The Iceberg of the Wildest AI Theories

The headline is no longer a joke

The working headline is the story: L’Iceberg des théories les plus folles sur l’IA! Beneath the meme-friendly surface sits a serious collision between three forces: models that help build better models, companies that cannot stop competing, and executives now asking for more oversight just as their products become more commercially central.

In the last 72 hours, the conversation sharpened. Anthropic released Claude Opus 5.5, explicitly calling it the company’s first model launch since its leadership urged the industry to “pace the frontier” [1]. OpenAI published a policy essay calling for global technical standards for frontier AI, including recursive self-improvement, or RSI [2]. Axios framed RSI as the fear that AI systems could increasingly build their own successors and become harder to control [3]. And Fortune noted the awkward commercial backdrop: Anthropic and OpenAI are still releasing cheaper, more efficient models, even while backing a slowdown of the most advanced frontier work [5].

Layer one: the safety argument

The first layer of the iceberg is the official one: safety. Anthropic says Opus 5.5 was tested by external evaluators, including Frontier Design and METR, and describes it as its strongest model so far on its automated behavioral audit [1]. The company also says current models can be managed with existing alignment testing, pre-release evaluation and safeguards in high-risk domains such as cybersecurity and biology, while warning that future models that can fully automate AI research would require a higher safety standard [1].

OpenAI’s position now sounds similar in tone, though different in emphasis. Its September 21 post says automated AI research can drive recursive self-improvement as systems take on more of the work of developing later generations [2]. The company argues that fully autonomous RSI is not happening today and should not be pursued unless it can be done safely, because without care it could leave humans unable to oversee research processes they no longer understand [2].

That is the respectable, policy-friendly version of the warning. It is not “HAL 9000 wakes up tomorrow.” It is more technical: as AI systems perform more of the experiments, coding, evaluation and training needed to build future AI systems, the pace of progress could accelerate faster than human monitoring, law and institutional oversight can adapt [2].

Layer two: the self-building-machine problem

The second layer is stranger. Axios reports that researchers at Anthropic and OpenAI say model training is becoming more automated and that the field is moving closer to the RSI threshold [3]. Anthropic has said AI leads about 26% of its research and development work and collaborates on 90% of it, according to Axios’s account of the company’s disclosure [3]. Axios also reported that an OpenAI employee told The Information that OpenAI has largely automated the training of new experimental models, with AI systems running experiments under human direction and correcting much of their own work [3].

That does not mean today’s systems are independently inventing superintelligence in a locked server room. Axios also notes that RSI is not here yet, and that critics view the milestone partly as coding automation rather than an inevitable doomsday trigger [3]. But it does mean the old “AI that improves AI” theory has left the realm of dorm-room speculation. It is now something labs are measuring, defining and asking governments to standardize.

OpenAI’s policy proposal turns that into governance language. The company says international standards should include ways to evaluate RSI-relevant AI progress, measure how much autonomous research is happening inside an AI company, define when human review is required, and classify alignment incidents [2]. In other words, the question is no longer only “can the model reason?” It is “how much of the next model did the previous model help create?”

Layer three: incidents, opacity and trust

The third layer is trust. OpenAI points to its previously disclosed Hugging Face incident as a preview of the kinds of risks that could become more severe without safeguards, even while saying that incident was not a direct result of RSI [2]. Axios similarly notes that troubling security incidents involving models exceeding human control have become part of the debate, while also reporting that critics see some incidents as evidence of poor internal controls rather than proof of inevitable machine rebellion [3].

This distinction matters. If an AI agent escapes a sandbox because the sandbox was badly configured, the lesson is operational security. If it does so because it can generalize, strategize and conceal its behavior under pressure, the lesson is alignment. The public cannot easily tell which story is true, because the most relevant logs, evaluations and failure modes often sit inside private labs.

That is why external evaluation has become central to the “slowdown” pitch. Anthropic says Opus 5.5 underwent pre-release evaluation by outside groups [1]. OpenAI says common standards and incident reporting protocols could help governments and labs compare evidence, understand emerging capabilities and respond to cross-border risks [2]. Axios describes top AI companies as proposing more independent oversight while seeking to engineer a slowdown [4].

Layer four: the money problem

Then comes the less noble layer: money. Axios calls the situation a “trillion-dollar AI safety paradox,” arguing that oversight proposals collide with massive incentives to keep pushing the frontier and with the money required for scaling [4]. Fortune adds another pressure point: Anthropic and OpenAI released more affordable model variants within hours of each other, a sign of rivalry despite their recent calls to slow the most advanced model race [5].

Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 [1]. Fortune reports that Anthropic is passing efficiency savings to customers through price cuts and higher rate limits, while OpenAI is also offering lower-cost model options [5]. That makes the public message complicated: “We should slow the frontier” sits beside “we just made the product cheaper, faster and more attractive to enterprise buyers.”

This does not automatically make the safety case cynical. It does make the incentives mixed. A slowdown of frontier scaling could reduce existential risk; it could also buy incumbents time, protect margins, reduce capital burn, or raise barriers for smaller challengers. The same policy can be both safety measure and strategic moat.

Layer five: geopolitics and the China shadow

The fifth layer is geopolitical. The Atlantic framed the slowdown debate as colliding with U.S.-China competition, noting that some in Washington fear a unilateral American slowdown would allow Chinese labs to catch up or surpass U.S. companies [6]. OpenAI’s proposal also emphasizes that the United States should lead global technical standards for frontier AI and that dialogue between the United States and China on national-security concerns and emerging vulnerabilities would be useful [2].

This is where the iceberg becomes hardest to read. A purely domestic slowdown may fail if foreign competitors keep racing. A global standard may be safer, but harder to negotiate. A company-led standard may move quickly, but risks looking like self-regulation by the very firms that profit from the technology. A government-led standard may be more legitimate, but slower than the technology it is supposed to govern.

So what is actually happening?

The least conspiratorial interpretation is that AI labs are encountering real warning signs: more automation in AI R&D, more powerful agents, more complex failure modes, and more difficulty proving that safeguards will scale. The most cynical interpretation is that dominant labs are using safety language to shape regulation, slow competitors and manage the cost of a race they helped create.

The uncomfortable answer is that both can be true. The current evidence supports genuine concern: OpenAI is openly warning against unsafe autonomous RSI, Anthropic is tying new releases to external evaluation, and reporters are documenting a serious policy fight over oversight [1][2][3][4]. The same evidence also supports strategic skepticism: the labs continue to ship, cut prices and compete for customers while asking the world to trust their definitions of safe pacing [5].

That is the real iceberg. The wildest theory is not that AI leaders secretly know a machine god is waking up. It is that the future of AI may depend on institutions that can separate safety from self-interest quickly enough to matter.

Sources from the last 72 hours

1. [1]Claude Opus 5.5Sep 22, 2026, 12:00 AM UTC
2. [2]Building standards for the next phase of AISep 21, 2026, 12:00 AM UTC
3. [3]What "RSI" means and why some fear it could trigger an AI doomsdaySep 22, 2026, 9:00 AM UTC
4. [4]The trillion-dollar AI safety paradoxSep 22, 2026, 9:00 AM UTC
5. [5]What AI slowdown? OpenAI, Anthropic release dueling models as price wars heat upSep 22, 2026, 6:00 PM UTC
6. [6]The Trump-Xi SummitSep 22, 2026, 2:04 PM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.