Shorter, freeform notes and reactions, kept out of the main blog. Subscribe via RSS.
Meta Superintelligence Labs shipped Muse Spark 1.1 this week. The turnaround is real and underrated, the pricing is a gift to developers, going closed is understandable but I hope temporary, and the benchmarks would...
GPT-5.6 (Sol) is out: a strong, thoughtfully positioned release with clear Sol/Terra/Luna tiers, flat pricing, and near-Fable capability at better token efficiency. The open questions I most want to stress-test are day-long autonomy and...
Grok 4.5 is the first genuinely frontier-adjacent Grok: fast (~110 tokens/sec), token-efficient, and competitive at real coding, with compelling pricing once you mind the context tier. Safety and trust are the area I most...
Fable 5 has been out for weeks, and it's the most capable model I've used for real work. The classifier friction is a solvable early-days problem, the cost is a positioning choice, and my...