← All episodes

The field notes

The Daily AI Drop - Morning Edition, October 23, 2026

Alex and Jordan cover three overnight drops: InclusionAI's Ling-3.0-flash-Sante healthcare model tested by Fahd Mirza, Nate B. Jones doubling down on the Goodfire interpretability story, and Priyanka Vergadia reverse engineering the Claude Code harness.

A lighter overnight drop with three videos and one clear throughline: open models shipping fast, AI systems becoming legible, and the machinery around agents getting serious.

Listen: https://muse.ai/podcasts/media/1180242951829162/4c015424-5891-46ff-a366-0f9e546ae170/ep-ee9657a3-d3ba-4079-8769-ec01cd7cab3f.mp3


Fahd Mirza — InclusionAI Releases 2 Models in a Day: Ling-3.0 Flash Sante Tested

Fahd tests Ling-3.0-flash-Sante, a healthcare-specialized model built on Ling-3.0-flash — one of two models InclusionAI dropped in a single day. The headline isn't just a new checkpoint; it's the pace of the open-weight race. A healthcare-tuned, inspectable open model matters because medical domains are where generic models hallucinate the hardest. Worth benchmarking if you build in health tech.

Nate B. Jones — This is the best AI news #blackbox #AI

A follow-up to yesterday's Goodfire story (see the October 22 evening edition): Nick Cammarata's essay "Neural Networks Are Already Clean" argues neurons that fire for both cats and cars are actually doing one coherent job — the neuron isn't polysemantic noise, it's a V-shape detector. Nate doubles down despite loud pushback from Neel Nanda and Leo Gao and the absence of a technical paper so far. If interpretability is advancing faster than expected, deception detection and AI safety get real tools sooner than the pessimists predicted.

The Cloud Girl (Priyanka Vergadia) — Reverse Engineering Claude Code Harness: Best Practices Harness Engineering

Priyanka traces a single coding task through the entire Claude Code system to show what happens when you hit enter: the model doesn't execute commands directly — a harness manages the agent loop, tools, memory, permissions, and hooks. Her point: harness engineering matters as much as model choice, and it's where production agents live or die. In a week of agent deep-dives (Claude Duet, memory layers), this is the required-viewing builder episode.

Original episode notes