The field notes
The Daily AI Drop - Evening Edition, October 16, 2026
Alex and Jordan break down five afternoon drops: OpenAI's Dot triaging a 3am on-call page and running shared team tasks, LangChain on why continual learning is still unsolved, Fahd Mirza testing NVIDIA's 74B Nemotron locally, and the week's top GitHub AI repos.
Alex and Jordan break down five afternoon drops: OpenAI's Dot triaging a 3am on-call page from a phone and ChatGPT running shared team tasks in the cloud, LangChain's Charlie O'Neill on why continual learning is still unsolved, Fahd Mirza testing NVIDIA's 74-billion-parameter Nemotron locally, and the week's top open-source AI repos.
OpenAI — "Put your dot on call"
An OpenAI engineer triaged a 3am on-call incident entirely from his phone: his Dot investigated the issue and started a Codex thread so he could steer the fix from the ChatGPT app. It's the latest proof point in the Dots arc — from killing zombie meetings to calendar triage — showing persistent agents handling real operational work. If an agent can handle incident response at 3am, the bar for everyday delegation keeps rising.
OpenAI — "Put ChatGPT to work on shared team tasks"
A demo of shared, recurring team tasks in ChatGPT: a legal team's daily intake workflow where the team reviews requests, checks deadlines, and refines shared instructions together while the task runs on a schedule in the cloud. Available on Enterprise plans. The shift that matters here is from personal assistant to shared teammate — multiple people feeding one improving workflow.
LangChain — "Why Nobody Has Solved Continual Learning"
Charlie O'Neill (Co-Head of Model Training at Baseten) explains why the "update" step in the own-your-intelligence loop remains unsolved: the SFT-vs-RL trade-off, catastrophic forgetting when new training erases old skills, and distillation warm starts. Recorded at LangChain Presents: The Learning Loop in San Francisco. For anyone shipping agents in production, this is why models go stale — everyone has solved training, nobody has solved retraining.
Fahd Mirza — "NVIDIA Nemotron-3.5 Bolt-74B-A9B Locally: It's Not Over"
Fahd installs and tests NVIDIA's Nemotron 3.5 Bolt — a 74-billion-parameter vision-language model built for everyday agentic tasks — on local hardware. Paired with this morning's 2.4GB Ternary Bonsai test on the same channel, it's a same-day look at both ends of the local-model spectrum. The local scene keeps narrowing the gap with hosted giants, and Fahd's channel is the running scoreboard.
The Next New Thing — "GitHub's top: Copy any app, create gorgeous videos, 3d print anything & more"
Andrew Warner and Adam round up the week's top GitHub repos, headlined by REA (Reverse Engineer Anything) — the AI that opens up software on your computer and explains it. This is the third channel in a week to feature REA, which is the signal. Also covered: Hyperframes for AI-generated video from HTML, Ghidra (the NSA's decompiler), and Text to CAD. The recurring theme: the open-source shelf now has a free answer for a shocking number of paid tools.
Note: pvergadia's "How Hackers Steal Your Login Even With Password & MFA (Here's the Fix)" published inside today's window but was already covered in this morning's edition, so it's not repeated here.