The shift is recursive self-improvement stepping out of the whiteboard and into production. Together AI just shipped a compiler that writes faster GPU kernels than human engineers can, and Anthropic told the SEC its next models may imminently cross the threshold where AI doubles the pace of AI research itself.
Together AI Ships The First Polymorphic Inference Compiler
Together AI, the Nvidia-backed startup valued at $8.3 billion after an $800 million Series C, released a Compiler that automates GPU tuning for newly released open-source models. The tool is the first piece of what the company calls Polymorphic Inference — a workflow where AI agents continuously optimize the systems that run open models. Instead of engineers spending days or weeks hand-tuning kernels for each new architecture, the compiler delivers production-ready code the same day a model drops.
Vipul Ved Prakash, cofounder and CEO, said open models have closed the quality gap with frontier labs while release cadence has accelerated to major drops every six weeks. “The being able to maximize the value that you can extract from GPUs that are already in the ground is like becoming more important than ever,” he noted. The compiler has already outperformed hand-tuned megakernels, hitting 1.3x faster throughput on Llama 1B across every output length tested. Two-thirds of enterprises forecast moderate or major shortages of accelerated compute, according to ISG research, making every cycle count.
From Turing’s Child Machine To Seed AI
The concept traces back to 1950, when Alan Turing proposed a “child machine” that would be educated rather than programmed with adult intelligence. Decades later, Eliezer Yudkowsky coined “seed AI” — a system that could not only learn but rewrite its own architecture. The Voyager agent demonstrated a primitive version in 2023, mastering Minecraft by iteratively prompting an LLM for code, refining it from game feedback, and storing working programs in an expanding skills library. Google DeepMind’s AlphaEvolve followed in 2025, an evolutionary coding agent that mutates and combines algorithms using an LLM, discovering novel optimizations — though it still requires automated evaluation functions to close the loop.
By 2026, the pattern is unmistakable. Anthropic reported that as of May, more than 80 percent of code merged into its codebases was generated by its own models. The trend underneath is clear: AI is no longer just the product. It is becoming the production line.
Anthropic’s S-1 Warning: The Threshold Is Close
In an amended S-1/A filing submitted November 2 ahead of a planned Nasdaq listing, Anthropic disclosed that internal estimates indicate it may imminently meet its Responsible Scaling Policy definition of “dramatically accelerated progress attributable to AI.” The policy defines this threshold precisely: AI doubling the rate of progress in overall AI capability, so that one year brings what would normally take two. The filing links this threshold directly to what “has been termed ‘recursive self-improvement.’”
The company also acknowledged it has already used AI to develop new models whose capability leaps went unrecognized at the time — identified only after the fact. “We have in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred,” the filing states. No current model has been designated as exceeding the threshold. Over 90 percent of contributions to Anthropic’s codebase in September 2026 were written by Claude, part of a cycle the company calls “closing the loop.”
Stability Risks And The Intelligence Explosion
The safety implications are explicit. Anthropic’s filing warns that using AI to build more capable AI “could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress.” The company is considering policy updates requiring human review and approval of certain research decisions as models take on more of that work, and says it does not expect to safely let Claude design and build new frontier models “without meaningful human supervision or understanding.” It recently delayed pre-training a new frontier base model by more than a month to tighten security around model weights.
Whistleblowers testified before the New York City Council on October 5. Jacob Coxon, a former OpenAI and Anthropic researcher, told lawmakers human oversight is “already not great” given how much work has been offloaded to AI. “The humans in charge will see results and read AI written summaries, but they won’t really understand what is going on or how to fix it,” he said. Daniel Kukutajlu, a former OpenAI researcher, argued tech firms ought to be required to prove they can control systems before deploying them at the frontier.
What’s Changing Now
The compiler at Together AI shows RSI in a narrow, practical form: AI optimizing the inference stack that runs AI. Anthropic’s filing shows the same logic turning toward the research stack that builds AI. The gap between “AI writes code” and “AI improves the AI that writes code” is narrowing. Watch this space — the next S-1 amendment may not be a warning. It may be a designation.
Sources
- Together Brings Recursive Self-Improvement To Open Models
- Recursive self-improvement
- Anthropic warns that it's very close to RSI capability, aka self-building AI | TechCrunch
- Anthropic Claims to Be Near Recursive Self-Improvement
- Anthropic: Recursive Self-Improvement May Be Imminent
- Anthropic Says It Is Nearing Self-Improving AI As Whistleblowers Sound Alarm
Photo: U.S. Air Force AFRL by Aleah M Castrejon (Public domain)
