"OpenAI just showed one of the clearest early signs of recursive self-improvement: GPT-5.6 Sol was used to post-train GPT-5.6 Luna. This is not an intelligence explosion (yet). Humans still defined the objective, infrastructure and constraints. But the loop is now visible: frontier..." — Chubby
SUMMARYOpenAI used GPT-5.6 Sol to post-train GPT-5.6 Luna, an example of frontier models helping with the engineering work needed to improve later models. Tejal Patwardhan and others framed it as an early sign that AI systems are starting to contribute to AI research and training pipelines, while still under human-defined objectives, infrastructure, and constraints.
GPT-5.6 sol post-trained luna! — Tejal Patwardhan
Source: https://x.com/tejalpatwardhan/status/2075272564629451110
OpenAI just showed one of the clearest early signs of recursive self-improvement: GPT-5.6 Sol was used to post-train GPT-5.6 Luna.
This is not an intelligence explosion (yet). Humans still defined the objective, infrastructure and constraints.
But the loop is now visible: frontier models are beginning to perform the engineering work required to build and improve the next generation of models.
Once AI meaningfully accelerates AI R&D, every generation helps produce the next one faster. Yes, read that again.
That is how recursive self-improvement begins, not with a model rewriting its own weights overnight, but with AI gradually taking over the research and engineering pipeline that creates better AI.
And this is further proof that the speed of releases is increasing and models are improving even faster. — Chubby I’d be careful calling this recursive self-improvement already. But models helping with training runs, debugging, experiment setup and post-training is still a big deal, since that loop used to be mostly people's glue work. — Yann Kronberg I deliberately didn't say it was RSI but "early signs", precisely for the reasons you mentioned. — Chubby
Source: https://x.com/kimmonismus/status/2075564241721946486