2025 - Week #28

A Framework for Understanding AI Companies

AI companies can be categorized into three distinct groups:

This framework is helpful for understanding how different players in the AI ecosystem create value.

Sources:

The Four Pillars of Product Discovery

Product discovery should systematically address four critical risks:

Source: chapter 33 of “Inspired” by Marty Cagan, which provides a clear framework for de-risking product development before significant investment.

Understanding Temperature in Language Models

Language models work by predicting the next token in a sequence of text. They calculate probability distributions across their entire vocabulary to determine what comes next.

For example, after “Hi! How are you?” the model might predict “Hi” with 70% probability, “Hello” with 15%, “Fine” with 5%, “Why” with 0.4%, and so on.

The challenge is choosing which token to present to users—a process called sampling. Greedy sampling simply picks the highest probability token (“Hi” in our example). Probabilistic sampling follows the calculated distribution, presenting “Hi” 70% of the time, “Hello” 15% of the time, and so forth.

The temperature parameter adds flexibility to probabilistic sampling by redistributing these probabilities. Higher temperatures increase the likelihood of rarer tokens, producing more creative and varied responses. Lower temperatures favor the most probable tokens, creating more consistent and predictable outputs.

Source: Chapter 2 of “AI Engineering” by Chip Huyen

What I’m Listening To

What I’m Reading

What I’m Working On

I’m continuing the development of my macOS app that lets users chat with multiple LLM providers. Following Indragie Karunaratne’s inspiring post, I’m building it using also Claude Code.

Using AI agents has improved my productivity in ways that are hard to appreciate without direct experience. The quality of Claude Code often impresses me, though success requires following best practices to avoid getting stuck in unproductive cycles (for example: Claude Code - Best practices).