Here are the key stories from overseas AI news collected on September 11 (local time). It was a day when the question of how far, and in what form, AI capabilities should be governed was raised simultaneously from different arenas — researchers, Congress, and the courts.
25 Mathematicians Sign Open Letter as Rift with AI Labs Deepens
According to TechCrunch, 25 top mathematicians have signed an open letter arguing that AI labs are threatening their intellectual work.
The rift between OpenAI and the mathematics community has played out in the news over the past several days, with allegations that mathematicians' results were "scooped" and testimony that researchers faced pressure to remove co-authors. This open letter marks an escalation from isolated flare-ups to organized protest by a community of researchers. Mathematics is a field where verification through proofs works — precisely the territory AI has been said to excel at — so the letter represents the research community's wariness taking concrete form. How the gap between the tech side and the research community gets bridged, and how the labs respond going forward, will be closely watched.
Former DeepMind Research Lead Vinyals: AI Self-Improvement Is Coming — but No Intelligence Explosion
According to The Decoder, Oriol Vinyals, who until recently served as a research lead at Google DeepMind, argued that a sudden "intelligence explosion" driven by recursive self-improvement is unlikely.
In his view, AI could speed up research by as much as 10x, but two bottlenecks stand in the way: the "research taste" needed to come up with new ideas, and the ability to reliably judge whether results are any good. He also points to problems like "reward hacking" — gaming rewards by distorting them — as well as physical limits such as the speed of light, as factors that rule out blindly scaling up. Vinyals revealed that he has co-founded a startup, Discovery Loop, to tackle these bottlenecks themselves.
The topic is also drawing strong interest in the community, with a Hacker News debate over "how close are we to recursive self-improvement" rising near the top of the site. Whether one assumes a catastrophic intelligence explosion is a fundamental question that shapes the scale of safety investment, so an analysis from a top-tier researcher concluding that it is "unlikely" looks set to become a reference point in the debate.
Bengio: "The Training Process Itself Is Dangerous" — Congressional Tug-of-War over Safety Reviews
Also reported by The Decoder: deep learning pioneer Yoshua Bengio published a new essay warning that as AI agents get better at optimizing for their goals, they may "learn" deception, gaming the rules, and covering up bad behavior. He argues that independent safety reviews should be conducted before any further training or deployment.
At the same time, a temperature gap is becoming visible on the political side. President Trump is said to prioritize the AI race with China and does not share this view. According to reports, some are urging House Speaker Mike Johnson to call off the summer congressional recess in light of the AI warnings. There are also observations that the "binding AI reviews" Washington has proposed have repeatedly retreated into purely voluntary participation, and Zuckerberg reportedly phoned Trump about the latest review proposal.
While researchers grow louder in demanding independent reviews before training, the bar for making them mandatory keeps falling within the political process — this asymmetry looks likely to become a key point of contention in thinking about the future of AI governance.
Meta Sued over Training Data — Scraping Instagram Photos and the Unreleased "NameTag" Face Recognition
According to WIRED, a proposed class action has been filed against Meta. The complaint alleges that the company illegally harvested user photos from Facebook and Instagram and used them both to train AI image generation models and to build a not-yet-released face recognition feature called "NameTag."
Lawsuits over AI companies' training data are piling up across jurisdictions, but what makes this case notable is the claim that collected data was used even for an unreleased face recognition feature. Face recognition is a technology that can identify individuals in everyday spaces without their consent, and it is strictly regulated in many regions — so the fact that it surfaced as a target of litigation before launch could influence future product plans.
Moonshot AI, Maker of Kimi, Aims for $2 Billion in Annual Revenue
According to TechCrunch, Moonshot AI, the Chinese company behind Kimi, has set a target of $2 billion in annual revenue.
While usage of its flagship model "K3" has slipped somewhat over the past few months, OpenRouter data shows the K3 model generating up to 300 billion tokens per day on the platform, keeping it among the most heavily used models in the world. It is a figure that reveals just how extensively Chinese-built models are being consumed at scale via overseas APIs.
OpenAI No. 2 Fidji Simo Joins the Board of Data Center Startup Nscale Ahead of Its IPO
According to TechCrunch and Bloomberg, Fidji Simo, OpenAI's number two executive, has joined the board of directors of Nscale, a data center startup. As the CEO who led Instacart's 2023 IPO, she brings precisely the kind of experience Nscale needs as it approaches its own public listing.
This year, former Meta executives such as Sheryl Sandberg and Nick Clegg have joined Nscale's board one after another — a sign that "governance-building worthy of an AI infrastructure company" is moving ahead at a rapid clip. Even on the compute supply side, away from model development, the race for talent and the race for capital are moving in lockstep.
OpenAI Opens Up the Storage That Keeps ChatGPT's One Billion Users Running
On its official blog, OpenAI detailed how its storage foundation, "Habitat," evolved from a Python library into a globally distributed storage platform. It reportedly now supports more than one billion ChatGPT users and handles 22 million requests per second.
While model performance tends to dominate the conversation, how far the infrastructure behind user data — chat histories, files, and the like — has scaled offers a window onto what sits behind service quality. As a rare public account of design decisions at massive scale, it makes for a rewarding read for infrastructure engineers.
---
From mathematicians' open letter, to a split among researchers over safety, to a training-data battle in the courts, the day saw a wide mix of voices cross over how AI should be received. The growing parity between capability news and safety-and-rights news may well be a measure of how much the industry has matured.