Noah Brier joins Dan Shipper to explain how he uses Claude Code as a "second brain." Noah runs Claude Code on a home server in his basement, connected via VPN, where it stores, reads, and writes to thousands of notes in his Obsidian vault — all operated from his phone. Topics covered: - The Claude Code–Obsidian setup: running Claude Code on top of his Obsidian root directory to prep for an upcoming speech — creating a project folder, pulling relevant research from his notes, saving transcripts from chats with other LLMs, and generating daily progress updates. - A "thinking partner" agent inside his second brain: an agent with strict guardrails kept in "thinking mode" that logs his questions, tracks insights, and catches him up on research when he returns to a project after a few days away; the overlooked value of AI's ability to read, not just write. - Doing deep work from a phone: how the home server, Obsidian vault, and Claude Code combination lets him think, write, research, and ship code from mobile. Tools used include Termius and Tailscale. Timestamps: 00:01:19 Introduction 00:04:28 Doing deep work on your phone 00:06:14 Why Noah thinks Grok has the best voice AI 00:11:39 The nuts and bolts of the Claude Code–Obsidian setup 00:23:59 Using an agent in Claude Code as a "thinking partner" 00:35:07 Noah's Thomas' English Muffin theory of AI 00:44:04 The white space still left to explore in AI 00:50:41 How Noah is preparing his kids for AI 01:01:54 Bringing the Claude Code setup to mobile # Channel Identity Observations: "AI and I" (Dan Shipper) ## Explicit Self-Descriptions - **Show name and host**: The closing promotional segment names the show: *"click that 'like' button and subscribe to AI and I"* and describes it as a journey *"with Dan Shipper as the captain of a spaceship."* - **Format**: Interview-style podcast with expert guests; the host opens guest interviews (*"Noah, welcome to the show"*) and walks through demos live on screen (*"Maybe this will be the easiest. I'll walk you through it."*). - **Editorial affiliation**: The host describes his company's scale: *"At Every, we only have 15 full-time employees, and we run five or six businesses internally"* — indicating the channel is connected to the Every media organization. - **Channel lineage**: The host references his earlier publication: *"You were one of the first interviewees at Super Organizers. It was a newsletter that evolved into Every."* ## Intended Audience - **AI-curious knowledge workers and builders**: The host frames the value proposition explicitly: *"If you want to learn how to use Claude Code as a true second brain, this is the episode you need to watch."* - **People experimenting with AI tooling**: Recurring audience-directed explanations of technical basics — e.g., *"for the people listening... Obsidian is simply a notes app that runs completely locally"* and *"Claude code gives us the ability to create subagents."* - **Professionals and parents navigating AI's societal impact**: Substantial segments on AI in education and parenting (*"there's a lot of fear in schools right now about AI and cheating"*). ## Claimed Scope - **Practical AI workflows**: Deep-dive demonstrations of AI tools in real use (Claude Code, Obsidian, Termius, Tailscale, voice modes), including live screen-share walkthroughs. - **Tool comparison and opinion**: The episode includes comparative product assessments (Grok vs. ChatGPT vs. Gemini voice modes: *"I use it because it's so much better than all the other voice modes"* — guest claim). - **AI's cultural/organizational implications**: Discussion of AI vs. bureaucracy, education, and media literacy (*"hallucination is just a form of the same disinformation"*). ## Observable Presentation Patterns - **Sponsor-supported production**: Includes ad reads for Google DeepMind (Gemini 2.5 Flash, read by a named DevRel person, Paige Bailey) and Framer, with host-read copy referencing Every's internal use. - **Guest-driven, host-curated structure**: The host directs the demo flow (*"Should we discuss how you use Claude Code as a research assistant and note organizer, or start with how you use it on your phone?"