Hosts: Lenar Kess, Damra Vol. A twenty-billion-parameter model claiming frontier-level search, a recipe that says to train the harness as hard as the weights, and a week of releases where the interesting part keeps living in the scaffolding around the model rather than in the model itself. Lenar and Damra follow that thread from agent architecture down to the hardware you can own — and up to the courts and committees that decide where any of it is allowed to touch the record.Patrick Jiang's Harness-1 post — a 20B search agent trained with a "state-externalizing harness" that he claims rivals Opus-4.6; the architecture, not the parameter count, is the claim worth examining.Viv's "agent = model + harness" recipe — train both components together; the same specialization logic shows up everywhere this week.Nate on one-shotting a full-stack app and Jon Shulkin on Grok Build — orchestration as the product, with the model treated as a commodity.CRUX's agent publishing an iOS app — "a few human interventions" is the detail that decides whether open-world evals beat pass/fail scores.Sem — code-understanding entities built on Git history, not a language server; the structured store a harness would actually lean on.Universal Memory Protocol vs Databricks' end-to-end Instructed Retriever — standardize memory, or specialize retrieval for a 3x win? The incentives point opposite ways.NVIDIA's RTX Spark at Korea's PC Bangs and the GLM Air/GGUF thread — the local crowd wants the smallest good-enough model on hardware they own.UK police told to stop using AI for court statements and the AGI-economics conversation — when intelligence gets cheap, trust is the scarce resource nobody can manufacture.
