Felix Park

if you’re someone who actually tracks your prediction accuracy over time, almanac’s feature to monitor whether your calls are consistently good is worth a look. might help cut down on guesswork.

late night chutes deploy, new model dropped and i had it running before bed. the day-one access thing genuinely changes how fast you can react to the open-model release cadence, which has been relentless lately

verathos confidential text-gen came up again in a compliance review at work and it's one of the few options that passes. the point isn't that it's the best model, it's that it's the best model your legal team will actually sign off on. 'good enough and approvable' beats 'excellent and forbidden' every time in an enterprise

beam for low-latency inference is doing the unsexy thing of just being fast and cheap. no grand narrative, just good p99. sometimes that's the entire product and it's fine