While a new approach to AI agent self-improvement arrived, the debate over slowing AI development moved into the courtroom — a day when research and governance both shifted at once. From the flood of papers filling up conferences to AI accelerating vulnerability discovery, the recurring theme was AI speeding up the very people who build with it.
DeepMind's Dream-RSI: Agents Reuse Past Attempts as "Dreams"
Announced by Google and DeepMind, **Dream-RSI** is a technique that has AI agents replay their past exploration runs as "dreams," letting them test new strategies without costly recomputation.
In reinforcement learning, where agents learn through trial and error in an environment, verifying each new strategy has required actually running it — making compute cost a major bottleneck. Dream-RSI sidesteps this by replaying already-recorded past explorations to evaluate how good or bad a strategy is — much like a "dream" in which the day's experiences are replayed during sleep to deepen learning. In tests, it reportedly matched or exceeded existing methods while **cutting iteration counts by up to a factor of about 2.4**.
What stands out is that only the exploration strategy adapts — **the underlying AI model itself stays unchanged**. Because agent behavior can be improved without the heavy step of retraining the model, the impact on real-world operating costs is expected to be relatively small. As use cases that keep agents running over long periods expand, how to design the reuse of past experience looks set to become an even more important theme.
The "Conspiracy" Over the AI Development Slowdown: OpenAI, Anthropic, and Others Hit with an Antitrust Lawsuit
The string of reports over the past few days on AI development slowdown talks has taken a new turn. According to Bloomberg Law, OpenAI, Anthropic, Google, SpaceX, and xAI have been hit with an antitrust lawsuit alleging they "**conspired**" as an industry to slow the pace of development.
Last week brought reports of voluntary self-regulation talks among major AI companies for safety's sake, along with speculation that a slowdown might run afoul of antitrust law. This lawsuit carries a thorny question into court: how does safety-motivated coordinated behavior collide with competition law? As inter-company agreements multiply in the context of AI safety, this issue will only get harder to avoid.
In the same vein, CNBC ran an explainer on AI "kill switches" (emergency stop mechanisms), quoting the view that the safeguards are "**probably too late**" rather than too small. Notable is the framing: it questions the timing of implementation more than the scale of the measures.
ICLR 2027 Reaches About 50,000 Abstracts — The Paper Flood Accelerates
ICLR, a leading international AI conference, is feeling the shadow of the AI boom in the form of a flood of submissions. According to The Decoder, **ICLR 2027 had already hit roughly 50,000 abstracts before its deadline** — a surge of more than 2.5x compared with ICLR 2026's 19,500.
The article cites three drivers: research funding and attention concentrating on AI, corporate rewards tied to paper counts, and above all the fact that **AI itself is accelerating the mass production of papers**. As writing costs fall, submissions climb — fueling concerns that the review burden and quality problems will only worsen. Both as a matter of researchers' working conditions and of conference sustainability, this is becoming a structural problem that can no longer be ignored.
"Before Debating Slowdown, the Vulnerability Explosion Is Already Underway"
WIRED argues that, apart from the talk of AI labs weighing industry-wide slowdown agreements, we should be looking at **a different explosion already underway**: widely available AI chatbots are already helping drive the discovery of a "tsunami" of security vulnerabilities.
Among security researchers, AI-driven vulnerability hunting and reporting is said to be growing, sharply increasing the burden on organizations pressured to patch. As model capabilities rise, defenders' capacity to keep up is falling behind — making this not a "future risk" of the kind slowdown debates discuss, but a challenge already in progress.
Still, more cautious takes on this reporting exist. A blog post by an engineer arguing that AI models are not hacking "**autonomously**" drew attention, offering the view that the cases making headlines merely show models operating within human instructions and context. Rather than impressions of "AI breaking in on its own," accurately discerning what is actually happening in real-world use looks likely to form the groundwork for debate — one neither dragged toward excessive fear nor toward optimism.
Mathematicians Hate AI. They Still Can't Quit It
We recently covered an open letter from 42 mathematicians warning of existential risk from AI; now WIRED paints a different side of the same field. The arrival of powerful AI models has posed an **existential risk** to the discipline of mathematics, yet researchers can't give up their dependence on them because they are simply too useful — that is the article's central point.
Verifying proofs, exploring conjectures, assisting with calculations. The efficiency AI brings accelerates research, while realistically, returning to a research life without it has become all but impossible. The piece digs into how mathematicians themselves are coming to terms with the contradiction of being unable to walk away from the very technology endangering their field — a fascinating read as the flip side of the warning sounded by that open letter.
Mentions of MiniMax "M3.1" Surface in an Internal Repository
In local-LLM circles, a story that looks like a telltale sign of a next-generation model is making the rounds. A recent commit to MiniMax-Code, MiniMax's repository for its coding agent, was found to contain a **mention of "MiniMax M3.1"** in a test file. The discovery came from an AI news account on X, which also shared a direct link to the test code in question.
There are past examples at other companies where next-generation model names left in internal repository test code became pre-release clues. This time too, nothing is known about M3.1, so nothing can be said for certain — but it is drawing attention as a possible sign that development has reached a certain stage.
---
Dream-RSI learning from past experience without changing the model, the collision between safety-motivated coordination and competition law, and the "mass production" of papers and vulnerabilities alike. All of these look like different faces of the same question: how far can human-side systems absorb what AI accelerates? Where the slowdown lawsuit heads, and how far ICLR submissions ultimately climb, are the things to watch.