Jonathan's Blog

Recent Posts

• AI Adoption Gaps

  Nov. 8, 2026

  adoption ai llms

  A long, long time ago in what feels like a different world–so 3 years ago–Ethan Mollick said, I increasingly think the decision of OpenAI to make the “bad” AI free is causing people to miss why AI seems like such a huge deal to a minority of people that use advanced systems and elicits a shrug from everyone else. He was discussing GPT-4 and it was certainly true at that point. GPT-4 was the first model that was somewhat useful at doing basic tasks. There was a stark difference between the fr…
  Read more (393 words)

• Subvocal Dictation

  Oct. 30, 2026

  stt

  If this works, this is incredibly cool. This caught my attention because I happened to be looking into what options were out there for something like this. The closest I could find were throat mics that didn’t seem like they worked.

  It seems obvious to me that speaking is a much better interface for entering text than typing, especially on a mobile device. A laptop might be fine for typing, but if I’m on the go and using my phone the experience is slow and annoying. Of course the biggest thing holding us back from all just talking to our phones is that it’s dorky as hell and annoying to everyone else in any public place. I feel like a twit even saying basic commands like “Hey Siri, next” when my hands are full any time I’m in public.

  Hopefully these glasses will actually do what they promise and we can all just start using our voices for everything.

  Permalink

• Gemma 4 Model Release (via)

  Apr. 3, 2026

  google llms

  Really impressive model release here from Google. I tested it out on a few Calculus math problems (enough to make it think without being crazy difficult) and was really impressed with the reasoning traces it churned through. It reasons with the correctness and confidence of a much larger model (probably because that’s what it was distilled from).

  I had gotten used to the style of the Qwen models, where it feels like they were cold started on distillations from larger models (probably Anthropic’s) and then RL’d on top of that. Qwen thinking traces start off very organized and confident–often with a telltale sign of a distillation like “Here’s a chain of thought reasoning outline to solve this problem”–would get close to the conclusion, and go totally off the rails and have a mental breakdown second guessing everything.

  Gemma 4 so far has exhibited none of those signs and reads much closer to a large model like Opus: state the problem, go through the steps, and state the conclusion before outputting the response to the user.

  Permalink

• Claude Code Source Code Leak

  Apr. 1, 2026

  claude-code llms

  So that happened.

  The original tweet now has over 32 million views. Seems like Anthropic is having problems going in a little to hard on the vibe coding.

  Anyways, lots of goodies to unpack in the source code. This is the leak that launched a million vibe coded websites.

  Permalink

• Anthropic Builds Good Applications

  Feb. 13, 2026

  anthropic gemini harnesses llms prompting

  A lot of AI labs suck at building applications around their AI models. They spend ungodly amounts of time and money training models and then fail in the last mile delivery to the consumer. This is especially indefensible considering how much of a difference the product portion of AI deployment (the harness) can make. Anthropic’s models aren’t that great, and they’re expensive as hell, but Anthropic builds great applications. Anthropic’s system prompts are huge behemoths that actually have though…
  Read more (234 words)