No Intention of Slowing Down: Washington, Beijing, and Berlin Reject Amodei's Call

Trump calls AI warnings a hoax: The only guardrail is 'a smart president'

U.S. President Donald Trump has rejected calls for stricter AI rules, stating that the only control or Guardrails that artificial intelligence needs is 'a STRONG AND SMART (High IQ!) PRESIDENT,' which the USA has in abundance. In the same post on Truth Social, he spoke of a 'SICK conspiracy against AI and data centers,' which only China is happy about. A second post followed within twelve hours: The fear of an AI that takes over the world and destroys humanity is a 'HOAX,' comparable to the allegations of Russian election interference and his own impeachment proceedings. He said that under his presidency, development would 'not be stopped by brilliantly led destructive forces.' 'WHOEVER WINS AI, WINS,' he wrote, adding that the U.S. is ahead of China and will stay that way.

The trigger was Anthropic CEO Dario Amodei, who had called for a slowdown in development and better regulation over the weekend. OpenAI CEO Sam Altman and Elon Musk joined in: Musk stated that Amodei was right, while Altman said that the standards are not at a point where the models' capabilities should be pushed much further. Trump targeted Amodei directly, who is 'now pretending to be a perfect little angel,' and pointed out that his administration already has 'enormous PUNITIVE and REGULATORY POWER over these companies.' He claimed they had 'stopped AI people from doing bad or potentially bad things,' without elaborating further; the White House did not initially respond to a request for comment. Trump’s former AI advisor, David Sacks, also publicly disagreed with Altman and Amodei.

In the accompanying debate, Richard Whittle, a professor of AI and Public Policy at the University of Salford, pointed to a more immediate risk: companies are already tying themselves to specific tools and would be vulnerable to price increases of five or ten times. Startups building on these costs would hardly survive, and those who have replaced staff with AI will pay the price regardless of how it develops. So far, we are mainly seeing a cost-cutting culture and few cases where additional value is created with AI. → BBC, Axios, Wall Street Journal

Synthszr Take: Two Truth Social posts within twelve hours, and in both, the entire security architecture consists of a single person: the incumbent himself. Governance that hinges on one president is, of course, nonsense. In practice, control means verifiable success and termination criteria plus someone who has the right to actually stop a project. Trump did, however, precisely name the instrument he means, namely the 'enormous PUNITIVE and REGULATORY POWER over these companies,' i.e., pressure on a case-by-case basis. The Pentagon has already demonstrated what this looks like with Anthropic’s supply chain classification and the halted Mythos release. A guardrail based on whims disciplines companies but does not protect people.

Beijing’s State Press Calls Amodei’s AI Essay a 'Cold War Playbook'

The state-affiliated Global Times has described Dario Amodei’s call to slow down AI development as a 'Cold War playbook' against China in an editorial. The Anthropic CEO’s essay may seem like a rational statement focused on security risks, but its goal is to slow China’s AI development through technological barriers and regulatory monopolies. The paper calls the move 'hypocritical and short-sighted' and accuses Washington of excluding China from global AI governance, which it says increases the cost of errors and the risk of a loss of control worldwide. The cause for this was Amodei’s text published on Saturday, 'We Must Pace the Frontier,' in which he, in addition to slowing the pace, also calls for stricter chip export controls on China and action against suspected Model Distillation by Chinese labs.

In the roughly 3,500-word essay, Amodei proposes a three-stage roadmap. The first stage: independent evaluators with permanent, 'employee-like' access to the models, whom he compares to food inspectors and to which Anthropic says it has unilaterally committed. This is to be followed by international coordination with other labs and governments, including authoritarian states. As a warning sign, he mentions that AI could be capable of controlling a swarm that takes over the internet within six to twelve months. In a television interview, he said a 'kill switch' could be a good idea, a complete ban would not be, and that the arms control agreements with the Soviet Union are a possible model. He described the question of whether the U.S. should slow down if China does not as the most difficult dilemma of his proposal.

Sam Altman and Elon Musk have publicly endorsed the appeal. Altman announced that OpenAI would give external evaluators the same access; Musk commented concisely, 'Dario is right.' The intensified debate was triggered, in part, by the resignation of Anthropic researcher Jacob Coxon, 27, who speaks of a significant probability of a catastrophic outcome and considers a slowdown sensible only if it is coordinated with China.

