highly recommend marrying your best friend

life update

All you need is Modu

I’ll take it one step further: each model is best optimized with an entirely different harness

Shameless plug: this is @askModuAI Unified access to the frontier coding agents like cline/codex/claude code/etc with no manual config/overhead and switching costs

Codex is now really good with gpt5 But you have to be much more intentional with prompting, and task execution is much slower than claude code w/sonnet 4 Very clear that we've already entered a world with different "tastes"/"flavors" of the same solutions

Elite ball knowledge is knowing that Buffalo Wild Wings has easily top 3 fast food chain burgers

Gabriel and I started this business programming in a tiny office for a year straight. We started with the idea that AI systems must be able to learn from real-world experience to solve real-world problems. Eventually it became clear that this was much more of a systems and data problem than a machine learning problem. The solution needed to be industrial-grade software that was freely available and easy to adopt for individual developers but could scale to fit the needs of large organizations. So we open-sourced the project and told our earliest users to stop paying us. It felt crazy. These days, we're rounding a corner. We have a small but incredible technical team: Aaron (Rust compiler maintainer, Svix, AWS), @anndvision (Columbia postdoc, Oxford PhD), and Alan (CMU PhD, VP at JPM AI Research) soon to be joined by Shuyang (staff SWE on LLM infra at Google, Palantir) and Cole (Cognition, Windsurf, Stanford). Our community is active and growing (soon to 10k stars!). There is a clear path towards building the agent that optimizes every TensorZero deployment against the real-world feedback it collects. If you told me when I started my PhD that one day lots of companies would voluntarily start storing RL trajectories in a data model that I helped build so that we could optimize their policies against the rewards they care about I would have been astonished. Soon, this won't be done by hand.

I'm very confident that we've accumulated the largest set of IRL task-based evals for coding agents like codex, claude code, cursor, amp, devin, etc. over the past few weeks with @askModuAI Need to figure out a way to benchmarking publicly accessible

Paradigm is the AI-native spreadsheet to eliminate menial work. Thousands of users have saved 10,000+ hours with Paradigm, and you can be next. Get your first month free today, then plans start at just $20/month.