Why Two Former OpenAI and Google Leaders Are Building a New AI Architecture

Jerry Tworek, once a staunch reinforcement‑learning advocate at OpenAI, and Rohan Anil, a former pre‑training lead for Google Gemini, left their companies to found Core Automation. Both concluded that merely enlarging reinforcement learning or Transformer scale will not bridge the gap to AGI, because today’s models cease learning once they are deployed.

The authors observe that while model benchmarks and complex reasoning scores keep rising, real‑world tasks remain ambiguous, evolving, and often undefined in advance. Laboratory‑trained models rely on pre‑collected data and fixed environments; when faced with novel, fuzzy problems they cannot adapt without being sent back to the lab for another training cycle.

From this they identify the fundamental flaw: models do not continue learning after entering the real world . This insight drives their decision to abandon the two dominant research routes—large‑scale reinforcement learning and pure pre‑training—and to ask whether a model that learns only in the lab can truly become AGI.

They examine the two existing post‑deployment learning mechanisms. In‑context learning lets a model use information supplied in the prompt, but its capacity is limited by the context window and the information is lost once the session ends. Continual fine‑tuning updates model parameters with new data, yet it suffers from catastrophic forgetting and demands massive data and compute, making rapid adaptation impractical.

To overcome these limits they propose a third possibility: meta‑learning that embeds “how to learn” into the architecture itself . Such a system would decide which new information to retain, how to integrate it with existing knowledge, and when to modify its behavior, all while preserving previously acquired abilities.

Rohan stresses that the next‑generation architecture must treat pre‑training and reinforcement learning as a single, unified process rather than a sequential pipeline. The current split creates mismatched optimization goals and computational patterns, preventing efficient end‑to‑end learning.

Another bottleneck they highlight is the computational depth of Transformers . Although Transformers have many parameters, they perform only a fixed number of forward‑pass layers, limiting the depth of reasoning for tasks that require many sequential steps. Techniques such as chain‑of‑thought reasoning extend depth by generating intermediate tokens, but this incurs linear time and cost growth.

Consequently, a future architecture should enable dynamic computation depth—allowing the model to allocate more steps only when needed—and support continuous learning without the prohibitive expense of full retraining.

Regarding AGI, they argue that the common definition—outperforming humans on economically valuable tasks—misses a crucial requirement: the system must be able to improve itself without human‑in‑the‑loop data collection and training. Current models still depend on researchers to gather failure cases, augment datasets, and release new versions, so they fall short of true autonomous general intelligence.

Core Automation’s broader vision is to build a highly automated AI lab where intelligent agents generate code, run experiments, manage data, and track results, freeing researchers to focus on hypothesis generation and analysis. By integrating such agents with a continuously learning architecture, they aim to accelerate research cycles from a few experiments per day to potentially hundreds, thereby exploring a wider space of architectural ideas.