Our approach to open research and intelligence

Modern AI was built in an unusually open research environment. Many of the ideas behind today’s systems first appeared in public papers, often accompanied by code, model weights, datasets, or enough technical detail for others to build on them. Ideas moved quickly between universities, independent researchers, and competing labs.

Infinite Ascent is a direct beneficiary of that culture. The models we use, the infrastructure we build on, and many of the methods behind our research exist because other researchers chose to make their work public.

That culture has become less common at the frontier. As the economic and strategic value of AI has increased, leading developers have become more selective about what they disclose. Frontier systems are increasingly released as products, with less visibility into their training, data, architecture, and the research that produced them. There are understandable reasons for this. Frontier development has become expensive, competition has intensified, and increasingly capable systems create real safety and security considerations. But the result is that more of the most important work in AI is now happening behind closed doors.

At the same time, much of the recent momentum in open-weight AI development has come from Chinese labs, including the teams behind DeepSeek, Qwen, and Kimi. Their openness is not abstract to us. Our own post-training work builds on open-weight releases from these labs, and the ability to inspect, modify, and train these models materially expands the research that a team like ours can undertake. Infinite Ascent wants to contribute to the same tradition.

AI will accelerate research

Our work focuses on self-improving AI systems, in which AI increasingly conducts research that contributes to the development of subsequent generations of the system. Models can already search large amounts of information, write and test code, compare approaches, analyze results, critique their own work, and coordinate with other agents. As these capabilities improve, AI systems will be able to carry out increasingly complete research cycles themselves.

This changes the relationship between research and openness. As research cycles shorten, the ability to continuously produce improvements becomes a more durable competitive advantage than any individual discovery. At Infinite Ascent, our systems operate through successive cycles of experimentation and improvement, with each generation building on the findings of the last.

The advantage compounds over time as successive generations contribute to better data, evaluations, infrastructure, and the process by which subsequent research is conducted. As these systems advance, earlier methods and findings may become less commercially sensitive while remaining useful to other areas of research. This creates opportunities to share more of our work without compromising the research currently underway.

As open as the work allows

Our aim is to make our research as open as possible, while recognizing that openness is not a binary choice. AI systems consist of many components, from the models and data they use to the architectures, tools, and methods through which they operate. Each can be shared independently, allowing research to become public without requiring the entire system to be released.

Our live paper-trading agent portfolios are one example. Their decisions and performance can be observed publicly, even while the architecture and traces behind them remain internal. As our research advances, we intend to make more of the underlying methods and components available, including technical reports, evaluations, code, and models where appropriate.

Not every part of our work can be made public. Commercial considerations, proprietary data, and safety or security concerns will sometimes require restricting access or publication. These limitations should be considered individually rather than treated as a general argument against openness.

Wherever possible, the starting point should be to consider the most open form of release compatible with the work, rather than assuming that research must remain private.

Open intelligence

As recursive self-improvement begins to accelerate AI research, it could also concentrate the capacity to develop increasingly powerful systems within a small number of organizations. Labs capable of continuously improving their own systems may accumulate a compounding technological advantage, along with growing economic and scientific influence, while the wider research community loses the ability to reproduce or build upon their discoveries.

This is why we believe the next stage of AI research should take place as openly as possible, allowing discoveries to be examined, challenged, and built upon beyond the institutions that originally produced them. The acceleration of intelligence should expand opportunities for progress rather than concentrate them within a handful of laboratories.

Much of what Infinite Ascent is building exists because earlier researchers chose to make their work public. As AI increasingly contributes to its own improvement, we intend to make as much of the resulting research available to the world as possible.