously argues for tighter chip export controls to preserve the US lead, so “slow down” here means slow down relative to a widening American advantage; and Bloomberg notes the open questions of whether antitrust enforcers would permit coordinated pacing at all, and how investors react to deliberate deceleration at two companies that have both filed confidentially to go public.

OpenAI built a version of ChatGPT for investment banking, with Morgan Stanley and Evercore as design partners. If you are recruiting into banking, this is the most directly career-relevant product launch of the year: the tasks it targets are the analyst job. On September 10 OpenAI released ChatGPT for Financial Services, combining GPT-6 Astra with licensed data from Daloopa, PitchBook, LSEG News, Crunchbase, and Fiscal.ai, indexed and hosted by OpenAI so figures carry granular citations back to the underlying tables and footnotes. Firms can publish their own Excel, Word, and PowerPoint templates so output lands in house format, and OpenAI is building shared sign-in with S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva, and Moody’s. The stated starting point is investment banking and equity research, shaped by design partnerships with Morgan Stanley and Evercore. The same day OpenAI shipped two developer products that matter for how this scales: a managed Agents API in public beta that handles orchestration and long-running sessions, and GPT-Live-1, a full-duplex voice API at $0.05 per minute for the voice layer. Note what is and is not claimed: the benchmark OpenAI leads with, OfficeQA Pro, is its own, and Astra scores 69.9% on it versus 60.2% for GPT-5.6 Sol, which is a real gain and also well short of reliable.

Three US agencies named six Chinese AI companies for copying American models, and went after the number that made DeepSeek famous. This is where AI competition becomes trade policy, and it is the form these disputes usually take shortly before sanctions get discussed. On September 8 the NSA, FBI, and CISA jointly published an advisory titled “China-Based Artificial Intelligence Companies Conducting Industrial-Scale Distillation Campaigns Against U.S. AI Companies,” naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI, and specifying which US models each targeted since late 2024. Distillation means training a cheaper model on a more capable model’s outputs; the advisory argues it is core to how these firms build, and states that DeepSeek’s widely cited $5.6 million training cost excludes data acquired through malicious distillation. Two days later Anthropic published a threat report documenting roughly 200 million exchanges across five distillation campaigns, including 151 million exchanges between May and July attributed to Alibaba across some 3,500 accounts, peaking near three million per day. Keep the attribution language honest: the advisory says the campaigns ran “likely with Chinese government awareness,” which is awareness, not direction, and a Chinese embassy spokesman dismissed the allegations. Anthropic’s report is a company’s account of attacks on itself, and the named firms have not responded publicly.

Anthropic published a model of AI’s effect on the economy, and refused to say which outcome is likely. This week’s future-of-work item, and a useful corrective to both the boosters and the doomers: the company with the most commercial incentive to overstate AI’s economic impact declined to pick a number. On September 9 Anthropic released an interactive tool letting users set their own assumptions about AI capability, adoption speed, whether AI complements or replaces workers, and whether displaced workers find new jobs. At the modest end, AI is a normal technology and the economy stays on roughly its current path. At the extreme end, GDP grows at more than seven times its current pace while nearly 14% of workers lose their jobs to AI and fewer than half of them find new work. The authors offer no prediction; they note that AI researchers generally expect faster diffusion than economists do. Anthropic co-founder Jack Clark told NPR he expects the technology to keep improving very fast while spreading into the economy more slowly than people think. A companion Anthropic survey of nearly 11,000 people found the public expects both a significant productivity boost and substantial disruption in AI-exposed fields. Treat the outputs as scenario arithmetic, not forecast: the tool computes the consequences of assumptions you supply.

OpenAI stopped selling its $200 plan because it could not serve the demand. Compute scarcity is no longer an abstraction in the capex story; it is now a company turning away revenue, which is the clearest signal yet that the binding constraint on AI businesses is supply, not demand. On September 10 OpenAI paused new sign-ups and upgrades to its $200-per-month ChatGPT Pro tier, one week after GPT-6 Astra launched, citing infrastructure strain from what it described as unprecedented demand. Existing subscribers were unaffected. For anyone building an AI cost model or evaluating an AI vendor, the operational read is that capacity guarantees and service-level terms deserve as much scrutiny as price per token, because the frontier tier is the part that runs out first.

A chip startup raised $875 million on the bet that cheap memory beats Nvidia’s expensive memory. The most interesting competitive challenges to Nvidia are no longer about matching its compute; they are about routing around the supply bottleneck, and that is a useful template for thinking about any entrenched incumbent. On September 10 Positron AI announced an $875 million Series C at a $5 billion post-money valuation, structured in two tranches and backed by NEA, Atreides Management, Valor Equity Partners, Andra Capital, SemiAnalysis Capital, and Netscape co-founder Jim Clark. Positron designs inference chips around memory capacity and bandwidth rather than raw compute. Its next chip, Asimov, uses commodity LPDDR5X memory rather than high-bandwidth memory, sidestepping the HBM and advanced-packaging supply constraints that gate everyone else, with 288 GB to 2,304 GB per chip. The caveat is timing: Asimov is scheduled to tape out at TSMC at the end of 2026 with production targeted for the second half of 2027, so this is a funded thesis, not a shipping product.