A personal note

AI has moved fast in the past few years. I remember diving as deep as possible into the world of LLMs back in mid-2018. It was a much shallower dive back then, that’s for sure. With the rapid pace of this technology’s progress, I feel like my own opinions have been only recently begun to take shape – though they remain constantly in flux.

Strong opinions, loosely held.

I’m writing this as much for myself as for anyone else. It’s simply a snapshot of my current thoughts, somewhat unfiltered and certainly disjoint and unconnected. I find recording my opinions useful, especially since they often take shape as I discuss them with others.

Opinions

• Google will win
  • Google doesn’t have to build distribution
  • Google has the money and doesn’t need to spend time building products, it only has to build models
  • It can afford to be slow
• If the person who is telling you about AI stands to gain from it, they will be selling it
  • This is an axiom
• Get as close to the bare metal as possible and develop your own opinions
  • Occaisonally, if you can find the smart people working on that bare metal, they’re a good source too
• AI requires developers
  • To build the next model
  • To operate effectively in any context (for programming, for integrating into new and existing systems, etc.)
• Executives all want to reduce headcount (spend) and retain output
  • AI is a clear and hyped way to do this
• Use different models for code changes (Claude seems best post-trained)
• Use different models for problem discovery and planning: Gemini, Grok
• Evaluate new models frequently
  • Determine your buckets of usage and ensure you’re using the right model for each bucket
• AI is good for spikes
  • When you will, and must delete and rewrite the code.
  • It’s actually a funny antidote to the “the demo becomes the product” because AI poisons the well and forces the rewrite that every engineer wanted.
• AI is good for potential options of small changes (take a block and refactor it a few ways)
  • Suggest alternative ways to do the same thing (TDD works well here)
• AI needs as little dead code and debt as possible the larger you get
  • A good context is everything
  • Cruft in a codebase is like poison to the AI context
• Convention over configuration made engineers more productive and will not make their AI models more productive
• AI for learning or grokking
  • Abides endless questions
  • Can provide a million examples
  • The patience of the AI, and its wide knowlege lets you cement learning by poking holes in your mental model
• AI is great for building out mental models of solutions
• AI is 17th century math + 21st century online data and every technical and compute brilliance we can throw at it
  • The second part does a lot of the heavy lifting on usefulness and thus: coding

These are just snapshots, of course. Ask me again in six months, and some of these opinions will likely have shifted again. That’s the nature of working with something evolving so quickly.