Sunyan Lee

Curious Person | Former Tech Portfolio Manager

[Sunyan Lee @sunyanlee]

Work is becoming a compression problem.

The important question is: which bits still need to come from you?

That set gets smaller as models get better.

Find those bits.

Push the rest into the model.

[Sunyan Lee @sunyanlee]

@martin_casado assuming the market does not become monopolistic, share gain incentives may offset this

[Sunyan Lee @sunyanlee]

@mgratzer I wonder if we can use skills as a guided template for human learning more broadly. Skills are like a Neo downloading kung fu for AI, but what if it could also be the same thing for humans with AI as the tutor

[Sunyan Lee @sunyanlee]

@max_paperclips why can't this be done with skills?

[Sunyan Lee @sunyanlee]

@fchollet True for a subset of SaaS (mission critical, infra heavy, etc.). But much of SaaS = encoding best practices in a GUI workflow sold as a subscription. What happens when 95% of best practices can be encoded into skills with 0 defensibility?

[Sunyan Lee @sunyanlee]

@mattzcarey Why do we need 2 separate tools for search vs. execute instead of making a single code field that can work with both?

[Sunyan Lee @sunyanlee]

@LakshyAAAgrawal @gepa_ai Underappreciated work. Do you find rubric grading effective even for X/100 type scores vs 0/1? Is there a concern that the scores are uncalibrated and you discard “good” solutions

[Sunyan Lee @sunyanlee]

@abcampbell Bloomberg provides an officially supported Python API for terminal users (https://t.co/m8AId3lIjg). Caveat is data has to stay on your computer. Workbench just acts like a code-assist tool to run BQL/BDP/BDH queries on demand. Hence desktop app only. Can't run this on the cloud.

[Sunyan Lee @sunyanlee]

@abcampbell this gets data from your terminal on your desktop via the Bloomberg (runs any BQL BDP BDH queries)

[Sunyan Lee @sunyanlee]

@abcampbell You can do this today on Workbench (link in profile). Login with ChatGPT to try. Let me know what you think.

[Sunyan Lee @sunyanlee]

@GavinSBaker Both work similarly behind the scenes: an agentic loop w/ tool use. For Claude, the tools hook into the official Microsoft add-ins API for 3P devs. Copilot probably has privileged access to native APIs.

My Q is whether Microsoft will nerf 3P APIs to gain an edge over time.

[Sunyan Lee @sunyanlee]

Code is AGI complete for white collar work

[Sunyan Lee @sunyanlee]

@pmddomingos Only true if you assume the distribution is human data. The whole premise of the RL scaling paradigm is to learn from the world itself.

[Sunyan Lee @sunyanlee]

@corbtt Interesting - intuitively you are capitalizing on the generation/verification gap and specifically making verification even easier by framing as a comparative problem. Does this help for domains where you have robust verification like math/coding?

[Sunyan Lee @sunyanlee]

@natolambert Even original v3 was distilling from both R1 and applied RL (though maybe not direct RLVR?). One apparent difference is maybe the lack of <thinking></thinking> tokens, but even that line blurs if the final responses are growing longer.

[Sunyan Lee @sunyanlee]

@natolambert @benthompson Fundamental problem is misaligned incentives and thus potential erosion of user trust. You see this in the tension for optimizing user experience vs. monetization with Google - do you put the most monetizable links vs. the most relevant links at the top?

[Sunyan Lee @sunyanlee]

@stratechery Amazing. Sam openly 1) admitting the core value is the brand/userbase and 2) calling AI models a commodity