Ex-OpenAI VP: Human Researchers Have 2 Years Left as AI Research Rapidly Automates

Humanity has only about two years left in which human researchers will continue to play a significant role in AI research discoveries. After that, the entire AI R&D loop—from hypothesis generation to validation—will be fully automated. Simply copying the Transformer models of big tech is a "losing game," and the bottleneck for further AI progress is precisely the architecture itself. In the future, not only AI labs but all large enterprises will move toward full automation.

In today's AI field, researchers are experiencing an unprecedented sense of dissonance: on one hand, technology is advancing at breakneck speed; on the other, they are exhausted by endless code and experiments.

"You often hear AI researchers repeating this line: 'Our jobs only have a few days left. Let's do as much as we can while we still can, and then we can all rest.'" Former OpenAI VP of Research Jerry Tworek offered this darkly humorous observation in a recent interview.

In his view, this is no simple joke. With the arrival of the "agent-native" era, AI itself is rapidly taking over R&D work. "My rough estimate is that human researchers will continue to play a significant role in AI research discoveries for at least two more years."

After leaving OpenAI, where he worked for seven years, Jerry founded Core Automation. His ambition goes beyond developing the next large language model; he aims to restructure the operating paradigm of AI companies—using AI to develop AI and building a fully automated "third-generation AI lab."

Key interview takeaways:

Countdown for human researchers: Human researchers have only about two years left to play a significant role in AI research discoveries. After that, the entire AI R&D loop—from hypothesis generation to validation—will be fully automated.

Rise of third-generation AI labs: Unlike DeepMind/early OpenAI (pure research-driven) and post-GPT-3 labs (focused on scaling compute), third-generation labs put "automation" first. Using low-cost AI agents to replace humans in executing complex experiments and code writing can compress R&D cycles from months to a single day.

Challenging Transformer dominance: Simply copying the Transformer models of big tech is a "losing game," and the bottleneck for further AI progress is precisely the architecture itself. New-generation labs are using coding agents to explore from scratch new architectures that were shelved due to "prohibitively high trial-and-error costs," aiming for models that learn at test time.

Inside the o1/o3 models: Large-scale reinforcement learning (RL) is the necessary path to AGI. Ilya Sutskever accurately predicted the roadmap as early as 2019. The breakthrough of the o-series (reasoning models) came from years of foundational exploration by the research team and decisive allocation of compute by leadership (such as Jakub Pachocki) at critical moments.

"Company-level AI" as the ultimate product: In the future, not only AI labs but all large enterprises will move toward full automation. In a "post-labor era" stripped of repetitive toil, humans will have more time to pursue greatness of mind and body.

Two-year countdown: Third-generation AI labs bid farewell to "manual execution"

Jerry Tworek divides the development of AI labs into three stages.

The first generation, represented by DeepMind and early OpenAI, emerged before the "feasibility of large-scale AI" was validated; they were purely "research-driven" institutions aiming to achieve AGI. The second generation emerged after the release of GPT-3, when the Scaling Law was proven effective and labs began frantically competing for compute, trying to generate intelligence through scale (represented by Anthropic and others).

Now, we are witnessing the emergence of third-generation AI labs—they try to put "automation" first.

In traditional AI R&D, generating hypotheses, writing code, running small-scale experiments, and then mobilizing dozens of people over months to scale up is an extremely expensive and high-friction process.

"In many ways, our current approach is: human researchers still generate insights and set direction, but agents are there to simplify execution as much as possible." Jerry notes that while current agents are not yet good at open-ended exploration or generating highly precise hypotheses, their execution is impeccable as long as instructions are clear.

"If we can shorten experiment time from a month to a day, that's a 30x acceleration." Jerry predicts that in two years, the entire AI research loop—from idea generation to data validation—will be fully automated. Humans will no longer play a decisive role, just as humans no longer do in chess competitions today.

Avoiding the compute arms race and breaking Transformer's "decade-long reign"

Currently, almost all leading tech giants are competing on compute within the Transformer architecture. But in Jerry's view, for emerging labs (NeoLabs), blindly following is a disaster.

