OpenAI Researcher: GPT-5.6 Can Now Handle Most Human Intern Tasks
Miles Bennett
OpenAI senior researcher Noam Brown said at ICML that he'd pick GPT-5.6 over a human research intern for most tasks — but a colleague's chip-allocation test suggests the model's capability still has clear limits. The replacement of human research roles is accelerating, yet far from complete.
"I'd pick the AI over the intern" — how big a statement is that?
OpenAI senior researcher Noam Brown said at ICML in Seoul: for most tasks, he would choose GPT-5.6 over a human research intern.
This means → Sam Altman's October prediction — that OpenAI would have an AI research intern by September — is shifting from slogan to internal consensus.
For AI PhD students queuing at OpenAI's booth at ICML, the signal is blunt: human intern slots may shrink sharply.
If AI is truly good enough, why not give it all the chips?
Fellow researcher Yash Pande offered a harder test: if AI truly matched a human researcher, OpenAI should be willing to allocate as many chips as possible to its AI coding tool Codex to do research autonomously.
In plain terms = saying "it's good enough" doesn't count — where you spend the chips is the real vote of confidence.
Pande's implication: OpenAI hasn't reached that allocation scale yet, which means GPT-5.6's capability still has a defined ceiling.
What does the gap between words and resources reveal?
Brown's "good enough for most tasks" is a qualitative judgment; Pande's chip-allocation benchmark is resource behavior. The two don't match.
This reflects a core tension in AI capability assessment: stated confidence and actual resource commitment are not aligned.
In plain terms = qualitatively "good enough," quantitatively "not ready to go all-in" — that gap marks the real boundary of GPT-5.6's capability.
Why keep the tooling simple?
Brown advised developers to keep their harness — the scaffolding tools and workflows built around a foundation model — lean, because the next model generation will likely make most current harnesses obsolete within months.
OpenAI's enterprise product lead Alexander Embiricos echoed this: the product team deliberately avoids hard-coding capabilities that future models should handle natively.
This means → OpenAI believes AI value will keep concentrating in the model layer, not the application layer.
What does this mean for application-layer startups?
Many startups build on OpenAI and Anthropic models, relying on differentiated toolchains as their competitive moat.
In plain terms = if the next foundation model can natively do what your toolchain does, your moat disappears.
This reflects a long-term pressure: every model-layer upgrade can directly erode application-layer differentiation.
Content is for reference only, not financial advice.