Thoughts on AI, engineering, and building products
Most non-technical AI marketing workflows fail not because the tools are weak, but because they skip the operating model. Here's the stack I'd actually hand a marketing team in 2026, plus the last-frame stitching trick that turns Seedance 2.0 from a 15-second toy into something usable for real campaigns.
Lambda Durable Functions absorbed the core value proposition of every durable execution platform. But Inngest's flow control primitives and developer experience remain genuinely superior for certain architectures. A systems-level comparison for teams choosing in 2026.
Anthropic just built the most powerful AI model in the world. Then they decided you can't use it. Notes on the Claude Mythos system card, Project Glasswing, and what it means when offense structurally outpaces defense.
Google went full Apache 2.0 with Gemma 4, Alibaba closed its flagship model behind an API, a startup got GPU futures on Bloomberg Terminal, and OpenAI bought a media company. Four developments, grouped into three strategic shifts, on Thursday, April 2, 2026.
The launch-era frontier AI results were below 1% on ARC-AGI-3 relative to its selected human action-efficiency baseline, normalised to 100%. The new benchmark abandons pattern-matching grids for interactive video game environments, exposing a fundamental gap between memorization and genuine intelligence.
Large language models waste enormous computational depth reconstructing facts they already know. DeepSeek's Engram module fixes this by adding a fast-lookup memory system alongside neural computation, yielding surprising gains not just in knowledge retrieval but in reasoning, coding, and math.
CopyAlpha is my personal exploration of automated signal processing, configurable risk checks and execution visibility. So I built CopyAlpha, a full-stack platform that processes signals, enforces risk rules, and executes across CEX and DEX.
SK Hynix's long investment in High Bandwidth Memory shows why AI capacity depends on memory, packaging and manufacturing execution as much as accelerator compute.
An analysis of Meta's VL-JEPA paper and Yann LeCun's vision for non-generative AI models that predict semantic embeddings instead of tokens.
My concern about relying too much on LLMs for critical thinking - and how to use AI as a partner rather than a replacement.
My takeaways from Tesla AI Day - from neural networks predicting in vector space to the Dojo supercomputer and the Tesla Bot.
