Sakana AI hires Jürgen Schmidhuber, inventor of deep learning, world models, and your next ChatGPT update

【Executive Summary & Core Event】

Sakana AI, the Tokyo-based artificial intelligence startup founded in 2023 by former Google and DeepMind researchers David Ha and Llion Jones, has made a high-profile strategic hire: Jürgen Schmidhuber, the Swiss-based AI researcher widely credited as a co-inventor of Long Short-Term Memory (LSTM) networks and a pioneering voice in world models, artificial curiosity, and self-referential learning. Sakana has bestowed upon him the title of Chief Scientific Advisor, and he will help steer the company's newly formed RSI Lab—dedicated to recursive self-improvement, the concept of AI systems that autonomously enhance their own capabilities through iterative self-modification. The lab's mandate is ambitious: to move beyond the current paradigm of human-engineered training pipelines toward systems that can, in principle, design, evaluate, and improve their own architectures, objectives, and learning algorithms.

Schmidhuber's intellectual pedigree is formidable and contested in equal measure. His 1997 paper with Sepp Hochreiter introduced LSTM, the recurrent architecture that dominated sequence modeling for two decades and powered everything from Apple's Siri to Google's voice search prior to the Transformer era. He has also authored seminal work on world models (1990s), artificial curiosity and intrinsic motivation, meta-learning, and self-referential neural networks that can rewrite their own weights. Sakana's characterization of him as the 'father of modern AI' is a marketing flourish that elides the collaborative and distributed nature of the field's foundations, but it reflects a genuine strategic alignment: Schmidhuber's long-standing theoretical interest in self-improving systems maps directly onto Sakana's research thesis that nature-inspired, evolutionary, and self-modifying approaches represent the next frontier beyond brute-force scaling.

【Technical Architecture & Key Innovations】

The technical premise of the RSI Lab rests on a convergence of several research threads Schmidhuber has championed since the early 1990s. Recursive self-improvement, in its strongest form, describes a system that can modify its own source code, learning algorithm, or architecture to become more capable at modifying itself—a positive feedback loop that, in theory, could yield exponential capability gains. Schmidhuber's 2003 paper on 'Gödel Machines' formalized this concept: a self-referential, self-improving system that can prove theorems about its own behavior and rewrite its code only when it can mathematically demonstrate that the modification will improve a user-defined utility function. This is conceptually distinct from contemporary reinforcement learning from human feedback (RLHF) or automated architecture search (NAS), which optimize within fixed algorithmic frameworks. The Gödel Machine framework requires a system capable of both execution and formal self-analysis—a dual capability that remains largely unrealized in production AI but that Sakana appears positioned to pursue given its prior work on evolutionary model merging and automated scientific discovery.

Sakana's existing technical portfolio provides architectural scaffolding for this ambition. The company has published notable work on 'AI Scientist,' an automated research pipeline that uses large language models to generate hypotheses, write code, run experiments, and produce papers with minimal human intervention. It has also pioneered model-merging techniques inspired by evolutionary biology, combining weights of different models to produce emergent capabilities without full retraining. These approaches, combined with Schmidhuber's world models framework—where a system learns a compressed internal simulation of its environment and can 'dream' synthetic training data—suggest a potential architecture where an AI system iteratively improves its own world model, generates novel training scenarios, evaluates its performance, and modifies its learning strategy. The technical challenges are immense: maintaining stability under self-modification, avoiding capability regressions, ensuring the system's self-evaluation metric remains aligned with human intent, and preventing the kind of reward hacking that plagues even static RL systems. No current production system achieves genuine recursive self-improvement; Sakana's bet is that Schmidhuber's theoretical foundations, combined with modern compute and language model capabilities, can close that gap.

【Industry Context & Competitive Landscape】

This appointment positions Sakana AI in a distinctive niche within the competitive landscape. While OpenAI, Anthropic, Google DeepMind, Meta, and the Chinese labs (DeepSeek, Qwen team at Alibaba, Moonshot) pursue capability gains primarily through scaling laws—more parameters, more data, more compute—Sakana is betting that algorithmic self-improvement could represent a qualitatively different trajectory. This is a contrarian position. The dominant industry consensus, articulated most forcefully by OpenAI's leadership and reflected in the GPT-4 and Gemini training regimes, holds that scaling transformer-based architectures with high-quality data remains the most reliable path to capability gains. Sakana's thesis implies diminishing returns to pure scaling and anticipates that algorithmic innovation in self-modification could yield discontinuous jumps. Schmidhuber himself has been a vocal critic of the narrative that credits modern AI breakthroughs solely to the Transformer, arguing that LSTM and his other foundational contributions enabled the deep learning revolution long before attention mechanisms arrived.

