Missing context + AI becoming Fuzzy Code

Joke: What do you call an AI that writes novels?
(Answer below ⬇️)

Thoughts to boost you from this week 📅

AI as Fuzzy Code
This week I spoke with a finance startup that uses LLMs as a “fuzzy router,” picking from a small set of actions at each branch instead of hard-coding rules.

They use Microsoft’s Guidance library to limit generation paths for their LLMs. Pretty interesting stuff. It uses Context-Free Grammar (more powerful tool than regexes)
▶️ Check it out: guidance-ai/guidance · GitHub

My tips for choosing the right model: A cascade approach

Start with a good, relatively inexpensive model. I'm a fan of Gemini 2.5 Flash for its balance of speed and cost. (Also, the Gemini 2.5 Flash Lite preview is now out)

1. get your LLM working. I use OpenRouter to compare outputs. Usually Gemini 2.5 Flash works well.

2. Coding. Write code to connect the model to your data / product.

3. add simple evaluations with an LLM judge.

For this, start by simply using a unit test framework (e.g., vitest, unittests). Create ultra-simple test cases with examples of desired output; call your function; pass the tests.

4. Iterate infinitely (or until you’ve solved the problem)

We need context, not more intelligence

It increasingly feels like AI progress toward higher intelligence is slowing down. The key blocker is context - i.e., the amount of relevant information we can feed models.

Pulling in external web data could help.

But this doesn't work for many use cases (most?), when personal context and individual preferences matter.

Recent AI gains are incremental. Better UX rather than breakthroughs. 

One question. If OpenAI is on the edge of reaching Superintelligence, why are they bothering to build more apps?

Or to buy coding IDEs? Or to buy hardware companies?

—

Sending you very warm wishes from London 🇬🇧 (It's 34'C here ☀️)

Tom

Joke: What do you call an AI that writes poetry?
Answer: A prose-essor 🎉