https://www.jmd.gg Justin DaCosta is a machine learning engineer at Amazon AGI. He designed reinforcement fine-tuning for Amazon Nova, previously built threat models for Route 53, and ships iOS products. https://www.jmd.gg/writing/moe-is-probably-good-enough.html https://www.jmd.gg/writing/moe-is-probably-good-enough.html Wed, 18 Mar 2026 12:00:00 GMT I spent a lot of time trying to make sparsity work in dense models. Mixture-of-experts leaves some intelligence on the table. For almost everything I care about shipping, it is enough. https://www.jmd.gg/writing/memory-is-all-you-need.html https://www.jmd.gg/writing/memory-is-all-you-need.html Thu, 06 Nov 2025 12:00:00 GMT Chat products grew a memory of you. Taken seriously, that pattern is how you get business intelligence that is actually useful. https://www.jmd.gg/writing/what-should-models-do.html https://www.jmd.gg/writing/what-should-models-do.html Tue, 19 Aug 2025 12:00:00 GMT Frontier models keep getting bigger so they can hold more of the world in their weights. I am not convinced that is the job.