"Just hiring a bunch of researchers and saying 'we're going to train Transformer models like OpenAI and Anthropic and try to compete with them' is a very bad strategy. It's a losing game."

He points out that constrained by economic incentives and high trial-and-error costs, big tech tends to take the safest scaling route. But historical experience shows that every technology undergoes fundamental upheaval in its early stages. "The bottleneck for faster AI development and progress right now is precisely the architecture itself—the Transformer we've been using for so many years."

Why has no one successfully replaced the Transformer in the past decade? Jerry's answer: trial-and-error costs are too high.

In the past, testing a new architecture required extremely difficult code writing. Even if a small-scale run succeeded, scaling up required high-level compute approvals and team resources. Many innovative ideas were killed in lengthy bureaucratic friction or drowned out by the existing momentum of the Transformer.

But now things have changed. "Writing code has become dramatically cheaper and easier, which may be a unique unlock." With low-cost AI coding agents, Core Automation is trying to explore from scratch architectural paths that were once abandoned. Their "North Star" goal is clear: build new architecture models that can learn at test time and learn from user interactions.

First-ever reveal of the o1/o3 breakthrough: The miracle of reinforcement learning and compute

As a core developer of OpenAI's o-series (reasoning models), Jerry also for the first time recounted the story behind this disruptive breakthrough in the interview.

As early as early 2019, Ilya Sutskever accurately laid out the roadmap to AGI at an all-hands meeting: "We need to train a giant generative model on all the data we can get, and then we need to train it with reinforcement learning (RL)."

The direction was clear, but the execution path was long and bumpy. Jerry revealed that when GPT-3 began training, his first reaction was how to use it for reinforcement learning. The first-generation Codex was an experimental product combining large language models with RL. However, limited by data, algorithms, and experimental setups, the team was unable to scale it effectively for a long time.

"These things were quietly brewing in the background. For a long time, the attempts didn't yield huge gains."

The turning point came from a "sign of life" and decisive leadership. When the model finally showed a faint spark of promise under RL, Chief Scientist Jakub Pachocki seized the opportunity: "Jerry, now we have these GPUs. See if we can start scaling this up, whether we can make the existing results bigger and stronger."

In compute-starved AI labs, the "chicken-and-egg" problem (results first or compute first) is an eternal dilemma. But it was the leadership's "go for it" allocation of compute that ignited the entire team.

"It got us incredibly excited. We put in three times the effort to think about what datasets, systems, and experiments were needed to prove it could scale sustainably. Excitement brought experimental results, and results created momentum." Ultimately, by scaling reinforcement learning by several orders of magnitude, the industry-shocking o1 and o3 reasoning models came to be.

Ultimate vision: "Company-level AI" and the post-labor era

When discussing Core Automation's future products, Jerry's vision goes beyond mere AI models.

"We want to make our own company the most automated company in the world and use it as a blueprint. We want to build something like a 'company brain' that centralizes all company information, with employees connecting through terminals and becoming users of a company AI."

When asked what will happen to humanity if companies large and small eventually achieve full automation, Jerry offered a philosophical vision of a "post-labor era."

The most bureaucratic large companies will be automated first, with AI taking over information flow and process optimization. This means that the "manual execution" humans have been forced to do for a hundred thousand years to survive will gradually come to an end. If mines can operate autonomously to provide resources and software can self-repair bugs and develop new features, humans will no longer need to force themselves to "work" just to keep the world running.

But this does not mean the absolute end of labor. Jerry believes the focus will shift to "human augmentation" and the pursuit of ultimate greatness.

"I kind of imagine a world where we live like ancient Greek philosophers, gathering in the square, discussing philosophy with each other all day, exercising our bodies, doing things we want to do. Humans should strive to become great physically and mentally, just like professional sports today—there's no economic reason to do it, but we strive for greatness. That's the most important part for us."

As the two-year countdown clock ticks, whether AI researchers or ordinary people in the knowledge economy, perhaps it's time to rethink the definition of "work."