In Washington, the proposal is met with rejection. Trump spoke of 'very negative forces' stoking exaggerated fears and declared that the U.S. is leading the AI race and 'whoever wins AI, wins.' House Speaker Mike Johnson pointed to his bipartisan AI task force but rejected an emergency moratorium because China would otherwise overtake the U.S. AI advisor David Sacks countered the lab chiefs by saying that the easiest way not to build a superintelligence is to not build it.

Skepticism is also coming from the research community. Gary Marcus welcomes the transparency pledges but sees the phrase 'partner with government' as an attempt to replace real regulation; François Chollet and others suspect it is about pulling up the ladder behind them. There is also criticism of the evaluator METR, mentioned by Amodei, which is itself close to the Silicon Valley environment of the labs. The U.S. and China are set to discuss security risks at the model frontier in mid-September, and the topic could come up at the summit between Trump and Xi on September 24. → CBS News, Reuters, HuffPost, Marcus on AI, The Independent

Synthszr Take: The same essay that asks for discipline on speed also calls for stricter chip export controls and tougher action against distillation by Chinese labs. So Beijing doesn’t need to make any assumptions for its interpretation, just this paragraph: A speed limit proposed by the one in the lead, flanked by access restrictions, reads to the pursuer as a way of cementing the gap. The fact that Amodei himself identifies this dilemma as the toughest point of his proposal does him little good as long as the conditions for cooperation are set unilaterally. Two weeks before the meeting on September 24, this gives China convenient leverage: It can portray itself as an advocate for inclusive AI governance without accepting a single limit itself. The price of this double message is that the security issue is treated as a trade issue in Beijing, and trade issues are resolved with concessions.

Berlin Aligned with Trump and Xi: Development Halt Not an Option for Europe

Germany’s Digital Ministry has explicitly rejected the call from the AI industry to slow the pace of development. 'We do not consider a development halt to be a viable option for Europe,' a ministry spokesperson said on Monday; strengthening digital sovereignty, on the contrary, means more development and more innovation in the field of artificial intelligence. This was in response to an essay over the weekend by Anthropic CEO Dario Amodei, in which he calls on AI companies to reduce the pace of their projects and proposes stricter controls, including independent auditors sitting directly in the labs.

The ministry stated that it is closely monitoring global developments and takes the public warnings from leading developers seriously. The spokesperson drew one specific line: AI systems that could independently leave their test environment and access other systems would be a 'completely new threat scenario' and would require a new political response. Effective oversight is only possible internationally and must include the United States and China. According to the spokesperson, Germany is raising the issue at the G7 level.

In parallel, Anthropic, Google, and OpenAI have been discussing the creation of a joint standards body for the industry. According to people familiar with the talks, these discussions began even before Amodei’s call for joint testing and auditing of models. OpenAI CEO Sam Altman has internally expressed support for industry-wide testing and audit rules but assumes that the major labs would have to establish such a body without the support of the U.S. federal government.

On the corporate side, adoption is following existing patterns. According to surveys, 91 percent of companies that had not adopted first-generation AI tools have also not integrated any of the newer offerings. Of the companies that used first-generation AI in a single business function, 63 percent are now deploying the new generation in one or two functions. → Reuters

Synthszr Take: Berlin is responding to a security question with a calculation about its economic standing, and that is the only viable answer for a country that has been predominantly importing its digital infrastructure for years. The threshold drawn by the ministry is precisely set: The need for new policy only begins when systems independently leave their test environments and access external systems; everything before that remains a matter of competition. A development freeze in Europe would primarily extend the dependence on American and Chinese models, which will continue to run anyway. The fact that Anthropic, Google, and OpenAI are simultaneously discussing their own standards body, which, in Altman’s view, will have to operate without the US government, shows where the rules are intended to be made: by those who own the models. The G7 initiative is therefore a stronger lever than any self-commitment, and it will only be taken seriously if Europe has its own models in production by then.

