Jerry Tworek’s new lab, Core Automation, is making a blunt wager: the next phase of AI progress will not be won by simply training bigger models on more data, but by automating the research process that produces the models in the first place.
That framing matters because it shifts the target from model scale to research throughput. Core Automation says it is aiming to build “the most automated AI lab in the world,” starting with its own internal research workflow. In practice, that means using AI systems to help generate hypotheses, run experiments, evaluate results, and iterate on architecture choices faster than a conventional lab staffed entirely by humans.
Tworek’s background gives the announcement weight. After seven years at OpenAI, he left in January 2026 and argued that the kind of fundamental research he wanted to do was no longer possible there. He has also been more categorical than most founders about the state of the field, saying in effect that deep learning research is already “done.” Whether or not that claim survives contact with the work ahead, it signals how aggressively Core Automation is positioning itself against the familiar scaling playbook.
A lab that automates its own research
Core Automation is not pitching itself as another team trying to eke out incremental gains from the same recipes. The company’s stated objective is to automate the research function itself, so that a small group of researchers, assisted by capable AI agents, can do work that would otherwise require a much larger organization.
That is a meaningful departure from the dominant operating model in frontier AI, where research output has often been coupled to headcount, compute budgets, and training runs that grow larger with every generation. Core Automation’s bet is that the bottleneck is no longer just model capacity, but the speed at which labs can search the design space.
If that assumption is right, then automating experimentation could become a force multiplier. Instead of humans manually managing every iteration, agentic systems could help orchestrate experiments, surface promising directions, and reduce the time between a hypothesis and a result. The immediate prize is not just cheaper research, but faster discovery.
Beyond pre-training and reinforcement learning
The technical ambition is narrower than the marketing language suggests, and that is a good thing. Core Automation says it is focused on new learning algorithms that go beyond pre-training and reinforcement learning, plus architectures that scale more efficiently than transformers.
That is the real break from the current consensus. Pre-training has been the workhorse of modern foundation models, while reinforcement learning has become a post-training tool for shaping behavior and improving alignment with user goals. But both approaches sit inside a broader paradigm that still depends heavily on brute-force scale. Core Automation is signaling interest in learning regimes that may extract more capability per unit of data or compute.
The architecture side of the bet is equally important. Transformers have become the default backbone for large language models, but they are not the only possible route to scale. A lab aiming to “scale better” will likely be looking at designs that improve efficiency in long-context reasoning, memory, modularity, or search. The point is not simply to replace one architecture with another, but to find systems that make automated experimentation itself more productive.
That is also why the lab’s emphasis on optimization and systems engineering matters. If you want AI agents to run a research loop, the stack has to be robust enough to support repeated trials, reproducibility, and tight control over evaluation. In other words, the architecture of the lab may end up mattering almost as much as the architecture of the model.
Neo Labs and the new R&D tooling landscape
Core Automation is arriving in the middle of a broader reorganization of AI research talent. Tworek’s venture sits alongside a growing set of Neo Labs founded by OpenAI alumni and other frontier-model veterans, including Thinking Machines Lab and Safe Superintelligence.
That ecosystem tells a useful story about where senior researchers now see differentiation. The old signal was scale: who had the biggest training runs, the most data, and the deepest infrastructure. The newer signal is research process: who can discover, test, and operationalize new methods faster.
For tooling vendors and infrastructure teams, that changes the market. Labs chasing automated discovery will need software for experiment orchestration, model evaluation, dataset generation, sandboxing, and traceable agent workflows. They will also need systems that can support a higher cadence of failed experiments without turning the lab into an expensive black box.
In that sense, Core Automation is not just a company formation story. It is also a hint that AI product roadmaps may increasingly optimize for research automation as a first-class capability, not a back-office convenience.
From lab to deployment: what automation changes on the ground
If Core Automation’s thesis works, the consequences show up first inside the lab and only later in products.
The near-term technical implications are straightforward. Automated research requires stronger evaluation frameworks, because a system that can generate many candidate ideas also needs reliable ways to rank them. It requires governance controls, because more autonomy in experimentation increases the risk of wasted compute, unstable runs, and hard-to-audit model behavior. And it requires toolchains that can coordinate agents, humans, and infrastructure without losing track of provenance.
That is a different operational profile from a conventional model shop. Instead of a handful of major training runs per year, the lab may be running a continuous pipeline of smaller experiments, each feeding the next. Product teams that emerge from that environment could be built around faster iteration cycles, narrower specialization, and more explicit instrumentation of what the model is doing during development.
For companies consuming AI models, the downstream effect could be more rapid turnover in capability and architecture choices. If a lab can discover better training or architecture methods faster, product roadmaps may become more fluid, with shorter intervals between major releases and less certainty about which design patterns will dominate.
Risks, horizons, and the pace of change
The upside of this approach is obvious: compress the research cycle, and you may accelerate progress without linearly increasing team size or compute spend. That is the promise behind the phrase “most automated AI lab in the world.”
But the risks are equally clear. Automation can speed up discovery only if the underlying evaluation is trustworthy and the research objectives are well specified. Otherwise, a lab may simply generate more experiments, not better ones. There is also a business-model question: if automated research becomes the new competitive edge, it may be harder for outsiders to tell whether a lab is producing durable scientific advances or just faster iteration on familiar ideas.
The broader strategic question is whether this shift actually displaces the scaling era or merely complements it. Core Automation is betting on a future where the frontier moves because the process of finding new methods becomes itself machine-assisted. That is a plausible next step, but it is not yet a settled one.
What is clear is that Tworek and the Neo Labs around him are pushing the conversation away from raw size and toward discovery speed. If that thesis holds, the most important AI systems of the next few years may not just answer questions better. They may help build themselves better, too.
