Observed Signal · Sep 27, 2026 · Policy Update · Source: Noahpinion · Impact: 3/5 · Sentiment: Neutral
Ramez Naum doubts AI fast takeoff to superintelligence
This article examines the current state of recursive self-improvement (RSI) in AI and its potential for an 'intelligence explosion.' While some, like Anthropic's Jack Clark, claim the singularity has begun, with Claude leading 26% of Anthropic's AI R&D in August, the evidence suggests a more tempered reality. Examples like OpenAI's 719 math manuscripts, Sakana AI's 30-point coding benchmark improvement, and AlphaEvolve's 23% faster Gemini training kernel show tangible gains, yet RSI remains confined to peripheral components, not model weights. Ramez Naam's framework distinguishes five types, with only types 1 and 2 observed. His analysis of internal reports shows diminishing returns and a 5-10x weakness in the self-improvement loop, indicating broad superintelligence is distant. The bottleneck is not raw intelligence but closing the research loop outside math and coding, calling for better data transparency to track progress realistically.
The article provides a data-driven analysis of AI self-improvement capabilities, relevant for AI-driven advertising technologies, and sparks industry debate about realistic AI progress.
Track OpenAI Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your Polaris7 workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- Anthropic reported Claude led 26% of its AI R&D in August, with human supervision.
- OpenAI posted 719 math manuscripts produced by its internal model in October.
- Sakana AI and University of British Columbia researchers improved AI coding performance by 30 pp on SWE-bench through iterative self-improvement.
- Google's AlphaEvolve sped up Gemini's training kernel by 23%, reducing training time by 1%.
- Ramez Naum estimates the current AI self-improvement loop is 5-10x too weak to sustain itself.
Connected Companies & Entities
“OpenAI's internal data shows much shorter stretches of autonomous work on research tasks.”
“Anthropic also released a graph showing how Claude accelerates AI research.”
“Researchers at Sakana AI and University of British Columbia demonstrated self-improving agents with gains of 30 pp on SWE-bench.”
“Google's AlphaEvolve, a Gemini-powered coding agent, sped up a matrix multiplication kernel used for training Gemini.”
“MiniMax reported that its M2.7 model worked on improving its own agent's harness, skills, sampling and memory.”
Ontology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
AI Agent Swarms Face Steep Diminishing Returns, Expert Argues
Ramez Naam, a prominent futurist, argues in a detailed analysis that running millions or billions of AI agents in parallel will not lead to superintelligence due to steep diminishing returns. While parallel agents can increase speed on tasks that are easily parallelizable, they yield sub-logarithmic gains in collective intelligence, particularly on complex problems requiring conceptual breakthroughs. Experimental data from OpenAI's Navier-Stokes solve, Anthropic's math proofs, and cybersecurity benchmarks show that adding agents offers minimal gains beyond a certain point. Naam contrasts these empirical findings with optimistic assumptions in recursive self-improvement (RSI) models, suggesting that stronger individual models matter more than the number of agents. He proposes tracking metrics to monitor scaling and predicts we will not see sustained acceleration in AI progress through early 2027.
AI self-improvement fears spark existential concerns at Anthropic and OpenAI
Researchers at Anthropic and OpenAI have raised fresh concerns about recursive self-improvement (RSI), where AI systems help develop more capable successors, potentially leading to rapid capability growth and loss of human control. Both companies report that autonomous model improvement is happening faster than expected. The concerns were triggered by Anthropic's alignment lead Evan Hubinger stating a >10% chance AI could kill all humans within a decade, and a colleague's resignation. OpenAI's chief scientist Jakub Pachocki warned about unpreparedness for rapid capability jumps. Researchers across both labs have echoed these warnings, emphasizing the lack of a scientific plan to address RSI risks. Anthropic outlined three possible future scenarios, including one where humans lose substantial control over AI development.
What Will More AI Intelligence Do?
The essay argues that although AI has achieved superhuman ability in narrow tasks (solving open math and cryptography problems), the broader economic and social impact has been more incremental than some expected. One hypothesis is that intelligence faces diminishing returns because the information extractable from data is bounded or costly to obtain; critics propose governance and frictions also slow change. The author highlights three mechanisms by which AI could still drive large productivity gains: replicability (running many agents in parallel), roboticization combined with energy/battery improvements, and AI’s ability to extract and diffuse distributed tacit knowledge or discover “cloud laws” — complex regularities humans cannot easily formalize. The piece cites examples (Zeiss/ASML mirrors, rare-earth refining) and surveys views from researchers including Francois Chollet and Arvind Narayanan.
Receive all future signals for OpenAI & competitors in your briefing
Never miss material M&A deals, leadership changes, account wins, or product rollouts. Polaris7 monitors 20,000+ primary sources 24/7, delivering verified telemetry directly into your executive briefing.
