This week’s model updates have me feeling a bit numb. GPT-6 Sol/Luna just came out, and then Opus 5.5 followed right after—40% cheaper, 30% faster. Our team no longer picks models based on leaderboards; we look at how much it costs per million tokens, whether the latency stays consistent, and whether the API is prone to acting up.
What actually catches my attention more is that Qualcomm one: the new chip can run a 30B MoE model locally on a phone. If that really holds up, many little utilities won’t even need to tweak their API calls.
And then there’s Snorkel’s valuation tripling to $350 million—what does that say? No matter how strong the models get, the data still depends on being cleaned by people. The least valuable part is the model itself; the most valuable thing has always been clean data.
What actually catches my attention more is that Qualcomm one: the new chip can run a 30B MoE model locally on a phone. If that really holds up, many little utilities won’t even need to tweak their API calls.
And then there’s Snorkel’s valuation tripling to $350 million—what does that say? No matter how strong the models get, the data still depends on being cleaned by people. The least valuable part is the model itself; the most valuable thing has always been clean data.