{"slug": "why-self-improving-ai-still-depends-on-humans", "title": "Why self-improving AI still depends on humans", "summary": "OpenAI reported on Thursday that its internal coding agents independently complete 46% of two-to-four-hour research tasks as of October, up from 23% in March, and 35% of four-to-eight-hour tasks, up from 13%, but success is uneven: agents hit 79% on monitoring training runs, 35% on launching them, and just 25% on writing inference and serving code. OpenAI said \"humans remain a core part of the AI research loop\" and that its reporting \"should not be read as an intention to pursue full RSI,\" while Google's DeepMind Institute published an essay the same day arguing recursive self-improvement could accelerate progress toward AGI but may stall where automated verification is weak. The disclosures arrive as California governor Gavin Newsom outlined rules around RSI in a regulatory framework published earlier this week.", "body_md": "Though frontier AI is becoming increasingly capable as it inches closer to being able to improve itself, not all capabilities are advancing equally.\n\nOn Thursday, OpenAI released a report [examining the state of research acceleration](https://openai.com/index/research-acceleration-jagged-frontier/) within the lab, tracking how coding agents have changed the way that AI is developed as the question of recursive self-improvement or the ability for AI to develop and improve itself, still looms. The findings show that, while AI is speeding up certain parts of model development, these agents still struggle with complex engineering tasks and coming up with original research ideas.\n\nThe lab noted that transparency about \"indicators\" of RSI is \"essential to inform a public discussion of the capabilities, risks, and required safeguards for increasingly powerful AI; however, our reporting should not be read as an intention to pursue full RSI.\"\n\nHowever, it's undeniable that agents are getting better at doing their research:\n\n- According to OpenAI, internally, agents now successfully complete 46% of \"two-to-four-hour\" research tasks independently as of October, up from 23% in March. Meanwhile, tasks that take four to eight hours of human work now see a success rate of 35%, up from 13% in March.\n- However, that success varied depending on the task. Agents were successful 79% of the time at monitoring training runs, 35% of the time at launching training runs, and only 25% of the time at writing code for inference and serving stacks. \"Overall, uplift is jagged, and humans remain a core part of the AI research loop,\" OpenAI said in its report.\n- Humans, meanwhile, are also still responsible for \"research taste,\" OpenAI said, as these models are currently not good at coming up with scientific ideas on their own, something that could create a \"bottleneck for recursive self-improvement,\" the company mentioned.\n\nWhile OpenAI's research points to an uneven course of reaching self-improving AI, a recent essay from the [DeepMind Institute](https://institute.deepmind.com/essays/what-comes-after-agi/?utm_source=x&utm_medium=social&utm_campaign=dmi&utm_content=sl) may see it as the key to advancing even further. On Thursday, Google's research arm published an essay exploring the paths to artificial general intelligence and, beyond this, artificial super intelligence. The essay notes that recursive self-improvement could dramatically accelerate the path to these breakthroughs through things like code and model architecture design, data generation and agentic collaboration.\n\nHowever, DeepMind notes that RSI itself may be jagged, as the feedback loops that allow these models to self-improve can easily \"stall or degenerate\" if the feedback itself is poor. \"RSI should advance fastest in domains that allow for automated verification without humans in the loop, such as mathematical proofs, while other domains might lag behind or taper off sooner,\" the essay reads.\n\n## Our Deeper *View*\n\nRecursive self-improvement has become a hot-button topic in the AI industry, as many buy into the promise that creating a way for these machines to rapidly improve themselves could open the door for technological and scientific breakthroughs beyond human imagination. However, with that kind of power comes an incredible amount of risk. It's why California governor Gavin Newsom [outlined rules around RSI](https://www.thedeepview.com/articles/newsom-pitches-california-ai-rules-as-us-template) in his regulatory framework published earlier this week, and why many of the industry's safety advocates have been calling out the issue of [containment and potential loss of control](https://www.thedeepview.com/articles/why-hundreds-of-ai-researchers-want-stronger-sandboxes) of these systems. The labs themselves understand this too, hence why OpenAI explicitly noted in its essay that its research does not mark an intention to pursue RSI, and why Anthropic called out the dangers of RSI in [research that it published over summer](https://www.thedeepview.com/articles/anthropic-s-rsi-warning-contrasts-with-ipo-filing). Though this research shows that AI has a long way to go to reach this level of capability, with so many calling out the dangers, the question that we should be asking shouldn't be If it's possible, but rather if the benefits of it outweigh the risks.", "url": "https://wpnews.pro/news/why-self-improving-ai-still-depends-on-humans", "canonical_source": "https://www.thedeepview.com/articles/why-self-improving-ai-still-depends-on-humans", "published_at": "2026-11-20 00:24:37+00:00", "updated_at": "2026-11-20 01:11:48.978686+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-research", "ai-safety", "ai-policy"], "entities": ["OpenAI", "Google", "DeepMind Institute", "Gavin Newsom", "Anthropic", "California"], "also_reported_by": [], "takeaway": "OpenAI's internal agents complete 46% of two-to-four-hour research tasks unaided (up from 23% in March) but only 25% of inference and serving code tasks, so keep humans on research taste and complex engineering while automating monitoring.", "recommendation": {"advice": "Track agent autonomy per task type rather than in aggregate: OpenAI saw 79% success on training-run monitoring versus 25% on inference code, so scope delegation to tasks with automated verification.", "playbook": "evals-testing", "playbook_url": "https://wpnews.pro/playbooks/evals-testing"}, "alternates": {"html": "https://wpnews.pro/news/why-self-improving-ai-still-depends-on-humans", "markdown": "https://wpnews.pro/news/why-self-improving-ai-still-depends-on-humans.md", "text": "https://wpnews.pro/news/why-self-improving-ai-still-depends-on-humans.txt", "jsonld": "https://wpnews.pro/news/why-self-improving-ai-still-depends-on-humans.jsonld"}}