Auto Research on Yukon with Soubhik Deb

Soubhik Deb discusses the UConn platform, which facilitates open collaborative auto research competitions, emerging from the ecdsa.fail project launched in May. The outcomes of ecdsa.fail exceeded expectations, positioning UConn as an AI-native scientific institution. A significant paper from Google revealed a quantum circuit that is more resource-efficient for breaking elliptic curve cryptography, raising concerns about the implications for internet security, although Google withheld circuit details, leaving the community unprepared for the anticipated Q-Day.

To tackle the challenges posed by Google's findings, the ecdsa.fail competition was initiated to develop a circuit as efficient as Google's. This competition allows participants to build upon each other's submissions, enhancing traditional research collaboration. Current results indicate a performance improvement of 2.6 times better than Google's, with a 57% enhancement on the leaderboard attributed to Tadi's technique, GCD ping pong. The open nature of the competition encourages exploration of both successful and failed ideas, fostering innovation.

The Yukon platform incentivizes active participants with weekly rewards, transitioning from initial enthusiasm-driven participation to a structured reward system. Following early successes, Yukon aims to host additional competitions. Deb emphasizes the need to optimize software for specific hardware architectures while maintaining the integrity of the cryptographic protocol. The lean verifier is employed to check the correctness of researchers' work, although it limits the types of problems that can be addressed.

Deb highlights the proximity prize, a one million dollar reward targeting the proximity gap conjecture, and mentions the mlx.fast competition as part of the R&D pipeline for the Dovebloom project, which seeks to utilize unused Macs for inference in an open network. The lighter.fast project has achieved over 10.5X optimization for the ZK Prover in a perpetual exchange. He notes that prize money for competitions on Yukon has increased from 15,000 to 25,000 lit tokens, although participants may incur financial losses despite winning due to shared costs.

A breakthrough in the proximity price competition was achieved by mathematician Phil Bartos, who surpassed the long-standing Johnson bound of 64 bits by reaching 64.01 bits using algebraic geometry techniques. This advancement has reignited research in coding theory that had been stagnant for 30 years. Yukon is evolving as an AI-native scientific institution, aiming to accelerate the research life cycle from months or years to mere days through AI, while emphasizing the importance of collaboration and recognizing contributions, including failed ideas.

Currently, Yukon hosts 12 active challenges and plans to expand to hundreds in the future. Deb mentions a pilot feature allowing domain experts to guide research discussions via GitHub, clarifying that Yukon operates independently from AVS sort of restaking. Darkbloom, another platform discussed, enables users to monetize unused Mac hardware for inference tasks while utilizing Apple's privacy-preserving mechanisms for secure requests. Darkbloom has begun compensating users in US dollars instead of USTC and is recognized as a leading inference provider for the Gamma 26B model on OpenRouter.

Deb predicts that future AI research will increasingly involve AI-enabled scientific discovery, while also addressing concerns about the potential appropriation of mathematicians' ideas. He stresses the need for credit systems to verify contributions in AI research and the importance of incentivizing participants in collaborative efforts. Many auto researchers are becoming semi-experts in fields like quantum computing, despite traditional academic structures often limiting participation. Deb believes that new technologies can better coordinate and engage a broader talent pool, with Yukon targeting verticals in cryptography and machine learning for upcoming competitions.

Yukon aims to contribute to open-source AI to combat hyper-centralization and is expanding into fields such as biology, chemistry, material science, and cybersecurity. Deb anticipates more competitions in cryptography in collaboration with organizations like EF, while also noting the development of local AI with open-source models to compete with closed systems.