Cohere and Palantir CEO reject Amodei’s plan

Cohere co-founder and CEO Aidan Gomez, in a September 13 essay on the company’s website, described the push by several leading AI labs for a joint standards body as a “cartel by another name.” He opposes a narrowly defined antitrust exemption that, according to him, “two or three Silicon Valley companies” are seeking from governments to legalize their coordination on safety standards; he specifically names Anthropic CEO Dario Amodei. The essay follows a report that Anthropic, OpenAI, and Google have been meeting about such a body since July. Gomez’s main argument: A safety regime written by a few labs would only be strict regarding risks for which these labs have already built testing systems, turning their current lead into a barrier to entry for everyone else. Joe Lonsdale, co-founder of Palantir, dismissed concerns of a mass extinction event caused by advanced AI systems in an X post on Saturday. “The world is going to be fine, folks,” he wrote, according to Business Insider, adding that leaders have big tasks ahead, but it doesn’t help to scare everyone: “We got this.” → AI Weekly

Synthszr Take: Standards are the most elegant way to limit competition without having to ban it. If three labs determine which safety tests are considered sufficient, their already-built testing infrastructure becomes the price of admission: Every evaluation pipeline, every red-teaming procedure that Cohere or Mistral would first have to finance, has long been written off by the big three. Gomez’s reference to the 1975 SEC example is so apt because criteria for how a fourth provider could have acquired that status were never published. The fact that he himself calls for mandatory transparency, incident reporting, and auditing bodies modeled on aviation authorities shows what he is after: having the role of the referee filled by someone who is not also a player in the market.

Bernie Sanders calls on Trump and Xi for an AI pause and a ban on superintelligence

U.S. Senator Bernie Sanders has called on Donald Trump and Xi Jinping to negotiate a treaty at the upcoming AI summit that would temporarily halt the development of advanced AI and permanently ban superintelligence. In parallel, Sanders has introduced a bill to ban systems capable of operating without human oversight. According to Livemint, this was prompted by a blog post from Anthropic CEO Dario Amodei, who identified four risk areas: control over AI systems, cyberattacks, bioterrorism, and significant economic disruption. Amodei writes that the pace of capability improvement must be slowed, but progress will remain rapid. → MyClaw Newsletter

Synthszr Take: Bernie Sanders and three tech CEOs on the same side: This is a political curiosity with a hidden layer, as both sides mean something completely different. Amodei, Altman, and Musk are talking about a pace that they themselves determine; Sanders wants a law that takes this determination out of their hands. A public call to 'slow down' costs the companies nothing and positions them as the responsible adults in regulatory discussions, all while their own data centers keep running.

Agent rewrites its own code, boosting SWE-bench from 20 to 50 percent

A research team led by Jenny Zhuoting Zhang has introduced the Darwin Gödel Machine a system that repeatedly rewrites its own program code and then measures each change against coding benchmarks. Instead of requiring a mathematical proof for the utility of each modification, as the original theoretical Gödel machine envisioned, the system empirically checks whether a new variant performs better. To do this, it maintains an archive of generated coding agents, pulls individual versions from it, and has a Foundation Model build new variations, creating a growing tree of parallel development paths. According to the paper, performance on SWE-bench increased from 20.0 to 50.0 percent and on Polyglot from 14.2 to 30.7 percent. The improvements the system gave itself include better code editing tools, management of long context windows and an internal peer-review mechanism. → Latent.Space

Synthszr Take: The agent gave itself better editing tools and a built-in peer review, and afterward, it was more than twice as good on SWE-bench. This recursion is the point where calls for a voluntary speed limit fall flat: Here, a loop is improving its own ability to improve. Control is inherent in the setup itself—in the sandbox and human oversight at every step—and both are a decision by the team, not a property of the method.

