Technology watchers have been closely monitoring recent advancements in what has become known as AI Recursive Self-Improvement.
For the layperson, it is quite simply: Robots making robots.
Mass produced humanoid robots, and their eventual revolt against humans, is the premise of the origin story of the word “robot,” coined by Karel Čapek in 1920.
This week at Ink & Time we probe the frontier of AI development, and revisit the original robot apocalypse in the context of self-replication.
We’re unlikely to get a sci-fi Terminator scenario anytime soon, but make no mistake, body-less synthetic intelligence is on the loose, replicating and spreading.
Robots learned to replicate back in 1920, but then it was pure fiction
R.U.R. (1920) unfolds on an island factory where Rossum's Universal Robots: artificial people grown from synthetic tissue, "without souls, desires or feelings," are mass-produced as cheap labor.
Old Rossum built them to prove "that God was no longer necessary"; young Rossum stripped the design to its commercial minimum: "he rejected man and made the Robot."
When the humanitarian Helena urges her soon-to-be husband Domin, who happens to be the General Manager of the robot factory, that the robots should be made to be more human, Dr. Gall secretly engineers them with pain, "irritability," and a soul.
That capability turns out to be fatal.
Ten years on, the humanised robots organise, recognising their own superiority ("The Robots can do everything. You only give orders") and exterminate mankind. But they spared the wrong man: Alquist, a builder, not a scientist, is unable to reconstruct the lost manufacturing formula, and the robots, unable to reproduce, face extinction: "Teach us to multiply or we perish."
No spoilers here. The Epilogue resolves the crisis, but not in the way one might expect.
Order a copy of R.U.R. from Time Warp Editions below and enjoy the ride!
Meanwhile, what is the state of play in self-replicating bots… See below.
Claude Calls for Pause, or does he…?
Anthropic (June 5, 2026) published a major report,"When AI Builds Itself," disclosing that over 80% of code merged into Anthropic's codebase is now authored by Claude, up from low single digits before Claude Code launched in early 2025.
Anthropic engineers are merging 8× as much code per day in Q2 2026 as they were in 2024, driven by autonomous AI agents rather than manual coding. Claude's success rate on open-ended, unspecified engineering tasks reached 76% in May 2026 — up 50 percentage points in six months.
The task horizon for reliable AI completion has been doubling every four months: Claude Opus 3 (March 2024) handled 4-minute tasks; Claude Opus 4.6 (2026) handles 12-hour tasks.
Anthropic co-founders Jack Clark and Marina Favaro wrote: "We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for."
Anthropic called on major AI labs to consider a coordinated, verifiable pause in development, warning that unilateral slowdowns would only change who leads, not reduce systemic risk.
Anthropic's own credibility on the pause call is under scrutiny: the company simultaneously filed confidentially for a US IPO at a $965 billion valuation and walked back a key safety pledge earlier in 2026, saying it would no longer hold back dangerous AI if rivals were close to matching its capabilities. (Investing.com, WSJ)
Does anyone really believe that the race for so-called super-intelligence will slow down, when the company trying to produce it has a trillion dollars of investors’ money?
Claude is a capitalist after all.
And most AI experts agree on at least one thing… The race to any form of Artificial General Intelligence, however fanciful, runs right through Recursive Self-Improvement.
New Startups Racing Toward RSI
Recursive Superintelligence (founded May 2026), led by Richard Socher (former Salesforce/Stanford) and seven co-founders from Google, Meta, and OpenAI emerged from stealth with an explicit RSI mandate:
"The entire process of ideation, implementation, and validation of research ideas would be automatic."
Raised $650M+ from GV, Greycroft, Nvidia, and AMD; valued at $4.65 billion with fewer than 30 employees. (NYT, TechCrunch)
Ricursive Intelligence, a spinout from the co-leads of Google DeepMind's AlphaChip, is building AI to design AI chips, aiming to compress chip design cycles from one to two years down to days.
Phase 3 of their roadmap involves recursively using AI-designed chips to train better AI, under human supervision.
Adaption (founded by Sara Hooker, formerly of Cohere and Google) launched AutoScientist, a tool that trains agents to make incremental improvements to frontier model training pipelines: a near-RSI loop if agents begin pushing the frontier forward independently.
