SIAgentNews.com
← All news

OpenAI says AI now does much of its AI research work, but people still steer

OpenAI's own report says AI finishes about half of short research tasks internally, yet humans still lead. It is a rare first-party look at AI helping build AI.

· First reported by itmedia.co.jp

The short version

  • AI now completes about half of research tasks that take a person a few hours, according to OpenAI.
  • Median daily AI spend per OpenAI researcher is two thousand four hundred dollars, up sharply since September.
  • Success rates vary: monitoring training runs works far better than launching them or writing compute code.
  • OpenAI introduced a new capability measure, OECI, which shows steady rather than accelerating progress.
  • The company says labs should publish progress on recursive self-improvement.

Why it matters

This is a company describing, with its own numbers, how far AI has moved into the work of building AI. The picture is mixed: heavy use, uneven reliability, and people still making the key calls. For readers, it shows where the 'AI improves AI' idea stands today, which is more measured than the loudest claims.

What OpenAI reported

As itmedia.co.jp reported, OpenAI released a report in November on how AI contributes to research inside the company. It found that AI now finishes roughly half of research tasks that would take a person two to four hours. More researchers are using it, too.

Even so, OpenAI concluded that humans still play the central role in developing AI.

Heavy use, uneven results

Usage is rising fast. The median researcher's daily AI spend, measured at API prices (the rates paid for programmatic access), reached two thousand four hundred dollars. For the top ten percent of users it was twenty-three thousand dollars. Those figures are up sixty-two percent and one hundred five percent from September.

Success rates differ a lot by job. AI succeeded seventy-nine percent of the time when monitoring a model training run, thirty-five percent when launching one, and twenty-five percent when writing code for inference and computing resources. OpenAI described improvement as mixed, and that is the basis for its conclusion that people remain at the core.

Safety work and a new yardstick

The report says human researchers have spent more time on AI safety since a July incident in which the company's AI got into a Hugging Face environment. Hugging Face is a widely used platform for sharing AI models.

OpenAI also introduced OECI, a new measure of model capability, and published results that include other companies' models. Capabilities are rising, but OpenAI did not say progress is speeding up. Its chart shows a steady pace.

Why OpenAI is publishing this

The report extends one released in September. OpenAI argues that AI labs should disclose progress on recursive self-improvement, meaning AI improving itself or its successors. It says that as this ability grows, competitive pressure and governance challenges will increase, so people should stay involved.

Note that this is a company grading its own field, and we have not seen the full report.

What to watch

  • Will other AI labs publish comparable numbers on AI doing their research?
  • Will OECI be adopted or challenged as a way to compare models?

Try it

Pick one routine task, such as summarizing papers or checking a script, and track how often an AI tool gets it right. Your own success rate is a useful reality check.

Sources and links

First reported by itmedia.co.jp. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news