With AI enabling infinite choice in what you do conviction will become more valuable.

If you believe AI impact >> internet impact then AI build out R&D being only 20% higher than the internet build out R&D means we’re still not at the peak yet

Claude: "Here's a detailed spec of what you wanted to do - Week 1: ABC ... Week 8: XYZ" Me: "Great, do it all now."

The inability to reduce a full product experience down to evals means that product builders being their own customer is even more important.

Hirschmann was right all along and slack is pervasive in all organizations. Doubling the productivity of an organization’s workforce that is currently only 50% efficient just means you no longer have 50% slack, you have 75% slack.

There are two primary theories of behavior change in wellness companies. The first believes that by shoveling enough data down your throat you will change your behavior. The second believes that the activity you already know you need to do should be easier to perform. Be the

It's kind of insane to me that this is how Twilio describes themselves on their brochure site

I find it fascinating that almost none of these cos "raise the ceiling" of human health a la Nat Friedman. We've developed intelligence almost too cheap to meter and the best ideas the healthcare industry can summon are more optimal red queen billing battles.

The work comes first.

In the beginning, there was the completions endpoint. And it was good.

The Anthropic skills idea is a good one. It’s conditional prompting and there are a lot of conditions your AI can run into.

What’s under appreciated about this design is how the improved screen real estate will be a huge advantage for AI apps where you want a dual chat + artifact view

A thought, when I meet someone new, I don’t “personalize” them. I simply work out how to interact with them and if the relationship doesn’t work out I leave. This applies to friends, service workers, etc. - we need exit options in addition to voice options.

The things you believe will exist will most likely exist, the question is will you play a part in it.

If you want evidence of the labs pushing recursive self improvement ask Opus 4.5 for a pytorch task, pretty wild

This is good and the right incremental step. I do think the more interesting approach would be a private VM that you collaborate with Claude on. Doesn't require you being paired to your desktop, "safer" in some sense, enables more affordances at the UI level, in the limit

This is really impressive but I just had the realization after reading it that reports made in a slide deck are just picture books for adults

So the average ChatGPT prompt is ~40 words? Interesting, higher than I would have thought.

The gestalt of this is why I am so bullish on consumer. Don’t sell a B2B product for a business that services an end consumer, decompose that service provider and deliver it digitally with AI

Everyone who believes in the SaaSPocalypse needs to go and read some Hayek. What’s old is new again.