Jerry Tworek, the engineer who personally led development of OpenAI's o1 and o3 reasoning models, has made a stark prediction: human AI researchers have approximately two years before their current roles become largely obsolete. The timeline isn't about benchmark scores or computing power. It's about how many parts of AI research are still being done by humans versus AI agents.

What Does Tworek Actually Mean by "Two Years"?

Tworek spent seven years at OpenAI before leaving in April 2026 to start a company called Core Automation, which aims to build "the world's most automated AI lab." His two-year prediction isn't about artificial general intelligence (AGI) arriving or humanity needing to retire. Instead, he's counting how many distinct parts of AI research work remain in human hands.

He breaks research into two main categories: idea generation and execution. Idea generation means coming up with new research directions and deciding what to try next. Execution means turning those ideas into working code, running experiments, and retrieving data. According to Tworek, AI agents have already taken over most of the execution work. The remaining bottleneck is idea generation, where human creativity still dominates, but not for long.

"Agents do have creativity, but it is an extremely low-quality, massively stacked kind of creativity. They generate all kinds of ideas, which are generally very poor," Tworek explained.

Jerry Tworek, Former Lead of OpenAI's Reasoning Model Team

At Core Automation, Tworek's team has compressed a full experimental cycle from one month down to one day, a 30-fold efficiency gain. This acceleration is possible because agents now handle the execution phase almost entirely. The remaining human work is narrowing to a single task: deciding what experiments to run in the first place.

How Many People Actually Understand Cutting-Edge AI?

Tworek made another striking claim during his interview: there are probably only 30 to 50 people in the world who have truly mastered the full process of training and deploying a cutting-edge AI model. Everyone else, including the vast majority of full-time employees at top AI labs, is assisting these dozens of core researchers.

This creates an unusual dynamic in the AI industry. All the major labs are competing for talent from the same small pool of elite researchers. Tworek's own company, Core Automation, recruited talent from Anthropic, Google DeepMind, and OpenAI itself, including Rohan Anil (co-founder, formerly at Anthropic and Google DeepMind), Anmol Gulati (worked on Gemini at DeepMind), and Julia Villagra (former head of human resources at OpenAI). The scarcity of these top talents means the two-year timeline applies specifically to them, not to all of humanity.

Why Hasn't Anyone Replaced the Transformer Architecture Yet?

Tworek believes the transformer architecture, which powers nearly all modern large language models (LLMs), is almost certainly not the optimal solution. The transformer has a fundamental limitation: it cannot continue learning after deployment. No matter how much a user interacts with a transformer-based model, it stays frozen in its original state. The context window, which determines how much text a model can process at once, also cannot support continuous learning.

Attempts to fix this through continuous fine-tuning, a process of updating the model with new data, create new problems. Fine-tuning is extremely inefficient and causes "catastrophic forgetting," where the model forgets old knowledge after learning new information. As a result, most of the industry's work on transformers in recent years has focused on making them cheaper, not more capable.

Yet despite these known limitations, the transformer remains firmly entrenched. During his seven years at OpenAI, Tworek said there were only three to four serious attempts to replace the underlying architecture. Each attempt required researchers to run small-scale verification experiments for at least three months. Only if results looked promising would they attempt to scale up, which meant convincing about ten people with decision-making power to invest three to six months into an uncertain project. Most architectural alternatives either got crushed by the accumulated momentum of transformer research or were partially absorbed as tiny components within the transformer framework.

How to Navigate the Rapid Evolution of AI Research

• Stay Flexible with Model Choices: The open-weight AI model landscape is now moving in days rather than months. On August 26, 2026, both Alibaba and Zhipu released major frontier-adjacent open models on the same day, signaling that your best model option can change between a Monday and a Wednesday.
• Keep Evaluation Suites Ready: Model retirements are becoming routine. OpenAI retired its o3 model from ChatGPT on August 26, 2026, after a 90-day sunset period, forcing users to migrate to newer models like GPT-5.6. Teams should maintain evaluation benchmarks that make switching between models cheap and fast.
• Monitor Open-Weight Releases: Chinese labs including Alibaba, Zhipu, DeepSeek, and Moonshot are now setting the pace for open-weight model releases. These models often come with permissive licenses like MIT, allowing commercial use without friction, making them worth regular re-evaluation.

What broke the deadlock for Tworek at OpenAI was a moment of leadership confidence. When his experimental work showed early promise, Jakub Pachocki, who later became OpenAI's chief scientist, approached him with a simple offer: "Jare, all these GPUs are for you, see if you can make the results you are working on bigger and more powerful." This access to computing resources allowed Tworek to scale up his work, eventually leading to the development of o1.

Tworek noted that most cutting-edge research directions face a paradox: you must first get results before you're eligible to apply for computing resources, but you clearly need those resources to get the results in the first place. The fierce hidden battles in AI labs for computing resources are essentially all attempts to escape this paradox. The exit, Tworek suggested, is sometimes just a small amount of confidence from leadership.

What This Means for the AI Industry Right Now

The broader context of Tworek's departure and predictions comes as the open-weight AI model space is accelerating dramatically. On August 26, 2026, Alibaba released Qwen3.8-Flash-Next, an open-weight mixture-of-experts model with 125 billion total parameters but only about 6 billion active per token, positioned as a technology preview for the upcoming Qwen4 architecture. The same day, Zhipu launched GLM-5.3-Flash, a 320-billion-parameter model with 18 billion active parameters, which was revealed to be the stealth model previously known as Ox Alpha.

Both models were released under permissive open licenses, with GLM-5.3-Flash available under MIT licensing. This represents a significant shift in how frontier-adjacent models are being distributed. The fact that two major Chinese labs released frontier-adjacent open models on the same day signals that the center of gravity in open AI has moved to China, and it underscores Tworek's point about the accelerating pace of AI research.

Meanwhile, OpenAI's retirement of the o3 model from ChatGPT represents a broader pattern in 2026: labs are pruning older models aggressively to concentrate usage on their newest, most efficient lines. This keeps the fleet lean and steers users toward current pricing, but it also means the ground shifts under teams that built on a specific model version.

Tworek's prediction about the two-year timeline for human researchers may sound alarming, but it reflects a specific reality: the bottleneck in AI research is narrowing. As AI agents take over execution, the remaining human work becomes increasingly specialized and concentrated among a tiny group of elite researchers who can generate novel ideas. For everyone else in AI research, the question is no longer whether to adapt, but how quickly they can transition to roles that leverage AI agents rather than compete with them.