Core Automation

Jerry Tworek is a mathematician who earned an MSc in Mathematics from the University of Warsaw after obtaining a Bachelor of Science in Applied Mathematics there. Prior to entering AI research, he spent the first five years of his career in the hedge fund industry as a quantitative trader, developing investment strategies in futures markets using optimization theory and advanced signal extraction techniques from noisy datasets. This experience led him to deep reinforcement learning; he joined OpenAI around 2019 and spent nearly seven years there, rising to Vice President of Research where he led work on reinforcement learning and reasoning models including o1 and o3.

X+4

Rohan Anil holds a Master's degree in Computer Science from UC San Diego and an MSc(tech) in Information Systems from the Birla Institute of Technology and Science, Pilani. He previously served as Distinguished Engineer at Google DeepMind, where he led significant work on the Gemini models including pretraining, architectures, and optimization for large language models such as PaLM-2 and Gemma; he also held a Distinguished Engineer role at Meta. He later joined Anthropic as a Member of Technical Staff before co-founding Core Automation.

X+4

Joanne Jang earned a Bachelor of Science in Applied Mathematics and a Master of Science in Computer Science from Stanford University. Prior to OpenAI, she worked on natural language understanding at Google Assistant, focusing on uncertainty modeling and conversational mechanics. At OpenAI she served as Product Lead for DALL·E, Head of Model Behavior, and General Manager of OpenAI Labs, where she founded and led a research team prototyping new interfaces for people and AI.

LinkedIn+3

Anmol Gulati earned an Integrated Masters degree in Mathematics and Computer Science from the Indian Institute of Technology, Kharagpur. He previously worked at Google DeepMind as Tech Lead for Agents Research on the Gemini project and was a co-founder of Adept AI; his earlier experience includes software engineering internships at Google and Directi, as well as contributions to projects like Gensim.

LinkedIn+2

Julia Villagra attended Boston University. She spent nearly 16 years at Hudson River Trading, rising to Head of People and leading recruiting and people operations starting in 2008. She then joined OpenAI in February 2024 as Head of Human Resources and was promoted to Chief People Officer before departing to co-found Core Automation.

LinkedIn+2

Adaption Labs develops adaptive AI systems centered on continual and gradient-free learning to enable models that evolve from real-world interaction rather than relying on periodic retraining or massive pretraining runs. The company targets efficiency through malleable datasets, adaptive intelligence modules, and interfaces that reduce the need for prompt engineering, positioning its approach as a direct alternative to the dominant scaling paradigm. This creates strong overlap with Core Automation in the shared emphasis on post-training adaptation, reduced data and compute demands, and elements of automated research workflows such as experiment orchestration. As a pre-revenue neolab, its traction stems from founder expertise in efficiency and adaptation drawn from prior roles at leading labs and rapid capital raises that support focused research on these paradigms. Durable strengths include a structural bet on lifelong learning architectures that could compound capabilities over time without repeated full retraining cycles, alongside talent concentration in gradient-free methods. Constraints arise from the high technical risk of proving continual learning at frontier scales and dependence on recruiting specialized researchers in a competitive talent market. The business model aligns incentives around pure research progress toward adaptable systems rather than immediate productization.

Adaption Labs+2

Ineffable Intelligence pursues superintelligence through reinforcement learning systems designed to acquire knowledge and skills directly from agentic experience rather than static datasets. Founded by a leading RL researcher, the lab emphasizes building superlearners capable of progressing from basic motor skills to complex intellectual tasks via interaction-driven updates. This overlaps substantially with Core Automation's focus on continual learning from real-world feedback and architectures that move beyond traditional pretraining and RL paradigms. As an early-stage neolab, its momentum derives from founder pedigree, substantial seed funding, and recruitment of talent from established labs to explore experience-based intelligence. Structural advantages include deep expertise in RL techniques that enable ongoing adaptation without full model retraining, supporting a durable edge in paradigms prioritizing interaction over scale. Key limitations involve the execution challenges of scaling RL to superintelligent levels and reliance on a concentrated team of specialists amid intense competition for such expertise. The singular research focus mirrors Core's automated-lab vision by prioritizing fundamental breakthroughs over diversified product development.

Ineffable Intelligence+2

Reflection AI builds agentic systems, including autonomous coding agents like Asimov, that index and reason over entire codebases and institutional knowledge to automate software development and related computational tasks. The lab combines RL, LLMs, and agent architectures to enable self-directed code writing, debugging, and evolution, targeting the automation of knowledge work as a foundational step toward broader superintelligent capabilities. This aligns closely with Core Automation's agent-centric small-team model and emphasis on automating research and complex processes through capable AI systems. Traction as a neolab comes from significant early funding rounds and positioning as an open frontier alternative, supported by recruitment of researchers experienced in agentic and RL approaches. Durable strengths lie in the structural bet on agentic automation unlocking recursive improvements in software and research productivity, with open models potentially broadening distribution. Weaknesses include the nascent state of reliable long-horizon autonomous agents and dependence on continued talent inflows in a crowded field. The pure-research orientation reinforces overlap by prioritizing paradigm-level advances in automation over near-term commercial applications.

Reflection AI+2

Thinking Machines Lab develops frontier AI systems optimized for human-AI collaboration, customizability, and multimodal adaptability to make advanced models more accessible and effective across diverse applications. The lab explores architectures and interfaces that enhance general capabilities while enabling users to tailor behavior without extensive retraining. Overlap with Core Automation arises through the shared pursuit of adaptable systems and agent-like collaboration that could support automated research workflows in small teams. As a prominent neolab, its position benefits from high-profile founders from OpenAI and rapid capital attraction that funds ambitious research agendas. Enduring strengths include expertise in multimodal and collaborative paradigms that structurally favor flexible, user-aligned intelligence over rigid scaling. Challenges stem from execution risks in translating research into reliable collaborative agents and heavy reliance on elite researcher networks. The model of focused, well-capitalized research labs competing on novel approaches directly parallels Core's strategy.

