← 2026 Work Log

Work Log

2026-04-23

It’s been almost another month since the last entry, time flies so fast 😭

I’ve almost forgotten how to write, still just jotting down whatever comes to mind.

  1. Workstation
    I applied for Suzhou’s Leading Talent Program, and through that, I got to know Senior Brother Liang, who is in charge of investment attraction at SISPARK. He is three years ahead of me from our university. He introduced us to the OPC shared workspace at the Dushu Lake Digital Economy Industrial Park, which is near my home. The first three months are free, and starting from the fourth month, the pricing is 500 CNY/person/month, though I heard it might still be free.

    After PGW heard about this, he was very interested and also joined. So now in Suzhou, there are three of us: me, cz, and pgw, each with our own workstation. Today is already the fourth day of working here. It feels a bit strange to be back at an office. As cz put it, “it feels nostalgic.” For me, it helps to clearly separate working from non-working states, making my focus sharper. I think that’s a good thing. Another benefit is that working alongside colleagues facilitates the exchange of ideas and sparks collaboration. Closer cooperation means a more unified product.

    As for other aspects, we’ll see how it goes.

  2. Iterating on legion-mind
    I feel that legion-mind tends to go off track easily, lacking a brainstorming capability. So I cloned Superpower and studied it thoroughly, absorbing its advanced concepts to iterate on legion-mind.

    This has significantly slowed my progress on everything else, because I imagine myself as a lumberjack, and legion-mind is my axe. Now that I’m sharpening the axe, I can’t efficiently cut trees. Often, I’m uncertain how a particular modification will affect the overall capability of the harness tool.

    Hence, I need a benchmark. I previously considered using the SWE benchmark, but that felt off track—I’m not testing the LLM’s tool-use ability; the LLM already handles that. What I need is to construct a few typical examples that I understand, then evaluate them in terms of stability and creativity.

    Lately, I’ve frequently seen the idea of asking an LLM to generate a script for a “pelican riding a bicycle” GIF. So I’ll use that as an interview prompt:

    "Using TypeScript, generate a 512x512, 4-second, 24fps, seamlessly looping GIF: a pelican riding a bicycle along a coastal road. It should have three background layers: sky, sea, and road. The wheels must keep spinning, the pedaling motion synchronized with wheel speed, with a slight up-and-down bob of the body and a subtle head sway. Support --seed and --out parameters."

    Then, let the LLM use this harness tool to build a small working product—understand the requirements, design, implement, test, all without bothering me, and finally check the results.

    Next, there’s the LLM’s ability to understand and process data—like parsing a CSV file and extracting insights? But I haven’t fully defined this yet; it might again be testing the LLM’s capability rather than the harness tool’s.

    Then come bug fixes and feature additions.

    In short, coming up with test cases is tough. Previously, I tested the harness in real projects, hoping for a two-birds-with-one-stone effect. But that essentially relies on nothing going wrong. I need a more scientific approach.

    cz is working on AIM (AI Project Manager), which is exactly the direction legion was heading. I plan to embed legion as the harness layer for AIM.

  3. Other things
    There are many other things I want to do but haven’t started yet: one is the AI-powered operation of 1earn, and the other is PMA sports arbitrage. I previously thought that with AI assistance, I could handle both simultaneously. But now I realize that what’s truly scarce is my attention. When juggling multiple important tasks, I risk neglecting one for another. As I wrote in earlier diary entries, I can realistically only manage 2-3 things at a time.

    In an AI-enhanced era, what matters most is managing human attention. I likely need to actively prune, focusing my attention only on a few tasks, not too many at once, and close out tasks before starting new ones.

    I’ve already set OKRs, but the review process is slow. Sometimes I’m just too lazy, or I crave that “perfect working state” so much that I slack off when it’s not perfect. I need to break these self-imposed constraints with schedules and persistence—like the old guy who, at 82, still went to the gym at 4:30 AM every day, “working through the pain.”