DeepSeek dropped a V3 update with sharper benchmarks and a price floor that makes the rest of the workhorse-tier market look expensive. The cost per million is interesting; the trajectory is the actual story.
GPT-4.1 quietly shipped this week and it's the most interesting OpenAI release of the year. Not because it's the most capable model, because it's the model the business actually runs on.
The legal frameworks around AI training data are still arguing about the wrong layer. The interesting ownership question isn't whether your text was scraped. It's whether the captured shape of how you reason is yours.
Llama 4 is the first time Meta has shipped a mixture-of-experts open-weights model at frontier scale. The release is more interesting for what it implies about Meta's strategy than for its benchmark wins.
Bedrock is built for the enterprise integration story. Most of the AWS-doc treatment assumes that audience. Worth working through what the small-shop or solo-developer version of using it actually looks like.
The framing has become a cliché. The actual mechanics, onboarding, scope of work, performance review, termination, change in interesting ways once you take the metaphor literally.
The workflow-orchestration tools of the early 2010s tried to solve a problem the LLM agents of the mid-2020s are now taking another swing at. The problem is the same. The substrate is different. Worth being honest about which parts of the previous attempt to bring forward and which to leave.
WWDC 2024 promised the most ambitious consumer AI rollout of the year. Most of what shipped is the easy part. The interesting part, the part that would have actually changed how iPhones feel, is still pending.
The headlines say GPUs are unobtainable. The headlines are about the hyperscalers. For everyone else, the picture in early 2025 is more interesting and more usable than it gets credit for.
MCP is the standard the integration layer of the AI stack converged on faster than anyone expected. Worth knowing what it actually is before you wire one in.
A framework I sketched a couple of years back: capture the priors that shape how someone reasons, ship the result as a portable artifact. Worth revisiting now that the substrate has landed.
OpenAI's o1 and o3-mini are the most public proof yet that scaling test-time compute beats scaling parameters at a category of problems. Worth being clear about which category.