Thinking Machines Lab+2

Safe Superintelligence Inc. operates as a dedicated research lab with the singular mission of developing safe superintelligence through integrated advances in capabilities and alignment, treating both as core engineering challenges. The company pursues revolutionary breakthroughs in model architectures and training methods without intermediate product distractions. This creates meaningful overlap with Core Automation via the common focus on next-generation paradigms that surpass current transformer and pretraining limitations in pursuit of advanced, reliable AI systems. Its neolab structure draws strength from founder reputation and aligned capital that enables long-horizon research on safety-capability co-development. Durable advantages include a streamlined organizational model that concentrates resources on foundational technical problems, reducing dilution from commercial pressures. Potential constraints involve the extreme difficulty of the superintelligence goal and dependence on a small, high-caliber team in an environment of rapid talent mobility. The straight-shot research ethos mirrors Core's automated-lab bet by emphasizing pure progress on paradigm-shifting intelligence.

Safe Superintelligence Inc.+2

Core Automation faces severe key-person risk due to its reliance on a tiny group of named researchers led by CEO and co-founder Jerry Tworek, who departed OpenAI after nearly seven years as VP of research focused on reinforcement learning and reasoning models in January 2026, with the company founded in late March 2026. Co-founders and early recruits include Rohan Anil (ex-Anthropic and Google DeepMind), Joanne Jang (ex-OpenAI model behaviour lead), Anmol Gulati (Google DeepMind), and Julia Villagra (former OpenAI Chief People Officer), forming the core of an 11-50 employee team explicitly described on the company website as a small lab rethinking neural architectures from scratch. This concentration means the loss or distraction of any one of these individuals—through departure, counter-recruitment, or competing demands—would directly impair execution of the core bet on automating research itself and developing continual learning systems beyond transformers and scaling. The company's own site stresses that its approach depends on "small teams with highly capable agents" and personnel who have built frontier models and optimization methods, underscoring structural fragility rather than scalable depth. No evidence exists of succession planning, broad bench strength, or contractual retention mechanisms that would buffer against this in a talent-competitive field. This risk is amplified by the explicit recruitment narrative where Tworek "nerdsniped" talent from rivals, creating reciprocal vulnerability.

Core Automation+3

Core Automation operates with no shipped product or revenue despite raising $100 million at a $1 billion valuation shortly after its late March 2026 founding, while pursuing compute-intensive frontier AI research that requires substantial ongoing infrastructure investment to compete with established labs. The company website outlines an ambitious program to automate its own research and develop new learning algorithms and architectures superseding large-scale pretraining, reinforcement learning, and transformers, yet reports confirm it remains far from any release. This creates a structural mismatch where burn rate for talent, compute, and experimentation could rapidly deplete runway without demonstrated milestones, especially as the firm targets an additional $300-500 million (with some materials citing up to $1 billion) at a $4 billion valuation within weeks of the initial close. Competitors it seeks to surpass already maintain giant compute budgets and deployed systems, heightening the execution pressure on an unproven timeline. The rapid back-to-back fundraising underscores investor appetite but also exposes the firm to any slippage in progress that could trigger funding gaps or valuation resets. No offsetting customer contracts, revenue streams, or phased product validations exist to mitigate this capital and timing asymmetry.

Dealroom.co+3

Core Automation's aggressive recruitment of senior researchers from direct competitors exposes it to ongoing talent retention and counter-poaching risks in an industry where mobility is high and compensation is escalating. Specific moves include Rohan Anil departing Anthropic after being "nerdsniped" by Tworek, alongside inflows from Google DeepMind (Anmol Gulati and others) and OpenAI alumni, building a team of 11-50 that the company positions as its primary differentiator for rethinking AI fundamentals. This strategy creates structural vulnerability because the same competitive dynamics that enabled these hires can reverse, with rivals offering equity, resources, or stability to lure back personnel critical to the small-lab model. The company's emphasis on a compact team with deep expertise in frontier models, optimization, and systems engineering means each departure carries outsized impact on research velocity and the self-automating lab vision. No public details indicate long-term vesting cliffs, non-competes, or cultural mechanisms that would durably lock in this talent against market pressures. This risk is inherent to the talent-war environment documented in the recruitment reports and remains unmitigated by scale or diversification.

Business Insider+3

Core Automation is pursuing an unvalidated technical approach centered on new learning algorithms and architectures that the company states will supersede current scaling paradigms of larger models, more data, and static deployment, including architectures that scale better than transformers and systems enabling continual learning from real-world interaction. The website explicitly rejects the prevailing recipe and bets on automating research itself to enable small teams to achieve breakthroughs that once required entire organizations, yet no prototypes, benchmarks, patents, or external validations support feasibility as of the July 2026 timeframe. This creates material execution uncertainty because frontier AI research has high failure rates, and the firm's $100 million seed plus targeted follow-on capital assumes rapid progress on this contrarian path against incumbents with proven infrastructure. The small-team, high-automation model amplifies downside if the core hypothesis—that self-optimizing agents can drive the next wave—does not materialize quickly. No concrete milestones, third-party collaborations, or phased results are disclosed to provide any buffer against scientific or engineering setbacks inherent to rethinking gradient descent and model development fundamentals.

Core Automation+2