Not Replaced Yet

This podcast is edited with Claude Code. Nick walks Jesse through how it works, starting with a premise he calls the meter: a scale from 0, where you do everything yourself, to 100, a magic button. You push your work as far toward 100 as today's tools allow, then cover the last mile yourself with taste, opinion and craft. One command in Claude Code starts an episode, and Nick's AI chief of staff makes the first cut and hands back a rough version to watch; Nick keeps the notes, the shorts picks, the hook text and the final approval for himself. The hard part was the shorts: plain-language cuts sometimes landed in the middle of a word, because a soft word ending can sound like quiet to a computer. So Nick had Claude build him his own editing tool, where he selects words, presses Delete and listens again. They close on the question of building your own tools or using what already exists, like Riverside's MCP server, which Nick has not tried. His answer: stakes matter. This one has a screen demo, so the video version on YouTube shows the tool: https://www.youtube.com/@notreplacedyet
Question for you: what is the last 10% of your work that you would never hand to an agent?
Everything we make: https://nryet.ai
Sponsored by NonFiction Agency: https://nonfiction.agency

If you run a lot of Claude Code sessions at once, the hard part is getting back to the right one. Jesse gave each of his work streams its own key on an Elgato Stream Deck XL, and he mostly uses it for client work, not for writing code. He explains how it works to Nick: one press opens a work stream's thread right where he left it, two presses clear it and restart it from a saved prompt, and three presses close it out. They talk through the ladder from prompt engineering to agents, why the deck sat unused for months until Claude could handle its settings, the honest cost of working this way (a lot more tokens), dictation instead of typing, and the payoff Jesse cares about most: less context switching. This one has a screen demo, so the video version on YouTube shows the layout Jesse walks through: https://www.youtube.com/@notreplacedyet
Everything we make: https://nryet.ai
Sponsored by NonFiction Agency: https://nonfiction.agency

Nick built an agent that reads every new business filed in Connecticut in a week, 1,117 of them, and hands back the 130 near Hartford that have no website yet and are worth a phone call. He runs it live for Jesse and explains the six steps behind it: search, filter, check, judge, decide, and remember. They cover how the data source was picked out of 26 states, what Claude Sonnet 5 does as the judge, why the cut list matters as much as the call list, and how their agency uses the same pattern every week to watch competitors. This one is a screen demo, so the video version on YouTube shows everything Nick points at: https://www.youtube.com/@notreplacedyet
Everything we make: https://nryet.ai
Sponsored by NonFiction Agency: https://nonfiction.agency

Jesse has three kids, ages sixteen, fourteen and ten, and three different results to report.
The first is his ninth grader, Lydia. She is taking AP World History, she came from a Montessori background, and she had no study habits built for a class like it. She told her dad she had read the same page of notes ten times and understood none of it. His suggestion was not something he built. It was live voice mode in the ChatGPT app. She spent Friday to Sunday talking to it about the opening unit, took the test, and came back with one of the top grades.
Jesse runs that conversation on the episode, on his phone, including the part where he interrupts the answer to ask what a term means. The interruption is the feature. Live voice mode is a step down in raw intelligence, and both of them think the trade is worth it for studying.
It works on this class because AP courses are documented in public. The model has read the same curriculum the teacher works from. That is also how Jesse built an Ableton Live course for his oldest daughter before her music technology class. Ableton publishes deep documentation, so he had Claude read all of it, screenshots and version numbers included, and write a step by step curriculum for someone who already knows music theory. She used part of it and stopped, and the class moved to second semester anyway.
The third example is the ten-year-old. Elliot was told to spend a week of summer building a game, asked if he could build an app instead, and built a soccer training app in ChatGPT Sites while his dad was out of the room. Drills, a timer, skill levels, and a camera mode he asked for because he assumed it would work, which turns on the camera and tracks a real soccer ball.
The back half is the harder conversation. The school says do not use ChatGPT for homework, each teacher adds a version on top of that, and Nick agrees the rule makes sense, because the same app that taught Lydia will hand another kid the answer. Jesse lands on this: a school has to build one system for everyone, AI can be personal to one kid, and the part nobody can delegate is the parent explaining the difference.
Links and everything else we make: https://nryet.ai

