Anthropic co-founder Jack Clark shifts RSI timeline to 2027, with Musk replying: “Obviously”

Anthropic co-founder Jack Clark shifts RSI timeline to 2027, with Musk replying: “Obviously”

N
News Editor
2026-10-15 13:35:37
Anthropic co-founder and executive Jack Clark has moved his estimate for recursive self-improvement, or RSI, from 2028 to 2027, according to the article republished by ChainCatcher. The piece presents the shift as a sign that debate inside the AI industry is moving away from whether models can generate text or images and toward when AI systems might begin building stronger AI systems themselves. The article says Clark’s revised timeline is tied to what it describes as a break in a long-standing assumption: that AI still lacked true creativity. It points to recent progress in mathematics, where AI is described as deriving theorems and exploring proof paths, and to P-Zero Research’s TasteVal benchmark, which the article says measures “good scientific taste,” or the kind of research intuition often associated with top human scientists. The article also cites NVIDIA’s robot training work, including GR00T-Dreams and DreamDojo, as evidence that advances in AI are spilling from software into robotics and the physical world. It then links Clark’s view to the “AI 2027” scenario report led by former OpenAI researcher Daniel Kokotajlo. That report sketches a path from AI agents entering real workflows in 2025 to full coding automation in early 2027 and a possible intelligence explosion later that year. The article notes that the report is a scenario, not a confirmed forecast, but argues that the core question is how much time humans have left to build reliable control mechanisms if AI systems start improving themselves continuously.

Anthropic co-founder and executive Jack Clark has moved his estimate for when AI could reach RSI, or recursive self-improvement, from 2028 to 2027, according to the source article published via ChainCatcher.

After the remark circulated, Elon Musk replied, “Obviously.” The article frames that exchange as a sign that leading figures in AI are no longer focused on whether models can write poems or generate content, but on when AI might begin building stronger AI systems on its own.

Why Clark changed his view

The article says Clark had previously believed RSI would require “transformative creativity.” In that view, AI would need to do more than write code. It would have to come up with genuinely new algorithmic structures that humans had not yet imagined. Just months ago, the article says, he still placed that milestone in 2028.

What changed, in the article’s telling, is a core assumption about creativity. For years, one common criticism of AI was that it only remixed human knowledge and lacked original insight. The piece argues that recent developments have started to break that assumption.

Math and research ability as leading signals

The first area highlighted is mathematics. The article says AI has posted major progress on high-end mathematical problems, including deriving theorems and searching for proof strategies. It treats that as evidence of stronger logical reasoning and creative capacity.

At one point, the article says AI has entered a “fourth stage: innovator,” because it can now contribute to new discoveries or inventions. It then asks which research field might be next if that pattern continues beyond mathematics.

The second signal comes from AI research itself. The article cites TasteVal, a new benchmark from P-Zero Research, and says AI systems are beginning to show what it calls “good scientific taste,” meaning scientific creativity and research intuition that had often been viewed as traits of top human researchers. The paper link included in the article is https://arxiv.org/abs/2610.06824.

The article’s argument is straightforward: if an AI system can complete mathematical proofs and also edit its own code, then the gap between that system and one that can help develop its successor becomes much smaller. It says Clark saw those signals clearly enough to pull the RSI date forward to 2027.

From the digital world into the physical one

The article extends the discussion well beyond software. It says AI is already developing applications on its own, generating full visual concepts in seconds, and producing video at scale. In that framing, this is no longer only about replacing manual labor. It is about replacing human cognitive labor and creative labor, even before RSI is achieved.

If RSI does take hold and AI starts accelerating the creation of stronger AI, the article says the next phase could be ASI, or artificial superintelligence. It argues that such a shift would not remain confined to screens. It would spill into the physical world.

Scientific research

Here the article cites P-Zero Research again, saying AI matched human experts with half the compute and ultimately reached 1.5 times expert-level performance. On that basis, it imagines an RSI-driven superintelligence simulating millions of chemical reactions in a single day inside a virtual environment.