The competitive implications are nuanced. Sakana is not directly competing with frontier labs on benchmark performance or model size; its released models, including the EvoLM series and various merged models, are relatively compact. Instead, the company is pursuing a research-direction bet that, if successful, could be paradigm-shifting. Recursive self-improvement is the holy grail invoked in discussions of artificial general intelligence (AGI) and is frequently cited as a potential pathway to superintelligence. If Sakana demonstrates even partial success—say, a system that can autonomously improve its own training data curation, architecture hyperparameters, or fine-tuning strategies with measurable capability gains—it would represent a significant proof of concept that could attract substantial additional investment and talent. The risk is that the problem proves as intractable as it has historically appeared, and that Schmidhuber's theoretical frameworks, while intellectually elegant, do not translate into production-grade systems under contemporary compute and engineering constraints. DeepMind's AlphaFold and AlphaCode achieved narrow superhuman performance through specialized architectures, but none demonstrated self-improvement; Sakana is aiming for a more fundamental breakthrough.

【Developer & Enterprise Implications】

For developers and enterprises, the immediate practical implications of this hire are limited but directionally significant. Sakana's current commercial offerings—primarily smaller language models and model-merging tools—are available through standard APIs and open-weight releases, and the RSI Lab's work is unlikely to produce deployable products in the near term. However, the hire signals that Sakana is investing in research that could eventually yield automated ML pipelines: systems that reduce or eliminate the need for human ML engineers to design training curricula, tune hyperparameters, or curate datasets. For enterprises spending millions on model fine-tuning and data preparation, the prospect of a system that can optimize these processes autonomously is economically compelling. If Sakana's RSI research produces even incremental automation of the ML development lifecycle, it could lower the cost of custom model development by orders of magnitude.

The integration complexity for any eventual self-improving system would be substantial. Current enterprise AI deployment relies on deterministic, auditable pipelines: models are trained, evaluated, version-controlled, and deployed through well-understood MLOps practices. A system that modifies its own weights or architecture introduces profound challenges for reproducibility, safety auditing, and regulatory compliance. Under the EU AI Act and emerging regulatory frameworks, systems that autonomously evolve their behavior may face heightened scrutiny or outright restrictions in high-risk applications. Enterprises considering adoption of any future Sakana self-improvement technology would need to invest in new evaluation infrastructure capable of monitoring not just model outputs but model evolution trajectories—detecting capability drift, emergent behaviors, and potential alignment degradation over iterative self-modification cycles. The hardware requirements are also non-trivial: recursive self-improvement, if it involves continuous training and evaluation loops, would demand persistent compute clusters rather than the batch-training paradigm that dominates current model development.

【Key Takeaways & Strategic Outlook】

Sakana AI's recruitment of Jürgen Schmidhuber represents one of the most intellectually ambitious bets in the current AI landscape: that recursive self-improvement, grounded in three decades of theoretical work, can be operationalized with contemporary techniques and compute. The formation of the RSI Lab signals that Sakana is positioning itself not as another frontier model provider competing on benchmark scores, but as a research-oriented organization pursuing a potentially paradigm-shifting approach to AI development. Schmidhuber's track record—LSTM, world models, artificial curiosity—provides genuine intellectual credibility, though his most ambitious theoretical constructs, particularly the Gödel Machine, remain unrealized in practice. The key strategic question is whether the gap between theory and implementation can be closed through the combination of modern language models, evolutionary search techniques, and Schmidhuber's architectural insights.

The broader industry should watch this development as an indicator of whether the field is approaching an inflection point where pure scaling encounters diminishing returns and algorithmic innovation reclaims centrality. If Sakana demonstrates that AI systems can meaningfully improve their own capabilities through autonomous iteration, it would validate a research direction that has been peripheral to the mainstream since the deep learning era began. Even partial success—automated architecture improvement, self-curating training data, or self-tuning learning algorithms—would have significant commercial implications. The risk for Sakana is that the problem remains as theoretically tractable but practically elusive as it has for thirty years. For the AI community at large, this hire re-centers questions about the long-term trajectory of capability gains and whether the next major leap comes from more compute or from fundamentally different approaches to how AI systems learn to learn.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.