loading...
Core Automation's Ceres model merges pretraining and real-world learning into a single continuous process – designed to need 100x less training data than current frontier models.
ENTRY ANGLES
Build governance and observability tooling for continual-learning systems in regulated industries · Build vertical applications on top of the Ceres API once commercially available
VERTICALS
CAPABILITIES
MLOps expertise, Compliance and audit framework knowledge, Continual learning theory
Every language model that runs in production today stopped learning when training ended. The model that answered your question this morning is the same model that answered it a year ago, with the same weights, the same knowledge cutoff, the same blind spots. When the world changes – a product updates, a regulation shifts, an organizational policy evolves – the model has to be retrained. An expensive, slow cycle that keeps every deployed AI perpetually out of date. Jerry Tworek, who led reinforcement learning and reasoning model research at OpenAI for seven years before leaving in January 2026, founded Core Automation on the premise that this architecture is wrong, not immutable.
Ceres, Core Automation's founding model, merges the distinct phases of model development – pretraining, fine-tuning, reinforcement learning from human feedback – into a single continuous process. The stated goal: a model that learns from production interactions without catastrophic forgetting, the long-standing technical failure mode where a neural network trained on new data degrades on what it previously knew. The company claims Ceres will require 100 times less training data than current frontier models.
Core Automation raised $100 million at a $1 billion valuation in late March 2026 from Nvidia, Spark Capital, and Accel – funded before launching a single product. By mid-August 2026, the company filed to raise an additional $300 to $500 million at a $4 billion valuation. Total capital has crossed $532 million.
Catastrophic forgetting is not a niche research problem. It is the structural constraint that prevents enterprise AI from self-updating. The workarounds enterprises currently use – retrieval-augmented generation that pulls context at inference time, periodic fine-tuning cycles, prompt engineering with manually updated system context – all exist because the underlying model cannot learn. Each workaround adds cost, latency, and surface area for failure. If Ceres eliminates the forgetting problem at production scale, the entire maintenance architecture of deployed enterprise AI changes.
The implicit competition here is not other startups. It is the research groups inside Google DeepMind, Anthropic, and OpenAI itself, all of which have published on continual learning. None of those systems are in commercial production. Tworek's structural argument for Core Automation as an independent company is that the established labs have strong incentives to preserve their existing model training business models, which run on expensive pretraining cycles that justify large GPU compute contracts. A company that reduces training cost by 100 times disrupts its own revenue base if that company is also a model API provider.
The $100 million seed at a billion-dollar valuation before launch is a significant prior. Nvidia in particular has strong financial incentives to maintain the current compute-intensive training paradigm – Nvidia's data center revenue is built on the pretraining cycle. Their participation suggests they believe Ceres can work within that paradigm rather than fundamentally against it, or that the competitive risk of not funding the most credible challenger is higher than the cannibalization risk. Spark Capital and Accel's participation is a bet on Tworek's track record; Nvidia's is a bet on structural necessity.
A model that learns continuously from production generates a kind of data asset that doesn't exist today: a live record of how the model's understanding of a domain is evolving, where it's drifting, which interactions are producing the largest weight updates, and whether the drift is directionally correct. Enterprises deploying a continual-learning model in regulated environments – financial services, healthcare, legal – cannot do so without tools to audit, monitor, and bound that learning.
No such tooling exists as a product. The governance and observability layer for continual-learning AI systems will need to be built before those systems can be deployed in any environment where model behavior has compliance implications. The team that builds it owns the choke point between Ceres and the enterprises that most urgently need what Ceres promises: the CFOs and compliance teams who will approve the deployment only after they understand exactly what the model is learning and how it's being bounded.