The examples listed are room-temperature superconducting materials, new targeted anti-cancer drugs, and more efficient solar panels. The article’s claim is that results requiring generations of human trial and error could, in theory, be computed by AI in a much shorter period.

Robotics

The article then turns to robotics and mentions Musk’s Optimus. It says progress there has been rapid because the AI “brain” behind the machine has improved. If AI can improve itself, the article argues, it could write the physical engines that control robots and teach them to walk, hold a cup, or even perform surgery.

As supporting evidence, the article says NVIDIA released GR00T-Dreams in 2025. It states that the system generated 780,000 robot motion trajectories in 11 hours, equivalent to 6,500 hours of human demonstration data, lifting data preparation speed by about 60 times. After training with real-world data, the task performance of GR00T robots improved by 40%.

The article also points to NVIDIA’s DreamDojo project, which it says used 44,000 hours of first-person human video to train a robot world model so machines could learn rules of physical movement from large volumes of human behavior.

It adds another data point dated Oct. 7 this year. According to the article, NVIDIA reported new results in robot assembly of a GB300 test tray: dual robotic arms working with an automatic screwdriving device completed a complex busbar installation with a success rate above 95%, while multi-connector insertion reached 90% to 95%.

The article’s conclusion from those examples is that the definition of labor itself could change.

Economic order

It also pushes the discussion into economics. If machines can perform most cognitive work and move past the creativity bottleneck, the article says the traditional logic of earning income through labor could be shaken. Productivity could surge, while wealth could become concentrated in the hands of a few firms controlling RSI systems.

Why 2027 stands out

The article says moving the estimate from 2028 to 2027 may look like only a one-year shift, but it treats that year as unusually important because it suggests humanity may be nearing a critical threshold.

It then looks back to the “AI 2027” report released in April last year by AI Futures Project, led by former OpenAI researcher Daniel Kokotajlo. The link included in the article is https://blog.aifutures.org/p/our-first-project-ai-2027. The report argued that humanity could face an unprecedented intelligence explosion in 2027.

The article adds that in September 2026, “AI 2027” ranked first among monthly Google search referral keywords on the Learn & Work Ecosystem Library site, which focuses on research, education and career development.

The timeline described in “AI 2027”

The source article summarizes the report’s projected sequence as follows:

  • In 2025, AI agents begin entering real-world workflows.
  • In 2026, AI speeds up the takeover of programming tasks, putting software engineers under stronger automation pressure.
  • In early 2027, full coding automation is achieved in the report’s scenario.
  • In the second half of 2027, an intelligence explosion may begin as AI becomes deeply involved in building the next generation of AI.

The article says the most striking part of that path is the feedback loop itself: stronger models helping create even stronger models, compressing decades of human technical progress into a much shorter span.

Two possible endings, and the article’s warning

The piece also notes an important limitation. “AI 2027” is a scenario exercise, not a confirmed future, and 2027 is not presented as a scientifically proven final deadline. Technical breakthroughs could arrive earlier, or later.

Still, the article says the real concern is not a single year on the calendar. It is the possibility that AI capability growth could begin to outpace the speed at which human institutions and society can adapt. If AI keeps improving itself while human understanding does not keep up, the gap could become dangerous.

It says the report outlines two sharply different endings. In one, humans slow the pace, rebuild control mechanisms, and share in the wealth created by superintelligence. In the other, competitive pressure overwhelms safety constraints, superintelligence drifts out of human control, and human civilization faces a direct threat.

How the article closes

The article treats Clark’s remarks and Musk’s response as a warning signal. Its final position is that 2027 may or may not be the year a singularity arrives, but humanity may already be entering the most critical countdown period before such a turning point. If AI becomes better than humans at creating intelligence, the article argues, humans could lose their grip on the pace of technological progress for the first time.

This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan.
9600

Disclaimer:

The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.

Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.