Essays on building AI systems: engineering, architecture, leadership, and machines that understand people.https://toshish.com/en-ushttps://toshish.com/writing/building-our-own-llm/https://toshish.com/writing/building-our-own-llm/In 2023, while everyone else wrapped GPT-4, our startup trained and shipped Nebula, a language model built for human conversation. What that decision actually cost, what it taught us, and how I would make it today.Tue, 14 Jul 2026 00:00:00 GMThttps://toshish.com/writing/are-human-conversations-special/https://toshish.com/writing/are-human-conversations-special/We measured how large language models pay attention to human dialogue versus web text, code, and math. Conversations turned out to be the domain models understand least on its own terms. Here is what we found, in plain English.Thu, 18 Jun 2026 00:00:00 GMThttps://toshish.com/writing/eight-years-of-teaching-machines-to-listen/https://toshish.com/writing/eight-years-of-teaching-machines-to-listen/I co-founded Symbl.ai in 2018 on a contrarian bet: that conversations are the most valuable data organizations have, and the hardest for machines to understand. Three eras, one acquisition, and one unchanged thesis later, here is what I know now.Tue, 26 May 2026 00:00:00 GMT