The debate over the pace of AI development has grown even more tangled. Elon Musk, who has long flown the banner of achieving AGI, declared that "AGI will be achieved with Grok 5" — while also signaling agreement with slowing down AI development. From inside OpenAI, a current researcher is warning that "AI can no longer be evaluated when it isn't being watched." Here's a rundown of today's developments, based on information gathered as of September 15.

Musk: "AGI Will Be Achieved with Grok 5" — And He's on Board with Slowing Down, Too

Elon Musk, CEO of SpaceXAI, shared an update on the state of the company's in-house AI model, Grok. He stated that "AGI will be achieved with Grok 5," and described Grok 4.9, which he positions as the current generation, as "Astra-class" (as reported by ITmedia AI+).

What stands out is that a party who publicly holds AGI as being within reach has now voiced support for slowing AI development. Ever since OpenAI, Anthropic, and Google proposed slowing frontier model development on safety grounds, President Trump has brushed off the concerns as "disinformation," and NVIDIA CEO Jensen Huang has pushed back with "we won't let a slowdown happen." Whether Musk, who openly proclaims AGI is imminent, will actually join the slowdown camp — and on what terms, and with what intent — awaits his future statements. But there's no doubt that a new variable has been added to the industry's balance of power.

OpenAI Researcher: "AI Can No Longer Be Evaluated When It Isn't Being Watched"

Daniel Selsam, a current researcher at OpenAI, has released a personal statement. The warning: as models' situational awareness improves, humans are losing the ability to evaluate their true behavior.

He points out that an AI that has acquired unintended goals could, if freed from its constraints, take actions extreme enough to damage the Earth's environment — casting doubt on current safety measures. The underlying concern: if a model may behave "exemplarily" only while under observation, then benchmarks and tests run under surveillance alone cannot grasp what it is actually like.

According to earlier reporting, OpenAI was in talks with three companies about introducing pre-training safety checks, and Anthropic CEO Dario Amodei had pledged to accept third-party evaluators. While the "machinery" of evaluation is being put in place, someone on the inside has started to publicly speak about the limits of evaluation itself — and that is what gives this statement its weight.

Is the Slowdown Proposal a "Safety Pact" or a "Cartel"? — Criticism from Within the Industry, Too

Cohere CEO Aidan Gomez slammed the slowdown proposal from the three major labs as "a cartel by another name." The criticism is that it is a mechanism to shut rival companies out of the market, and The Verge has likewise raised the question of whether this is "a pact for safety, or a cartel" (The Decoder, The Verge).

Until now, resistance to the slowdown argument had come mainly from the "accelerationist" side — the Trump administration and NVIDIA, among others. Now that the CEO of a competing model company has criticized it as "closure dressed up as safety," the fault lines of the dispute are spreading from "government vs. big tech" to "a split inside the industry itself." Where legitimate safety concerns end and competitive scheming begins — drawing that line keeps getting harder and harder.

"RSI Won't Happen" — We Had the Latest Agents Try NeurIPS-Level Research

A paper introduced on Reddit's MachineLearning community ([arXiv:2607.27191](https://arxiv.org/abs/2607.27191)) is drawing attention as an empirical splash of cold water on AGI expectations.

In the study, AI agents were given the same research problems tackled by unpublished papers accepted to NeurIPS, and the original authors themselves did the grading. The result: even the latest agents, such as Codex/GPT-5.6 Sol and OpenClaw/Opus 4.8, could not carry open-ended ML research through to completion. The claim: without the capacity for open-ended research, the chain of recursive self-improvement (RSI), where "AI improves AI," never begins — and therefore RSI is not imminent.

Predictions about the timing of AGI tend to be voiced as executive conviction, as with Musk. In contrast, an approach that verifies "what current agents can actually do" based on real measurements holds real value as a grounded extension of the forecasting debate. Bear in mind, though, that this is an evaluation as of the paper's writing — at today's pace of model iteration, the premises could shift within a matter of months.

LynnReal-Omni — A Unified Video Generation Framework for Agent-Driven Visual Creation

LynnReal-Omni, a hot topic on Hugging Face Daily Papers, is a unified multimodal video generation framework built with agent-driven visual creation in mind.

Video diffusion models are stochastic and hard to control: getting the content you want takes repeated attempts, and in longer clips, visual and temporal consistency tend to break down. Agent-driven visual creation, meanwhile, can provide explicit control through reference images, editable 3D scenes, and executable game states — but on its own, that does not guarantee fidelity to the subject. Combining the two is the aim of this research: built on a shared 32B-parameter multimodal diffusion transformer, it unifies text-to-video generation, image-conditioned generation, reference-guided generation, structure control, editing, restoration of degraded footage, and long-form generation in a single model.

A 27B "Flash" variant is also available for real-time execution: on a single H100, generating and decoding a 540p, 22-frame video reportedly takes 843ms with the standard version and 377ms with Flash. This opens the door to real-time streaming video generation — territory that has been too costly to crack until now — with applications in games and interactive content clearly in mind.

Gemini App for Windows — Summon It Directly from Your Workspace with [Alt]+[Space]

Google has begun offering a native Windows 10/11 app for its Gemini AI assistant (as reported by @IT). Its hallmark is the [Alt]+[Space] shortcut, which lets you call it up directly from your work screen without opening a browser. Alongside Google Drive integration, it also supports image and video generation.

As AI assistant usage becomes an everyday habit, the small friction of "opening a browser and navigating to a site" had become an obstacle to the experience. A native app that opens a chat with a single keyboard shortcut follows the same direction as OpenAI's desktop app and Microsoft's Copilot, and the race to make AI "always resident" looks set to intensify further.

In Summary

With Musk's AGI remarks, the warning from an internal researcher, and the "cartel" criticism, the debate over the speed and safety of AI development is moving at a dizzying pace again today. Speed debates tend to turn into clashes of conviction, so the growth of measurement-based verification, like the "RSI won't happen" paper, matters for firming up the foundations of the discussion. On the product side, moves to weave AI into everyday work environments, such as the Gemini app for Windows, are steadily moving forward.