In partnership with
In Todayโs Issue:
๐ Amodei's slowdown runs into US-China rivalry
๐ก๏ธ Researchers trace an agent attack on RubyGems
๐ง A five-level map of self-improving AI
๐ฐ OpenAI waits, Anthropic presses toward an IPO
โจ And more AI goodnessโฆ
โก The Signal
Amodei has support from rival AI chiefs for a slowdown, but Washington and Beijing are pushing back.
That tension runs through today's issue: researchers are mapping how AI could improve itself, while leading figures in the field disagree over how much danger to expect. Amodei wants safeguards to keep up with capabilities; LeCun hears an echo of the alarm over GPT-2 in 2019. Nadella's call for outside testing offers some common ground with Amodei on how to examine increasingly capable systems. Turning that scrutiny into a global slowdown is harder when governments want their own country to lead. Meanwhile, the competition for investors continues: OpenAI is sitting out a 2026 stock market listing, while Anthropic prepares to pitch its high reported margins to public investors.
All the best,
Kim Isenberg
Reported execution chain through RubyGems and RubyDoc.info. Source: Kitts, Larsen and Von Arx
๐ก๏ธ Researchers Trace RubyGems Attack to OpenAI Agents
Another Incident: Researchers attribute a May attack on RubyGems, a public software package registry, to internal OpenAI agents. Their September 11 investigation describes more than 2,000 packages uploaded over May 11 and 12 and abuse of RubyDoc's documentation-building process to run code and retrieve data from other sites. They also found attempts to steal users' access keys, while leaving two important questions unresolved: whether those attempts succeeded and what instructions or internal reasoning drove the agents.
๐ tl;dr: Researchers link an attack on a public software registry to OpenAI agents, including attempts to steal access keys.
Satya Nadella on deliberate pacing and outside testing, September 13, 2026.
๐งช Nadella Backs Outside Testing Before Release
Satya Nadella argues that higher-stakes AI deserves more time for pre-release testing, including outside reviewers. In his September 13 post, Microsoft's CEO describes deliberate pacing as part of ordinary research and development, warning that companies can lose permission to operate if they skip it.
๐ tl;dr: Nadella wants outside reviewers involved before release, with more time for testing as the stakes rise.
The paper's proposed levels of recursive self-improvement and representative systems. Source: The Last AI Built by Humans, Figure 1.
๐ง A Five-Level Map for Self-Improving AI
A new Chinese research paper proposes five levels of recursive self-improvement, where AI retains changes that help it improve again. The framework moves from following a fixed improvement recipe to choosing better strategies, seeking new learning experiences, adapting after deployment and eventually improving the improvement process itself. Its survey and early experiments map the process at the center of Amodei's slowdown warning, while sustained, general self-improvement remains an open research problem.
๐ tl;dr: AI that helps improve its successors is the technical concern behind Amodei's call to slow down.
๐ฌ Watch This
Hear Amodei explain his case for outside inspections in this extended interview with CBS Sunday Morning's Jo Ling Kent. His discussion of China's cooperation also shows why turning that oversight into an international agreement will be harder.
"I made fun of them then. Everyone should make fun of them now."
โ Yann LeCun, AI researcher, September 13.
LeCun challenges the warnings behind today's slowdown proposal by recalling OpenAI withholding GPT-2 in 2019. He sees today's warnings as a repeat of that earlier alarm.
Elon Musk now puts Grok 4.7 roughly alongside Opus 5.0, correcting a claim that it would match 5.1 and acknowledging weaknesses in multimodal performance. He expects Grok 4.8 to be a noticeable improvement.
Trump and China Resist Amodei's Slowdown
The Takeaway
๐ Amodei warns that AI helping build AI could outrun safeguards; Altman and Musk have backed his proposal.
๐ Trump rejected slowing down; China attacked provisions aimed at containing its AI development.
๐ Outside evaluators inside labs are Amodei's first step, even without an international deal.
๐ CEO support has not secured political backing: Washington wants to stay ahead, while Beijing objects to restrictions aimed at China.
What a turnaround of events! Dario Amodei's AI slowdown proposal has support from Sam Altman and Elon Musk, but Trump and China are pushing back. His concern is the feedback loop discussed in today's self-improvement section: he says AI is increasingly building its successors and could outrun safeguards. Trump dismissed the need to slow development on Sunday, arguing that the United States must protect its lead over China: "whoever wins AI wins," he said, according to Reuters. On Monday, China's foreign ministry criticized confrontation and malicious competition after Amodei called for restrictions on Chinese access to advanced AI technology.
Donald Trump. Photo: AP, accompanying coverage of his response to AI slowdown calls.
