OpenAI Chief Research Officer Mark Chen said the firm is nearing AI models that can conduct their own research, placing artificial general intelligence increasingly within reach.
Key Points:
- Chen argued scaling laws remain intact, with pretraining and longer reasoning chains still driving progress toward AGI.
- He said models capable of self-sustaining research are close, a shift that would reshape what human researchers do.
- Chen named a deepening evaluation crisis and unsolved continual learning as the field's biggest obstacles.
Chen Maps The AGI Path
Chen laid out his thinking in a recent podcast interview, where he cooked on camera while explaining OpenAI's research strategy.
He pushed back on the claim that scaling has stalled. The argument resurfaces, he said, whenever the field hits a fresh bottleneck.
The company sits on an exponential curve that has held across nearly 10 orders of magnitude, and little suggests it will break, he claimed.
Chen also pointed to OpenAI's wager on reasoning. He said early doubters inside the company questioned the o1 project before Jakub Pachocki, Ilya Sutskever and a few others pushed it forward.
Now he expects models to take on research tasks that stretch for weeks, producing ideas that move past the blind spots of human experts.
OpenAI's roadmap runs three years, he indicated, ending with models that handle research end to end, from the first idea to the finished result.
Also Read: BitMine Defies The Selloff With A $43M Ethereum Bet, Strategy Blinks
Why The Vibe Researcher Idea Matters
Chen floated a term that drew attention, the vibe researcher.
In that future, he told listeners, the best researchers stop writing every line of code and instead steer models that handle execution and scheduling. Human work narrows to two tasks, asking sharp questions and judging whether an answer carries real taste.
That vision rests on shaky ground, and Chen does not pretend otherwise.
He warned of an evaluation crisis, describing teams that chase benchmark scores without real gains, a habit he labels benchmaxxing. Older tests now sit saturated, and fresh ones lose value almost as soon as they go public.
Continual learning remains the harder gap. Chen called it a basic ability the field still must unlock, even as he said many efforts already target the problem.
If that arc holds, Chen suggested, the scarcest human resource shifts from raw intelligence toward judgment and lived experience.
Chen has made versions of this case before. Around the GPT-4.5 launch he argued the scaling paradigm could keep going, and he has long insisted there is no evidence that scaling laws are dead.
Read Next: CZ Says Binance Was Days From MiCA Approval Before Politics Hit