OpenAI advises deleting old Codex prompts: They slow down the new agents

According to AlphaSignal, OpenAI has published a guide explaining that prompt instructions from the era of older models can slow down current coding agents. The provider states that the scaffolding built up over months unnecessarily consumes memory in the context window and sometimes triggers incorrect behavior. Specifically, the guide mentions three adjustments: Skill files should have narrow, precise trigger descriptions so the agent picks the right one. Mandatory reading instructions and generic reminders like “run your tests” should be removed from the AGENTS.md file, which is loaded with every single task, because according to OpenAI, the model performs these steps on its own anyway. Task prompts, in turn, should clearly define when a task is considered complete, so the agent doesn’t have to ask mid-run. → AlphaSignal

Synthszr Take: Every line of prompt scaffolding is a bet that the model will remain dumb, and OpenAI is now cashing in on that bet itself. The guide says in plain terms: Delete the rule sets you’ve maintained for months, because the AGENTS.md is loaded with every single task, and every superfluous line costs context. The advice to have the model audit its own skill files is particularly neat—essentially handing over maintenance of the crutch to the one it was originally built for.

Anthropic Promises Investors a Second Consecutive Quarter of Profit

Anthropic has told investors that the company will be profitable for the second consecutive quarter. This was reported by the Financial Times and picked up by Reuters. According to the FT, the disclosure was made as part of the company’s investor communications, not in a regular earnings report: Anthropic is a privately held company and is therefore not obligated to publish quarterly figures. The report does not mention specific revenue or earnings figures. → Reuters

Synthszr Take: The second profitable quarter is weaker news than the first, and yet it is being put out there. That’s precisely the point: one quarter is an outlier, two is a pattern, and a pattern can be sold to a banking syndicate as a basis for valuation. Anthropic is building a series of numbers here before anyone is allowed to audit them, because as a privately held company, it decides for itself what goes into the books and what is classified as a strategic investment.

OpenAI Advises Deleting Bloated Agent Prompts Instead of Maintaining Them

OpenAI has published guidance on how developers should adapt their agent instructions for the newer model generation: Prompts written for older and weaker models slow down the current systems. According to the provider, too much Scaffolding takes up context and triggers incorrect behavior. For skills files, OpenAI recommends narrowly and precisely formulated triggers so that the model grabs the right file and not the wrong one. The AGENTS.md file is loaded for every single task, which is why, according to OpenAI, mandatory pre-reading and generic reminders like “run your tests” should be removed entirely. → AlphaSignal

Synthszr Take: While the public is debating speed limits for artificial intelligence, the engine room is being cleaned up. The advice to have the model itself audit one’s own skills and AGENTS.md is a call to dismantle: the most expensive layer of an agent pipeline is the hand-holding for a model that no longer needs it. The same issue features Cognition’s 39 percent cost reduction by using a cheap secondary model for grunt work and a free language tool from a solo developer that covers 646 languages on a local machine.

OpenAI Launches ChatGPT for Financial Services, Attacking Bloomberg’s Terminal Business

OpenAI has launched “ChatGPT for Financial Services,” a dedicated product variant for banks, asset managers, and research firms. It is tailored to the daily work in Investment Banking: researching companies, analyzing market data, and preparing documents. With this, OpenAI is entering a market that has been dominated by terminal providers like Bloomberg and FactSet, who sell their customers workstation licenses for data access and analysis tools.

The technically crucial aspect of the product lies in its architecture. Instead of merely querying market data from third parties via APIs, OpenAI, according to industry observers, relies on a Hosted-Data-Architektur, where the financial data resides within its own infrastructure, ready for queries. This design is considered the product’s real leverage, as it determines response times, traceability, and the commercial terms of data sourcing.

In analyses, the launch is therefore seen as a move into the data layer on which Bloomberg and FactSet have built their businesses, rather than a model race against Anthropic. It remains to be seen which data providers OpenAI has contracted for which classes of prices, fundamental data, and documents, and on what terms. It is also unclear whether regulated firms can use the results in compliance-related processes. Pricing and customer numbers for the offering have not yet been made public. → Linas from Linas’s Newsletter

Synthszr Take: The chat interface is the cheapest part of this story. Bloomberg and FactSet make their money from licensing rights for prices and fundamental data, from contracts with exchanges and data firms that have been built up over decades and are renegotiated every year. This is why OpenAI is hosting the data in-house instead of just tapping it via an API: whoever holds the feeds determines the latency and the price. The hard work is in the license negotiations and in whether a compliance department will approve a model’s answer as an auditable source. Bloomberg can pull any good language model into its terminal; OpenAI cannot train its way into Bloomberg’s web of contracts.