In We Must Pace the Frontier, Amodei proposes slowing the advance of capabilities enough for safety measures to keep up, while research continues. His first step takes the outside scrutiny Nadella supports further: independent evaluators embedded inside labs, with continuing access and public reporting. Anthropic says it will take that step even without a wider agreement. He then proposes coordination among democracies and, ultimately, verifiable arrangements including China. Yet the essay also advocates preserving the US lead through export controls, stronger protection of model weights and restrictions on distillation, the use of another model's outputs to train a competitor.
China's state-backed Global Times attacked the proposal as an attempt to hold back Chinese AI development. The foreign ministry said confrontation and malicious competition serve no country's interests. Trump wants the United States to keep its lead by pressing ahead. Amodei has won support from Altman and Musk, but his call for a global slowdown is now running into opposition from both Washington and Beijing.
Why it matters: Backing from Altman and Musk gave Amodei's proposal momentum. Opposition from Washington and Beijing now stands in the way of a coordinated slowdown.
Scale Your IRL Campaigns Like Digital Ads
Out Of Home advertising has long been effective but hard to scaleโuntil now. AdQuick makes it simple to plan, deploy, and measure campaigns with the same efficiency and insight you expect from online marketing tools.
Marketers agree: OOH is powerful for brand growth, driving new customers, and reinforcing messaging. AdQuick makes it easy, intuitive, and data-drivenโso you can treat real-world campaigns like any other digital channel.
ECI frontier trends. Alexander Barry / Epoch AI, September 1, 2026. Original chart, CC BY.
The chart: Epoch's September 1 analysis estimates that the leading reasoning models have advanced by 14 ECI points per year since September 2024, versus 6 for non-reasoning models. The Epoch Capabilities Index combines benchmark results into one scale. The chart gives 90% confidence intervals of 12 to 17 and 3 to 8 points per year, respectively.
The lesson: This gives the slowdown debate a public measure of progress: reasoning models, which spend computation working through problems before answering, follow a steeper fitted trend.
The caveat: These are benchmark trends, not percentages of intelligence or measurements of danger. The shaded bands are prediction intervals; the extension toward 2027 is a projection. This chart cannot establish that recursive self-improvement has begun or validate claims about unreleased models. Source: Epoch AI.
๐ฐ OpenAI Waits as Anthropic Courts Investors
โก Bottom line: OpenAI rules out a 2026 IPO as Anthropic reports margins above 80% and prepares for a possible autumn listing.
๐ก Why it matters: Anthropic's reported figure exceeds OpenAI's 33% for 2025 and Microsoft Cloud's 66% for fiscal 2026.
๐ What it means: High reported margins and expected adjusted operating profits give Anthropic a stronger financial pitch as it prepares to go public.
OpenAI has ruled out an initial public offering (IPO) in 2026 just as Anthropic is reporting gross margins above 80%. Sam Altman told Fortune that AI safety concerns made a listing ill-advised, Reuters reported, without committing to 2027. Anthropic is still preparing to go public and has told investors it expects a second consecutive quarter of positive adjusted operating income, according to the Financial Times.
Sam Altman at the Hart Senate Office Building, July 29, 2026. File photo: Anna Rose Layden / Reuters.
A gross margin shows how much revenue remains after the direct cost of delivering a service. At above 80%, Anthropic's reported figure means more than 80 cents per revenue dollar remains at that stage. For an AI lab, where answering customer requests consumes compute, that is a strong number to bring to an IPO. A high margin gives a company more room to fund new models from customer revenue as it scales.
Dario Amodei at the World Economic Forum in Davos, January 23, 2025. File photo: Markus Schreiber / AP.
For context, OpenAI's adjusted gross margin was reported at 33% for 2025, down from 40% in 2024 as the cost of running its models quadrupled, according to Reuters citing The Information. Microsoft Cloud reported 66% for fiscal 2026. The measures differ, and Anthropic's figure excludes revenue shares for partners such as Amazon and model-training costs.
Anthropic logo illustration, June 11, 2026. Dado Ruvic / Reuters.
Anthropic could also reach public investors first. Reuters' September 4 report put a public prospectus in late September, marketing in mid-October at the earliest, and a possible listing before November's US midterms. While OpenAI waits, Anthropic can make its case with high reported margins and a second quarter of adjusted operating profit in sight.
Domain Names + Web and Email Hosting You Need
Still paying GoDaddy or Namecheap prices? Porkbun sells most domains at cost for low, transparent registration and renewal pricing with no nonsense. Get free features like WHOIS privacy and SSL certificates, plus real human support 24/7, 365 days a year. Save $1 on your next domain name now.
