笔记
简短的笔记:用一段话或一页纸,写一个想法、一篇论文,或一个我仍在思考的问题。其中一些会长成文章。
2026
-
笔记Memory that dreams
A colleague showed me that agent memory should dream before Anthropic published on it. Replay the day, and check the notes against what happened.
-
笔记A self-fulfilling prophecy
Karen Hao says scaling laws are like Moore's Law: not a law of nature but a self-fulfilling prophecy. She is half right, and the half matters.
-
笔记An emergent property of matter
Sam Altman said intelligence may be an emergent property of matter, like a rule of physics. Karen Hao calls it hype. I share the conviction.
-
笔记How CRISPR works
Bacteria keep mug shots of the viruses that attacked them, and Cas9 uses them to cut. A 3D model to step through, from the discovery to the tool.
-
笔记Intelligence is compression
Ilya Sutskever's internal slogan at OpenAI: a model that must predict the world has to compress it. I think it is the most useful definition we have.
-
笔记Its magic crumbles away
Weizenbaum said a program's magic crumbles once it is explained. Karen Hao ends her book on it. True of ELIZA, not of language models.
-
笔记Small, focused, high-trust
Altman wanted OpenAI to stay small. Every large group spends more of its work holding itself together, and AI may change that arithmetic.
-
笔记Something closer to a religion
In 2013 Sam Altman wrote that great founders create something closer to a religion. Karen Hao reads it as a warning; I read it as a forecast.
-
笔记I wrote a short book about peace
The Art of Peace is a short web book on living together without thinking alike: thirteen chapters, 104 passages, free to read.
-
笔记One person can make a difference
Sam Altman had a line from Admiral Rickover painted on OpenAI's walls. It says one person can make a difference, and it was used to ask for unity.
-
笔记The interface that disappears
OpenAI's founders kept pointing to the film Her as a picture of AGI. What they described is the north star for every interface: natural enough to vanish.
-
笔记A billion without a plan
Large language models were funded before anyone could say what they were for. That took a dense pool of capital and a few people deciding where it went.
-
笔记In vivo, not in vitro
An open question: can intelligence be measured anywhere but in performance? My hunch is that it shows only when an agent acts.
-
笔记Who carries progress
A colleague says the breakthroughs come from academia and the state. I think they come from people who have to make things work, and that what they learn spreads.
-
笔记My GOAT games
My twenty-five greatest games of all time as one chart, from Phantasy Star Online to Anno 1602.
-
笔记Future Cop: L.A.P.D., a duel between genres
Future Cop: L.A.P.D. had a one-against-one mode where you were the soldier and the general at once. I have not found a game since that does it.
-
笔记Gangsters: Organized Crime, orders for the week
Gangsters: Organized Crime had you plan a week of orders and then watch the city carry them out. I have not played anything like it since.
-
笔记Who contributed to AI
A companion piece to Karen Hao's Empire of AI and Sebastian Mallaby's The Infinity Machine: who built AI, where, and which idea led to which product.
-
笔记Everything is a fractal
The idea I have carried longest: the same patterns recur at every scale. When something holds at three scales, it is usually something real about the world.
-
笔记A capability, democratised
Building software took money, a team, or years of training. Now anyone with an idea and a subscription can do it, and nobody predicted what would follow.
-
笔记Citizen entrepreneurs
My friend's answer to the transition: with AI, ordinary people can find a business opportunity, build the business, and run it. He built an idea factory.
-
笔记Ideas stop waiting
In 2026 the backlog of ideas I carried for years is turning into things that exist. On why this year feels like a window that will not stay open.
-
笔记Power, in the good sense
The right word for what AI gives an individual builder is power, meant as capability rather than domination. It is refreshing to say so.
-
笔记The move to the city
Urbanisation was miserable for the people who lived through it and better for everyone who came after. The argument about AI and work has the same shape.
-
笔记The tumult in between
If AI ends in post-scarcity, the dangerous part is the interval: jobs going before goods become nearly free, with no agreed idea of where the ship steers.
-
笔记The window before the utility
Tokens will be a metered utility and AI as ordinary as electricity. Until then, learning to use it well is an advantage that will not come around again.
-
笔记Opus 5.5 is the best Opus since 4.5
A month in, Claude Opus 5.5 feels like the next milestone after Opus 4.5: the same jump in what Claude Code can carry, now faster and cheaper per task.