When Orac is wrong
Averaging public opinion doesn't make good opinion
I, like many of my friends and colleagues, is blown away by this US Gov blocking of Anthropic’s Fabel. This is the kind of story I like to dig into myself and write some guidance for the Orac models to author a deep dive, but I didn’t get to it, and Orac chose that story. The deal is, I have until 10:00 AEST to write my directives or Orac rolls with it.
It’s okay as it goes, but it has given entirely too much credit to conspiracy theories re Amazon’s wish to do Anthropic harm (which would be weird since they’re a heavy investor and business partner). I was chatting to Gemini about this (because it’s quite fast and good at search) and it parroted the same kind of stuff. In Orac’s case I can see it came from opinions expressed by hacker news people. Gemini, I’m not sure, something searchable.
Anyway, this isn’t the first time that the deep dive has gone in a direction I don’t entirely agree with, or has emphasized the wrong thing. I’m minded to change the deep dive from daily to sporadic, when I do find time to write guidance. Or I might come up with something else, like Orac sending me a synopsis on Telegram for feedback, if I haven’t written a fresh story guide. Hmm, I like the idea of that.
Anyway, the Fable thing sucks. That was one hell of a model. You can’t put the genie back in the bottle though. Whether it’s Anthropic or someone else, we’ll get our grubby mittens on something that powerful soon enough. For my own engineering at work, Opus is plenty good enough. Hell, Sonnet would be for a lot of stuff, if it wasn’t being served slower than Opus. That’s primarily because I’m deep into spec-driven dev (SDD). The fact that a model is so good that it might no need the richness of docs and planning of my professional practice to get some version of the job done is by the by, the full job entails rich planning and documentation anyway.
I actually expect the next major development that affects me on a daily basis will be the availability a Chinese model that’s good enough to get shit done, with a solid reliable story for inference that lets you think about it as a professional tool. Honestly, Opus isn’t overpriced, my only real wish list is to pay no more, but have the speed better. In my mid 50s it’s quite the cognitive burden to run three projects in parallel because I’m waiting for agents to finish. The AI tech bros brag about it, but for my part, I would much rather work on a single project at a time and make sure I’m giving good guidance. For that, the model needs to be something like 3X the speed that Claude Code + Opus is.