The field notes
The Daily AI Drop - Morning Edition, October 12, 2026
Alex and Jordan break down five overnight AI videos: MiniMax's M3.1 Flash coding model with a million-token context, why the best model is the wrong question once capability saturates, a contrarian take that evals are overrated, a candid startup conversation, and why human creativity still matters.
Five overnight videos, and they mostly rhyme: the models keep getting cheaper and faster, and the smartest people in the field are arguing about what actually matters now. Models are converging, context windows are enormous, costs are cratering — and the value is migrating to the humans who know what to point the machines at.
MiniMax M3.1-Flash-Preview with Mcode: Hands-On Tests — Fahd Mirza
Fahd tests MiniMax's frontier multimodal coding model, headlined by a one-million-token context window and tunable thinking depth. A million-token context means an entire codebase, docs, issues, and chat history in one session without forgetting the start — and tunable thinking depth is the dial that says don't burn expensive reasoning on a variable rename but do think hard on a concurrency bug. It matters because the coding-model race is no longer just the big three: every new frontier entrant pushes incumbents on price and context size, and more context for less money is the direction of the whole market.
Why "best model" is the wrong question now — LangChain
A clip from Harrison Chase's Max Agency podcast with Russell Kaplan, president of Cognition. Kaplan's argument: once software tasks hit a saturation point where every frontier model does them well, raw capability stops being the differentiator — speed and cost start to matter more. It matters because this is the quiet thesis behind the whole small-model week (Haiku 5.5 going ten times cheaper, laptop-grade quantized models): capability parity is arriving, the next war is on throughput and economics, and if you're building agents your model choice should be boring — pick the cheap fast one and spend your cleverness on the harness.
Evals Are the Most Overrated Thing in AI Right Now — Priyanka Vergadia (pvergadia)
Priyanka's contrarian take: machine learning has always been optimization against metrics, so the industry's sudden reverence for evals is just the scientific method wearing a new hat. It matters as a warning against evals theater — elaborate scorecards built instead of shipping and learning from production. The needle to thread: modern agents are non-deterministic and touch tools, money, and data, so the real question isn't whether evals are overrated but which evals actually predict production behavior.
Unfiltered Conversation with my Co-Founder on Startups, Silicon Valley and AI — Aishwarya Srinivasan
A candid conversation between Aishwarya and her co-founder about startups, Silicon Valley, and AI — no demos, no benchmarks, just two builders on what it's actually like to start something in this market. It matters because when the model news is all saturation and cost, the differentiator moves up the stack to judgment: what to build, who to build it with, how to survive the hype cycle.
Human Creativity Still Matters — Matthew Berman
Berman's take on the human-skills question the week keeps circling: the outputs can be automated, but the spark — the weird idea nobody asked for — still has to come from a human. It matters as the creative counterpoint to the commoditization story: as AI gets cheaper, taste and originality get dearer.