Nick has an AI chief of staff. Her name is Ada. She organizes his work, executes a real share of it, and helps him think through what is possible. Ask what she actually is and the answer is deflationary on purpose: a series of markdown files, which is just text files, and a database. Nick did not build a model. He built the context that Claude, the intelligence layer, reads and acts on.
Jesse takes the interview chair this episode. He gets the 30 second version, the Ada Lovelace naming story, and then a walkthrough of a 3D model of Ada that Ada made of herself, in one shot, in under an hour, while Nick was doing other work for this show. Inside it: the memory pages, the working files pulled fresh per session, the Postgres database underneath, and the inbound lane where forwarded email and call transcripts arrive.
Then the two ideas worth taking home. First, the split. Focused work is the thing only Nick can do. Background work is the three factory loops running while he talks. His goal is to always have at least one thing running that does not need him. Second, the honest misses. The Telegram chat lane he set up and never used. The check in messages he built and abandoned. A 16 bit village view of his agents that he calls more fun than useful, right up until he explains the one case where it earns its keep, which is when you are running five thousand agents and no ledger of text helps at a glance.
Jesse asks whether the gain is speed. Nick says speed is real, but the bigger gain is width.
If you are thinking about building something like this, Jesse's advice closes the show. Markdown brick by markdown brick.
Links and everything else we make: https://nryet.ai

Jesse turned his house into data using a consumer drone, phone photos, and his county property tax card, then had Claude build it into a 3D model he can question and plan against. We get into why a house with cameras still forgets everything, and what changes when it stops.
Jesse spent a few Saturday evenings turning his house into data. He flew a consumer drone over the property for shape, stood at the back of the yard and shot elevation photos on his phone, and pulled his county property tax card for square footage and lot size. Then he handed all of it to Claude and got back a 3D model of his home that runs in a browser. He can spin it, zoom it, watch the sun move across it at the real latitude and longitude, and plan actual construction against it. He calls it property intelligence. Nick calls it GitHub revision history for your home.
The idea underneath is simple. Your house is already a system. Outlets, breakers, valves, square footage, property lines, warranties, and every quote a contractor ever handed you. Today that information lives in your head, in a drawer, or in Sharpie marks on a breaker box left by whoever owned the place before you. Cameras and a smart thermostat do not fix it, because none of it remembers anything.
In this episode: what Jesse actually captured and how. Having Claude design the survey before collecting a single measurement. Reading a line item contractor quote you do not understand and pulling cost per square foot out of it. A house that never forgets what it was quoted. Working out which tree takes out which room if it falls. Why Jesse thinks realtors, contractors, and commercial property owners want this more than homeowners do. Where Home Assistant fits. And the dream outdoor kitchen he cannot build for years but has already designed to the square inch.
If you build a version of this, tell us. We want to see it.
Links and everything else we make: https://nryet.ai

Anthropic just got labeled a supply chain risk by the US government — the first time a US company has ever received that designation. Meanwhile, OpenAI shipped GPT-5.4, venture capital hit a record $189 billion in a single month, and Apple is quietly becoming the AI hardware play no one expected.
In this episode, Nick and Jesse break down the biggest stories of the week: the Anthropic vs. Department of War standoff (and OpenAI's role in it), whether GPT-5.4 actually matters to the average user, the massive VC concentration in just three companies, Apple's unified memory architecture advantage for local AI, the DeepSeek V4 delay, Tennessee's proposed AI companion ban, and Claude finding 22 zero-day Firefox vulnerabilities.
Chapters0:00 Intro0:18 Anthropic blacklisted as a supply chain risk1:59 OpenAI's deal with the Department of War2:47 The #QuitGPT movement and Anthropic's consumer surge3:20 Anthropic's safety promise — are they really ditching it?6:33 GPT-5.4 ships — benchmarks and what it means9:02 Hot take: Was this an OpenAI attack on Anthropic?12:30 "We see no wall" — Sam Altman on scaling12:42 $189B VC record — and it's wildly concentrated15:40 Apple's AI hardware play (UMA + local models)17:00 OpenClaw and why Mac hardware is sold out21:29 DeepSeek V4 delay + Chinese open source models25:31 AI regulation heating up — Tennessee's companion ban27:02 Trump administration vs. state-level AI laws32:22 Claude finds 22 Firefox vulnerabilities (zero days)36:17 Closing thoughts
Watch the video: https://youtu.be/zHvkArQszBc

