Automated AI Lab: Core Automation Launches to Revolutionize Research

The artificial intelligence landscape underwent a seismic shift on April 22, 2026, as the industry witnessed one of the most significant talent migrations in its history. While the era of “Scaling Laws” and brute-force compute dominated the early 2020s, a new paradigm has officially arrived. At the center of this revolution is Core Automation, a high-stakes startup founded by former OpenAI Vice President Jerry Tworek. By successfully poaching elite researchers from Anthropic and Google DeepMind, Tworek has signaled the end of the manual research era and the beginning of the Automated AI Lab.

The founding of Core Automation is not merely a personnel shuffle; it represents a fundamental philosophical departure from how neural networks are conceived, built, and deployed. For years, the industry has relied on human-led trial and error—manually tweaking hyperparameters, curated datasets, and static model architectures. Core Automation intends to turn this on its head, treating the research process itself as a massive optimization problem to be solved by AI.

The Genesis of the Automated AI Lab: Beyond Scaling Laws

For half a decade, the mantra of “more data, more compute” reigned supreme. However, the Automated AI Lab movement suggests that we have reached the point of diminishing returns for traditional scaling. Jerry Tworek, who previously spearheaded critical breakthroughs at OpenAI—including the Chinchilla scaling laws and the early foundations of ChatGPT—exited the firm in early 2026, citing a need to pursue “high-risk, high-reward research” that larger organizations had begun to sideline in favor of commercial stability.

Tworek’s vision for an Automated AI Lab is built on the premise that AI should not just be the product, but the researcher itself. To achieve this, he has assembled a “dream team” of researchers, including Rohan Anil and Anmol Gulati. Anil, formerly a key architect at Anthropic, and Gulati, a lead on Google DeepMind’s Gemini project, have been “nerdsniped” into the venture, a term they used to describe the irresistible intellectual challenge posed by Tworek’s roadmap. Their collective goal is to build systems that automate the discovery of new learning algorithms, effectively allowing the AI to design its own successors.

According to internal communications and the company’s recent unveiling, Core Automation is focused on three technical pillars:

• Self-Evolving Architectures: Moving away from the static Transformer stack to models that can dynamically restructure their internal connections during training.
• Continual Learning: Solving the “catastrophic forgetting” problem, allowing models to learn from real-world interactions in perpetuity without needing to be retrained from scratch.
• Automated Hyperparameter Discovery: Utilizing meta-learning agents to handle the millions of micro-decisions that currently require thousands of human research hours.

Technical Deep Dive: The “Ceres” Model and Continual Learning

At the heart of the Automated AI Lab vision lies a proprietary model dubbed “Ceres.” Unlike the massive, frozen snapshots of intelligence we see in models like GPT-4 or Gemini 1.5, Ceres is designed for lifelong learning. In current paradigms, updating a model requires a full training run or expensive fine-tuning on a fixed dataset. Core Automation claims that Ceres can update its weights seamlessly while operating in production, maintaining a balance between plasticity (learning new info) and stability (retaining old info).

One of the most provocative claims from the Core Automation team is that their approach will require 100 times less training data than today’s frontier models. This is achieved by moving away from simple gradient descent in favor of more biologically inspired learning rules. By mimicking the efficiency of the human brain—which does not need to see “half the internet” to understand a concept—Ceres aims to achieve higher-order reasoning with a fraction of the traditional energy and data costs.

The Architecture of Research Automation

To power this Automated AI Lab, the startup has developed an orchestration layer that functions as a “meta-scientist.” This system uses automated evaluation and reproducible experiment tracking to run thousands of parallel simulations. While earlier attempts at research automation, such as Sakana AI’s “The AI Scientist,” showed that AI could write research papers and run experiments for as little as $15, they often suffered from hallucinations and logical inconsistencies. Core Automation is raising the bar by integrating formal verification and “digital-twin” simulations to ensure that the discoveries made by the AI are technically sound and physically viable.

ICLR 2026: The Intersection of Cognitive Modeling and Automation

The move toward specialized, self-evolving architectures was further validated at the 2026 International Conference on Learning Representations (ICLR). Researchers from the École Polytechnique Fédérale de Lausanne (EPFL) presented a groundbreaking study titled “Inducing Dyslexia in Vision Language Models.” This research utilizes what the team calls “digital twins”—Vision-Language Models (VLMs) that are intentionally architected to mirror specific human cognitive structures.

By identifying and perturbing “visual-word-form-selective” units in these models—analogous to the Visual Word Form Area (VWFA) in the human brain—the EPFL researchers successfully simulated dyslexic reading behaviors. This work is pivotal for the Automated AI Lab concept for several reasons:

1. Architectural Specialization: It proves that AI models are becoming sophisticated enough to serve as precise proxies for biological systems, requiring highly specialized, rather than general, architectures.
2. Causal Testing: Unlike human subjects, these “digital twins” allow for targeted ablations and invasive testing, providing a closed-loop environment where an automated system could theoretically “debug” cognitive disorders by iterating on model designs.
3. The Shift to Vision-Language Synergy: The study highlights the move toward models that process text and pixels in a unified stream, a necessity for the next generation of autonomous agents that Core Automation is currently building.

Industry Implications: The Exodus from Big Tech

The launch of Core Automation is symptomatic of a larger trend in 2026: the “Great Researcher Exodus.” Elite scientists are increasingly fleeing the “compute-first” culture of Google, Meta, and OpenAI for smaller, agile labs that prioritize architectural breakthroughs over raw scaling. The recruitment of Joanne Jang (former GM at OpenAI) and staff from the Gemini and Claude teams suggests that the industry’s most brilliant minds no longer believe that simply “adding another zero” to the parameter count will lead to AGI.

The economic stakes are massive. Core Automation is reportedly seeking funding in the range of $500 million to $1 billion, with a post-money valuation expected to exceed $5 billion. This war chest is not for buying more GPUs—it is for building the software infrastructure that will eventually make those GPUs 100 times more efficient. If Jerry Tworek’s Automated AI Lab succeeds, it will effectively commoditize the very research that currently commands multi-million dollar salaries.

Challenges and the “Catastrophic Forgetting” Barrier

Despite the optimism, the path to a fully Automated AI Lab is fraught with technical hurdles. The most significant is catastrophic forgetting. In a self-evolving system, how do you ensure that the AI doesn’t “evolve away” its core safety protocols or its ability to perform basic tasks as it optimizes for complex ones? Tworek’s team is betting on a “modular reasoning” stack, where different parts of the network are specialized for different tasks, much like the modularity seen in the human cortex.

Furthermore, there is the question of hallucination in automated research. If an AI is designing its own architecture, humans may eventually lose the ability to interpret *why* a certain design is superior. This “black box” problem is being addressed at Core Automation through a commitment to interpretability-by-design, where every evolution step must be accompanied by a human-readable justification and a verifiable proof of performance.

Conclusion: The Dawn of the Self-Evolving Machine

As of April 22, 2026, the artificial intelligence industry is no longer just about building the biggest model; it is about building the smartest Automated AI Lab. The transition from human-led research to AI-driven discovery marks the most significant milestone since the original Transformer paper. With Core Automation leading the charge, and academic centers like EPFL providing the cognitive blueprints, the next generation of AI will be characterized by its ability to learn, adapt, and evolve without human intervention.

The era of static deployments is over. In its place, we are seeing the rise of Ceres and its successors—models that don’t just answer our questions, but actively redesign themselves to understand the world more deeply. For Jerry Tworek and his team, the goal is clear: to create a system where the research never stops, the learning never ends, and the Automated AI Lab becomes the primary engine of human progress.