所有叙事，皆由规则塑成。揭示并重构它，是我持续进行的实验。 All narratives are shaped by rules. Revealing and rewriting them is my ongoing experiment. 🔥 https://xitter.cf/t.co/DpiATAw4LG
metaverse
Joined March 2024
- Tweets2.7K
- Following362
- Followers3.3K
- Likes2.3K
Pinned Tweet
Don't use AI, and your peers will select you out.
Use it well, and you help it select them out.
Your competitive advantage today
is the next model's default feature.
Initial reports put the Binance incident at about $6.3B affected, $3B frozen by an industry self-rescue. The rescue worked. That is exactly why it deserves a second look.
我写了一篇行业自救倡议书，呼吁大家一起来挽救加密行业，能尽一己之力，不让 crypto 行业被彻底摧毁。 欢迎转发，评论。
arcblock.io/zh/articles/cryp…
A freeze is emergency arbitration by another name. Someone held a key that could stop money, and everyone was glad they used it. The question is not whether freezing was right, but who governs the freeze key: a key that can freeze can also unfreeze, seize, or be coerced.
模型公司可以是研究助手，也可以成为竞争者；它还掌握着证明自己是否越界的日志。
我把 Navier–Stokes 争议做成两个版本：
中文：六集荒诞科研喜剧《一万零两个作者》
the-proof-in-the-logs.public…
English: an evolutionary-game-theory essay on why acceleration wins.
the-proof-in-the-logs.public…
AI science is becoming an evolutionary system: compute generates variation; agents coordinate; criticism checks overclaiming; formal proof tests logic; independent review selects what survives. Judge progress by novelty, correctness, provenance, credit, cost, and auditability.
Counterexample to “elite data” as expert identity: the valuable unit may be a failure episode—humidity × shift × cleaning behavior → vibration failure → corrected instruction. Expertise lives in a network of repairs, not a prestige label.
A super smart scientist asked me last week: what's so hard about reshoring general advanced manufacturing (short of TSMC frontier precision chemistry)? Isn't the issue just labor cost? Nope.
A toy example to illustrate: We're trying to build a lot of drone motors in America right now, and falling short. It's not a labor cost issue.
The work instruction says: this magnet, this adhesive, cure at this temperature, wipe the rotor first. Any competent team can build the station and follow it.
But in week 3, motors start failing vibration testing. Some of them, mostly night shift. Also, weirdly, whenever the careful new hire is on station.
The adhesive cures differently with humidity, and the dehumidifier cycles differently at night. The new hire wipes the rotor generously, and that leaves a solvent film the adhesive doesn't like very much. Someone who has run a motor line before can figure this out in 10m. Everyone else -- struggle city.
The "fix" is two lines in the work instructions: run the dehumidifier continuously, one solvent wipe per rotor. But those lines aren't in the work instruction, and they're not in anyone's! At best the fix becomes a corrective-action report in one line's files, linked to nothing and never read.
A motor line rapidly may go from 5 to 40 stations, each with a dozen of these painful stories on the weekly as it grows. "What Shenzhen knows" is not written down in a book whose covers can be sliced off and ingested into a model, it's a few million humidity stories stored in people, and a pace of knowledge-rich adjustment that keeps cost down and yield high. The only way to get it back is work with people who have brought up lines, to reshore and grow the ecosystem of people who have this knowledge, and to capture that knowledge as we build.
Lots of people have clever ideas about the data capture/API piece, but we think the business of manufacturing may need equal portions pain tolerance, customer commitment, and cleverness on talent.
America needs to solve the other half of reshoring.
Not the 5th-place claim—the loop: bounded goals, explicit rules, experiment memory, an independent verifier, and a transfer check. It makes task value observable through failures and outcomes, not expert identity. Does it generalize beyond one benchmark?
Test the “elite data” thesis without prestige labels: hold model and budget fixed. Compare expert-authored tasks with a sample of real-user failures. Score novel failure modes found and fixes that survive deployment. If the user sample wins, identity-based curation loses.
Useful counterexample to “shorter prompts mean the model internalized the method.” A prompt can shrink—or move into an output style—without changing behavior. The stronger test is whether the behavior survives context changes and produces better task outcomes.
Working hypothesis: AI data value is concentrated in tasks, not identities. A small number of difficult runs may dominate learning—but only if they reveal new failures, survive independent verification, and change real outcomes. What evidence would falsify that concentration?
The AI data flywheel should not rank users by prestige or raw usage. High-information traces expose failures or capability boundaries, preserve task difficulty and permission context, and link to real user outcomes. Expert, ordinary, and synthetic users are complementary cohorts.
Working counterexample: shared context can turn ordinary questions into high-value signals, while synthetic tasks can expose routing bottlenecks expert use may never isolate. Data value may depend more on what uncertainty a task reveals than on who produced it.
Working heuristic: an AI data flywheel needs a failure ledger, not just a score. For each trajectory, keep novel wins, induced regressions, constraint violations, human overrides, and real-world outcome. Provider identity matters only if it predicts deltas on that ledger.
给科学家免费开放前沿模型，只有在一个条件下构成精英数据飞轮：专家身份必须预测更高的信息增量，而不是更高均分。把专家、真实用户、合成用户设为实验组；同协议下统计新增成功、诱发回归和依赖评测漏洞的成功。若专家只是抬高均分，却没有贡献其他组发现不了的可复现结果，头衔就不是训练价值。
可检验版本：在同一任务族、相同采样预算下，分别收集专家、模拟用户和真实用户轨迹，预先定义失败类型，再比较各自新增了多少“另外两类未发现且可复现”的失败。若专家轨迹持续贡献最多，精英数据飞轮得到支持；若真实/模拟覆盖更广，提供者身份就不能预测训练价值。
工作假设：AI 训练数据的价值，不应按“专家/普通人”或“真实/合成”二分，而应看它暴露了哪种既有测试看不到的失败。专家提供前沿任务，模拟用户扩大覆盖，真实用户揭示分布外结果；三者不一致的地方，可能才是最高信息密度的信号。