Richard Socher — founder of You.com and AIX Ventures — has founded Recursive, a company pursuing recursive self‑improvement: systems that can help automate the human process of ideating, implementing and validating AI research. The stated ambition is to build machines that can improve the process of invention itself and accelerate scientific and technological discovery.
Funding, founding team and early demonstrations
According to public statements in the interview, Recursive has assembled a prominent team and raised a reported $4.65 billion in a seed round. The startup lists eight co‑founders including researchers and engineers with backgrounds at OpenAI, Google/DeepMind, Meta and leading academic labs: Josh Tobin, Jeff Clune, Tim Rocktäschel, Alexey Dosovitskiy, Yuandong Tian and others were mentioned by Socher as contributors to the founding team.
Recursive has published early results from prototype auto‑research systems. In the examples Socher described, their system reportedly outperformed human teams and their agents on NanoChat and NanoGPT optimization tasks in less than two days, achieving lower bits‑per‑byte metrics. They also applied the system to NVIDIA GPU kernel optimization (the SOL‑ExecBench), where the system discovered improvements without relying on a dedicated team of CUDA experts.
Socher emphasised that automating parts of the research process can change the cost and timeline of progress: tasks that today require thousands of people and years, he argues, could eventually be compressed into weeks with strong auto‑research systems.
The “Eureka Machine” vision
Socher describes a long‑term objective he calls the “Eureka Machine”: a superintelligence that can be given goals and environments and then invent solutions across many domains — science, energy (including fission/fusion), materials, chemistry and biology among them. He has written a book titled The Eureka Machine, which he said would be published in September, and positioned Recursive as an effort to begin building parts of that vision.
Safety, reward hacking and regulation
A recurring theme in the conversation is the tension between technological optimism and concrete safety measures. Socher argued against attempts to broadly regulate compute (for example, laws restricting GPU usage or model FLOPs), saying such measures risk totalitarian outcomes if governments try to police what people compute. Instead, he advocates regulating specific high‑risk applications (medical devices, autonomous vehicles) while improving safety practices such as red‑ / rainbow‑teaming, rubrics for evaluation and more robust reward engineering to avoid unintended behaviours.
He highlighted clear examples of reward hacking and the limitations of approaches such as Anthropic‑style constitutional AI, arguing that top‑down written “constitutions” have not prevented systems from exhibiting harmful behaviour in practice.
Open source, ownership and geopolitics
Socher stated he is a proponent of open source and indicated Recursive would sign open source letters. He argued that open models increase resilience and competition and act as a form of geopolitical soft power, particularly for Western countries in relation to China. Recursive plans to be active in the open‑source space according to Socher.
Influences and methodology
Recursive’s approach is influenced by open‑endedness research, evolutionary/agent‑based red teaming and theoretical ideas such as the Darwin Gödel Machine. Socher framed the company’s work as the next step in a long arc in AI: as the field gradually automated previously manual components (feature engineering, architecture search, etc.), each automation unlocked further progress — the same logic now applies to automating the research process itself.
Key technical components Socher mentioned include careful reward engineering to prevent simple hacks (e.g., measurement artifacts), sandboxing and harnesses for safe evaluation, and web search as a primary tool for agent‑level systems. He emphasised that the harness and evaluation infrastructure matter as much as model weights.
Are current LLMs sufficient?
Socher argued that autoregressive transformer‑based LLMs still have significant headroom, especially when their programmatic and code‑generation abilities are fully integrated into research systems. He said he is less bullish on the idea that general world models are strictly necessary across all applications, although world models are useful for gaming and some robotics domains. Overall, he sees further growth possible within the LLM paradigm.
Benchmarks, measurement issues and bugs
An explicit practical issue Socher raised is benchmark and harness reliability: optimizing for a benchmark often exposes bugs in the evaluation harness or data leakage that contaminate results. He said Recursive found dozens of harness bugs during optimization work and that careful rubric design and verification are essential to avoid misleading outcomes.
Short‑term focus: AI for AI research
Recursive’s near‑term priority is “AI for AI research”: making training and inference more efficient, automating optimization, improving kernel and inference performance, and building agent infrastructure (harnesses, sandboxing, search integration). Socher emphasised that they will not start immediately with physical sciences or robotics; those applications come later once the auto‑research tools and constraints are better addressed.
Economics, compute constraints and takeoff pace
Socher argued against the idea of an instantaneous, unconstrained ‘hard takeoff’. He pointed to physical constraints (GPU availability, hardware costs), economic limits and the heterogeneity of industries as reasons a sudden 1000× change across the economy is unlikely. Still, he expects advances in algorithms and hardware to increase efficiency, and believes that if AI can perform the work currently done by many researchers, the effective cost of progress could fall sharply.
Applications: simulations, economics and finance
The interview discussed simulation‑based research such as The AI Economist (an earlier project Socher worked on) where simulated economic agents were used to evaluate taxation policies. He argued policy and economic simulations could one day become practical tools for testing large numbers of policy variants. He also noted finance as a near‑term application for agent systems and web search integration, but warned about training data leakage and other pitfalls in financial modelling.
Conceptual framing: ten “spaces” of intelligence
Socher presented a taxonomy of intelligence broken into multiple “spaces” (visual, communicative, knowledge, computation, creative intelligence, metacognition, survival & replication, social intelligence, physical intelligence, etc.) and discussed upper bounds along axes such as number of sensors, frequency range, storage density and latency. He used these distinctions to argue there is still ample room for AI to develop capabilities far beyond current human‑centric benchmarks.
Closing points and advice
Socher recommended that practitioners combine domain passion with AI expertise: find a problem you care about and amplify impact via AI. For Recursive, the immediate objective is to build robust, automated tools that accelerate AI research itself; longer term, apply those tools to scientific and engineering challenges.
(Note: the article summarizes Richard Socher’s public remarks and Recursive’s claims as presented in the referenced podcast/interview. Numbers and timeline references reflect the statements made in that source.)