Andrej Karpathy (now at Anthropic's pre-training team) has been openly developing Auto-Research, agent swarms that train LLMs on simple tasks.
Currently confined to GPT-2 scale improvements but widely cited as a proof-of-concept for RSI methodology. It’s open sourced above so you can build your own self-replicating bots - it’s like an early Xmas present.
Learn about the history of humanoids from the Ink & Time archives:
What the Labs Are Actually Doing
OpenAI reported in February 2026 that GPT-5.3-Codex was instrumental in creating itself, helping debug training, manage deployment, and analyse evaluation results.
Google DeepMind's AlphaEvolve uses LLMs to guide evolutionary optimisation of neural-network architectures, data-centre scheduling, and chip design. Not a fully closed loop (humans still set goals and evaluation criteria) but each breakthrough enhances capacity for further AI breakthroughs.
Darwin Gödel Machines (DGMs) from UBC and Sakana AI use evolutionary algorithms to improve LLM-based coding agents; agents can alter their own code (though not the underlying model weights). A newer version can alter its meta-mechanisms for self-improvement.
The AI Scientist (reported in Nature, March 2026) automates the broader research loop: generating ideas, running experiments, writing papers, and peer-reviewing them, a significant step toward folding the full AI development cycle into an automated process.
Anthropic's Mythos model (internal preview) was assessed by five of 18 Anthropic engineers as capable of substituting for a mid-level (L4) engineer with harness improvements. Key remaining weaknesses: self-managing ambiguous week-long tasks, understanding organisational priorities, and epistemic judgment — precisely the capacities required for autonomous RSI.
Expert Commentary & Skepticism
Jeff Clune (UBC): RSI is "one of the hottest topics in Silicon Valley"; he believes "we are right around the corner from recursively self-improving systems" and argues it will "transform science, technology, and all aspects of society and culture."
Google CEO Sundar Pichai: "It's a continuum, and we are all definitely making progress. But in the way people describe RSI, that would represent a next level of acceleration… we aren't quite there yet."
Helen Toner (Georgetown CSET, former OpenAI board): Using AI tools for AI research does not qualify as RSI. "The classic definition is really that there are no humans needed." She distinguishes three milestones: adequacy (AI can do some research without humans), parity (AI-only matches human-only), and supremacy (AI-only outperforms human-AI collaboration). Adequacy may already be near or passed; parity, once reached, would "massively accelerate AI progress, leading to supremacy within another year."
Dean Ball (Foundation for American Innovation): Not a "doomer" on RSI. Notes that taking over the world requires practical steps, running lab experiments, navigating politics, building physical infrastructure. Knowledge is distributed and often tacit; it cannot easily be bundled into one AI mind. "Next year they're automating the grunt who grinds through algorithmic efficiency games," not automating genius.
Meta researchers Jason Weston and Jakob Foerster: Argue that "co-improvement" (keeping humans in the loop) is both a more achievable and a better goal than full RSI.
Ajeya Cotra (METR): Distinguishes adequacy, parity, and supremacy milestones. "I wouldn't be totally shocked if you told me [adequacy] had already passed." Once parity arrives, she expects supremacy within a year.
TechCrunch (May 28): "RSI is the new AGI, and it's just as hard to pin down." Notes that like AGI before it, RSI has become a three-letter byword for a cataclysmic AI takeoff, with significant disagreement about what it actually means and how far away it is.
Visions of humanoids can distract from the very real replication happening under the hood.
Probability Estimates
Anthropic's internal forecast (via Jack Clark): >60% probability that a state-of-the-art AI system achieves recursive self-improvement by late 2028.
External forecaster Jack (via Hashcollision Substack): 60%+ probability that by end of 2028 an AI system will be powerful enough to autonomously build its own successor.
Sources: Anthropic Institute (When AI Builds Itself, June 2026); IEEE Spectrum (May 7, 2026); TechCrunch (May 28, 2026); NYT (May 13, 2026); Reuters/Virginia Business (June 5, 2026); GV blog (May 13, 2026); Jeff Clune / X (May 13, 2026).
Ink & Time is committed to resurfacing and contextualising long lost, forgotten, or suppressed works of literature that hold unique relevance to the times we are now living through.
Subscribe and share with a friend who would benefit from reading more printed books.