LLM-generated AGENTS.md / CLAUDE.md files are hurting your coding agent.
A brand-new study tested repo-level context files across real tasks (AGENTBench + SWE-bench Lite) and found something surprising:Auto-generated context files cost more… and solve less.Human-written files help a little, but only when they contain what the code *can’t* tell the agent.
In this video, I’ll break down the results, explain the 3 failure modes (redundancy, attention, anchoring), and give you a simple, practical playbook to make your AGENTS.md genuinely valuable.
Chapters0:00 Auto-generated context files made things worse0:08 The first rigorous AGENTS.md study (what it tested)1:15 AGENTBench + SWE-bench Lite setup (ETH Zurich)1:49 Results: no context vs LLM-generated vs human-written3:03 Why another paper found the “opposite” (efficiency vs correctness)3:32 The 3 failure mechanisms: redundancy, attention, anchoring4:29 The 1-line filter: “If the agent can discover it from code, delete it”5:10 The landmines-only workflow (start empty, add 1 line when it trips)5:26 Beyond static files: ACE + optimization loops + layered routing6:27 The real takeaway: not more context—right context6:55 Like / comment / subscribe
Watch the video: https://youtu.be/miDg-3rSJlQ

Mitchell Hashimoto (Terraform’s creator, HashiCorp co-founder) lays out a simple 6-step framework for adopting AI in a way that actually compounds: end with one rule — always have an agent running.
In this video, I break down the 6 steps, what each one looks like in real workflows, and the meta-lesson most people skip: earn the right to delegate.
Links- Mitchell’s post: https://mitchellh.com/writing/my-ai-adoption-journey- Mitchell on X: https://x.com/mitchellh- Ghostty: https://ghostty.org/- Hashicorp: https://www.hashicorp.com
Chapters0:00 I don’t want AI writing code for me0:07 Why Mitchell Hashimoto matters1:11 The 3 phases of adopting any real tool2:02 Step 1 — Drop the chatbot2:42 Step 2 — Reproduce your own work4:13 Step 3 — End-of-day agents5:04 Step 4 — Outsource the slam dunks6:05 Step 5 — Engineer the harness7:05 Step 6 — Always have an agent running7:40 The meta lesson — Earn the right to delegate7:53 What to do tonight (the prompt)8:47 Wrap-up
Watch the video: https://youtu.be/ipuxmO7dj0Y

In this episode of Not Replaced Yet, Nick and Jesse Bushkar share 10 predictions for AI in 2026, from why Claude Code dies, to agents becoming real economic line items, to world models rising, and more on Perplexity and Meta.
Chapters0:00 Intro1:10 Nick’s Prediction 1 — Claude Code dies4:06 Jesse’s Prediction 1 — Agents and Agentic Impact6:16 Nick’s Prediction 2 — World models10:08 Jesse’s Prediction 3 — Foundation models commoditized14:23 Nick’s Prediction 3 — The future of Perplexity19:27 Jesse’s Prediction 4 — A new class of AI-native jobs emerges24:03 Nick’s Prediction 4 — China doesn’t overtake the US or pop the AI bubble in 202630:29 Jesse’s Prediction 5 — The first AI-native solopreneur billionaire35:10 Jesse’s missing prediction — Will Google monetize AI with ads?40:48 Nick’s Prediction 5 — The future of Meta and AI44:53 We skipped Elon and Grok — what’s xAI’s role in AI this year?46:11 Wrap-up and viewer question
Watch the video: https://youtu.be/FJbRdh-8eW0