Observations

There is leverage in having technical knowledge and learning faster than others, even if everyone has superintelligent AI.

In the long term, all the ambitious ideas I am interested in will be absorbed by the labs. In the short term, a small number of them will be solved by a startup.

90% of the notes I've written are outdated, either by (a) my own updated worldview, which stems largely from what we now know about AI, or (b) my new ability to iterate/validate ideas with AI. This is incredible, and I couldn't have dreamed of it 10 years ago.

Rocketships with smart people are intrinsically valuable and tend to have good outcomes. If one knocks, join it next time.

This is the first time in history a person can create any piece of software, and we're starting to see this for video games and videos. This has primarily led to an acceleration of creation on existing platforms, and it will probably continue that way. AI concentrates (next point).

AI concentrates in almost every way. Existing tools, marketplaces, platforms all benefit from AI because it is trained on them, they have the existing distribution, and they can quickly extinguish new competition by iterating using AI. Labs will solve a greater percentage of the world's problems, and naturally talent will concentrate at them. Increasing competition will cause startups to occupy shorter timelines, with more capital involved, and with bigger Pareto distributions.

AI will soon roughly have a human-like inner monologue at any task, including social ones and physical ones (like moving with dexterity). This is currently the main problem in misalignment and intent failures, and we are potentially only a few breakthroughs away from solving it in most areas. The initial solutions might not even require architectural changes, but training and data ones.

The labs can bring together training and product better than you can in the long run, but there is still potential to innovate before they do. Cursor and OpenClaw both built new interfaces that the labs later adopted. My guess is that if there is a new interface it probably involves moving past threads to something higher-level like a single spec, but things get complex with Git, PRs, CI, and I don't have a good abstraction yet.

Different fields are not so different. Humans organize naturally. For example, YC and SNL have many parallels.

Time is the bottleneck in life and in algorithms, because you can go back in space but not time.

The zeitgeist can feel insane. But things eventually settle to the truth, and often abruptly.

The computational complexity to solve a problem is roughly the same for humans on notebooks, computers on RAM, and LLMs on GPUs, barring small reduction overhead.

Bias towards taking the meeting and doing the thing.

Society is very capable of change. Things have changed a lot for insiders, and they are about to change a lot for outsiders.