Recursive Self Improvement Hack: Evals

Sundai: - a room full of people building moonshots - every sunday.- born out of MIT & Harvard. 135+ hacks. - aspire to beat the norm. aim higher. ------the loop is already spinning. More than 80% of Anthropic's code is written by Claude. Jeff Dean left Google after 30 years to build the loop for science. Direction-setting is the role left for humans. the hard part: Lilian Weng's survey of harness engineering argues the near-term path to RSI runs through the harness rather than weight-rewriting. A self-improving meta-harness is within reach. The bottleneck is evals, because a loop is only as good as its evaluator. That's the hack. in the room: researchers from MIT CSAIL, Harvard, DeepMind and Meta Reality Labs. Co-hosted by Autolab and Maritime (MIT spinouts), EdotEnv (YC S26), YN.app and Sundai.------We will build agents that recursively self improve, or train models using autoresearch principles. ✨Schedule✨ 10:00 Intros 10:15 Idea pitching and team formation 14:00 Check in #1 and lunch 20:00 Check in #2 and presentations ✨What is Autolab✨ A platform for Autoresearch. Autolab.ai runs your ML experiments for you - agents plan, launch, and score thousands of training runs in parallel, on your GPUs, against your evals. ✨What is Sundai✨ Sundai is the largest AI hacker community at Harvard & MIT, now building weekly moonshots in SF. Every sunday we ask each attendee to work on a aspirational larger than life moonshot idea that they always wanted to work on. One sunday at a time. This is your chance to push yourself to next level. We had done 130+ hacks back-to-back in the last 2 years - one for every sunday without stopping ✨What is Maritime✨ Maritime.sh makes it easy to host personal AI agents, agentic apps, and specialized agents for groups of users. You can build an app where every user gets their own dedicated agent, managed and hosted through Maritime, with simple tooling like the Maritime CLI and SDK. ✨What is E.env✨ E.env builds self-improving RL envs to teach LLMs general research skills from Quant Research tasks. Market data is a great source of hard data science tasks and is self improving: as models trade better, the market becomes more efficient, which makes trading well harder again. E.env provides RL envs for post-training and eval services for agents. ✨Our mission at Sundai ✨ "Push the envelope of what is possible to do with AI in a day." 1. Grow the Best Hackers learn to build fast with AI 2. Build things that matter - measure in users $$$ / github ⭐️s 3. Spread the Hacker Mindset teach the skills that we learn ✨Our Principles✨ Build. Ship. Shortcuts - Build: The purpose of Sundai is to get together and build moonshots. Every Sunday. - Ship: get off the localhost. Deploy your learnings as a webapp or a post and a demo that works on the internet. Go for less, to do more. - Shortcuts: Shortcuts is an ingenious move that a hacker makes to build a system that works. In order to build fast - you always need to think in shortcuts. ✨History of Hacking at MIT✨ https://www.sundai.foundation/about/hacker Terms of Participation: By joining this event you agree to take care & vigilance of the following: - leave the room as you found it. All whiteboards, chairs and tables must be cleaned and re-arranged as you found them. All trash must be in the bin. - do not let any strangers enter into any building behind you where sundai is taking place. if you see anyone you do not recognize who is suspicious, please let us know immediately. - use only the dedicated spaces to a sundai hack. if unsure - ask explicitly. - be respectful and considerate of others who may be in the building and are studying or working. - if unsure - always ask. You agree to these rules by coming to the event. Please read full terms carefully here: https://tinyurl.com/sundaiterms