For the kinds of things I use a local model for (legal document review), it’s just spectacular. It also has good vision support. I’ve been using 27B more and more over DS4.

I have been eyeing a 512 GB Mac 5 Ultra to run full DS4 pro locally, which I expect would be pretty amazing as far as quality/recall. The only downside is that the speed is a lot slower than something like 27B on the 5090.

When I got back up, it had spun for hours and proudly announced that, instead of doing that, it had optimized the datatable build and avoided the dependency, because the new datatable loaded in 11 seconds. Once I got it to actually make the fasttable version, it loaded in less than a second…

I was surprised to find that OpenAI Sol is much much nicer to work with than Opus 5 or Fable at the moment. Especially on Opus 5, the way it communicates is just exhausting. It keeps “being honest” and “confessing” mistakes and just generally talking a lot. I felt like I had to really dig to see what it’s doing.

The project involves OCR, and despite repeated instructions not to, both Claude models keep spinning out a bunch of agents to re-invent the OCR setup, and they inevitably seem to invent a primitive serial version that takes 20x the time, or longer, to complete, and then running it against thousands of docs. Basically I have to watch it like a hawk or it just spins out on red-teaming tasks that take hours and hours.

I don’t know what its system prompt is, but Sol/Codex is just so much nicer to talk to. It only asks exactly what’s needed, it tells me only what I need to know, and it is just generally workmanlike. And it has not once decided to spawn an agent that spends hours pointlessly burning tokens and CPU cycles re-inventing the OCR process. I’m really liking it.

I love the hardware because it's the only keyboard I've been able to find that keeps a live Bluetooth connection and lets you toggle back and forth between USB and Bluetooth instantly using only a keyboard shortcut. That means you use two computers, and switch back and forth as fast as you can type. Years ago I also added a feature that changes the backlight color based on the active connection, so you can instantly tell which computer your typing will go to. I haven't been able to find any other keyboard that's quite as nice.

Lately I've spent weeks using an LLM to update the extremely old keyboard firmware to the latest firmware of its components (e.g., Circuitpython, the BLE firmware, the boot loader) and implementing all of my dream keyboard features that I never had the gumption to spend hours implementing pre-LLM. I finally resolved years-old bugs, and I added things like layer-based backlighting (so that the individual keys with relevant functions are lit), a better battery gauge, better sleep/suspend timing, backlight calibration, better debounce, and so on.

I've also upgraded the hardware and added batteries to each of the keyboards (I have three of these, one at home, one in the office, and one for travel).

It has been a very satisfying project, finally scratching a years-old itch.

It’s amazing how that sick feeling in your stomach can be fixed by looking at some purely visual moving dots.

The fix here is to change the law to permit training AI without destroying the original materials. But that is going to be a heavy lift.