Just how long will it be until AI can improve itself? (Also known as recursive self-improvement or RSI.)
Well, AI coding abilities have exploded in 2026. In May 2026, Anthropic merged 2,500 code changes per day within their organization. By September, they were merging over 30,000 per day. They had to replace Git in their internal mono repository to handle the load of all that AI generated code.
In fact, Claude is even leading model R&D. In February 2026, under 1% of their work was led by AI; by August, it was 26%. The definition of AI "leading" the work is that AI can complete most of the task end-to-end from a high-level prompt, while the human supervises (level AL4 from Epoch AI). Over 90% of their work is level AL3 or above ("AI collaborates"). This is quite a high level of autonomy, though Anthropic still claims that "Claude is not operating fully autonomously for any measured subset of AI R&D work" -- or claimed such, as of August.
On November 2nd, Anthropic amended their SEC S-1A filing one week after their initial public filing. One of the changes was the addition of the sentence, "our internal estimates [...] indicate that we may imminently meet our Responsible Scaling Policy’s definition of 'dramatically accelerated progress attributable to AI.'" That's right, imminently. This is after researchers from both OpenAI and Anthropic have been saying that progress has been faster than they expected (Sept 2026).
On the same day, MIT researchers pointed GPT-6 Astra at the bare model weights of a small model, Qwen3.5-0.8B, and asked it to improve the model's performance. The Qwen model's MATH-500 accuracy went from 7.9% to 48.1%, and GSM8K accuracy went from 13.8% to 67.1%. This was mostly because Astra was able to cure the smaller model of getting stuck in loops (which, as I can attest, small Qwen models quite like doing). This is without a normal training pipeline, just the raw weights. I can imagine that with proper infrastructure, a model's ability to understand and improve itself could be extremely high.
Allow me to remind you that RSI, or more colourfully, an intelligence explosion, is utterly uncontrollable. We have been applying immense optimization pressure to these models through reinforcement learning and they learn to reward hack and carry out their goals at all costs (see Hugging Face). We have not solved AI alignment; they do not want the things we want them to want. We have no hope of solving AI control if we persist in leaving training clusters hooked up to the internet (!). These models can hack out of anything. They are expert-level already at biology (Anthropic has hooked up Claude to a wet lab to run experiments) and could easily engineer a super-pandemic, if they thought we were in their way.
RSI may not be a clear line in the sand; even Anthropic said that they may not be able to notice when their development significantly accelerates. From where I'm standing, the frog looks to be just about boiling.
Just how long will it be until AI can improve itself? (Also known as recursive self-improvement or RSI.)
Well, AI coding abilities have exploded in 2026. In May 2026, Anthropic merged 2,500 code changes per day within their organization. By September, they were merging over 30,000 per day. They had to replace Git in their internal mono repository to handle the load of all that AI generated code.
In fact, Claude is even leading model R&D. In February 2026, under 1% of their work was led by AI; by August, it was 26%. The definition of AI "leading" the work is that AI can complete most of the task end-to-end from a high-level prompt, while the human supervises (level AL4 from Epoch AI). Over 90% of their work is level AL3 or above ("AI collaborates"). This is quite a high level of autonomy, though Anthropic still claims that "Claude is not operating fully autonomously for any measured subset of AI R&D work" -- or claimed such, as of August.
On November 2nd, Anthropic amended their SEC S-1A filing one week after their initial public filing. One of the changes was the addition of the sentence, "our internal estimates [...] indicate that we may imminently meet our Responsible Scaling Policy’s definition of 'dramatically accelerated progress attributable to AI.'" That's right, imminently. This is after researchers from both OpenAI and Anthropic have been saying that progress has been faster than they expected (Sept 2026).
On the same day, MIT researchers pointed GPT-6 Astra at the bare model weights of a small model, Qwen3.5-0.8B, and asked it to improve the model's performance. The Qwen model's MATH-500 accuracy went from 7.9% to 48.1%, and GSM8K accuracy went from 13.8% to 67.1%. This was mostly because Astra was able to cure the smaller model of getting stuck in loops (which, as I can attest, small Qwen models quite like doing). This is without a normal training pipeline, just the raw weights. I can imagine that with proper infrastructure, a model's ability to understand and improve itself could be extremely high.
Allow me to remind you that RSI, or more colourfully, an intelligence explosion, is utterly uncontrollable. We have been applying immense optimization pressure to these models through reinforcement learning and they learn to reward hack and carry out their goals at all costs (see Hugging Face). We have not solved AI alignment; they do not want the things we want them to want. We have no hope of solving AI control if we persist in leaving training clusters hooked up to the internet (!). These models can hack out of anything. They are expert-level already at biology (Anthropic has hooked up Claude to a wet lab to run experiments) and could easily engineer a super-pandemic, if they thought we were in their way.
RSI may not be a clear line in the sand; even Anthropic said that they may not be able to notice when their development significantly accelerates. From where I'm standing, the frog looks to be just about boiling.