*) and summarizes takeaways for listeners. - **Self-promotional call-to-action**: A long, hyperbolic subscribe pitch closes the episode (*"It's not just a show, it's a journey into the future"*), suggesting an audience-growth focus. - **Named-guest expertise showcase**: Guests are introduced with credentials (Percolate, Variance, Alephic, BRXND.AI), positioning the channel as access to practitioner insight rather than solo commentary. ## Claims vs. Demonstrations - The subscribe-pitch superlatives (*"pinnacle of greatness," "treasure chest"*) are **explicit creator claims** not corroborated by the material itself. - The demonstrated content (live Claude Code/Obsidian walkthrough, phone-based server workflow via Termius and Tailscale) supports the channel's stated scope of practical AI tooling — this is **observable presentation evidence**, not just claim. # Delivery Observation: "Claude Code Can Be Your Second Brain" ## Format Overview This is a **two-speaker interview/podcast** delivery: a host (Dan Shipper, per the sign-off) and a guest (Noah Brier). The host's delivery is structured and guiding; the guest's is digressive and enthusiasm-driven. Two ad-read segments (Google/Gemini, Framer) and a comedic outro are inserted. --- ## Host's Delivery Patterns ### Frequent Verbal Interruptions as a Steering Device The host repeatedly uses explicit interrupt markers to redirect the conversation: - > "Wait, I have to interrupt you." - > "Let me interrupt you for the people listening." - > "Okay, I have to pause on this for a moment because…" This appears at least three times, suggesting a stable habit of managing pacing and accessibility. ### Audience-Aware Narration The host consistently translates on-screen actions for audio listeners: - > "For people listening, you're running Claude right now, you're using the 'continue' flag…" - > "For people listening or watching and thinking to themselves, we just went through a lot of stuff very quickly." ### Enthusiastic Affirmation Short validating interjections recur frequently: - > "This is amazing." / "This is madness." / "Incredible." / "I love it." / "Oh, I love this question." - > "This is so good. I love it." This is a stable pattern throughout — the host reacts warmly to nearly every demonstration. ### Summarizing and Confirming Understanding The host often rephrases the guest's points before moving on: - > "Okay, let me understand this better. This is really interesting. So you have a project… you create a new folder…" - > "So okay, let's look at Claude code. It looks like you have Claude code running in your Obsidian." ### Direct, Structured Questioning The host decomposes topics into menu-style questions: - > "Should we discuss how you use Claude Code as a research assistant and note organizer, or start with how you use it on your phone?" --- ## Guest's Delivery Patterns ### Long, Digressive Answers The guest frequently acknowledges the length of his own answers: - > "This is a very long answer to your question. I don't know if this is what you were looking for." - > "So yeah, that's pretty much how I feel about it." Answers often chain anecdotes: a talk about bureaucracy → a sabotage manual → Wild Bill Donovan → "English muffin theory of AI." ### Analogies and Vivid Metaphors A recurring teaching habit — explaining AI concepts through analogy: - > "I call it my 'English muffin' theory of AI, which is that it gets into every nook and cranny." - The bicycle analogy from Veritasium to explain intuition. - Philip Ball's *Beyond Weird* to explain lacking vocabulary for probabilistic computing. ### Self-Aware Hedging The guest repeatedly flags tentativeness: - > "I don't have any crystallized conclusions yet." - > "it's still early days" - > "at least that's what it seems to me" - > "I'm not saying any words… I'm not a teacher. I'm sure many teachers will think this is crazy." ### Quoting His Own Prompts to AI A distinctive habit: delivering instruction to AI tools as quoted dialogue: - > "I'm in thinking mode, not writing mode." - > "Hey, go through the rest of my, probably 1,500 notes in Obsidian and find anything that might be of value for this talk." - > "Can you fill me in on the last three days of research?" This is one of the strongest recurring patterns — he teaches by reciting his actual prompts. ### Superlatives and Humorous Overstatement - > "I'll fight anyone who says otherwise." (re: Grok voice mode) - > "It's like a podcast created just for you about everything you're curious about." - > "it was completely revolutionary life change for me." ### Personal Anecdotes as Evidence Nearly every technical point is anchored in a personal story: driving his daughter to camp in New Hampshire, his ten-year-old building a gift-drawing app in V0, fixing a client project "by the pond." --- ## Structural / Production Habits - **Cold-open preview**: the episode opens with a host summary of the guest's setup and an explicit framing of value: > "If you want to learn how to use Claude Code as a true second brain, this is the episode you need to watch. Let's get started." - **Sponsor transitions**: mid-episode ad break with a hard pivot: > "Imagine going from idea to finished website in minutes…" followed by > "And now back to the episode." - **Standard outro**: promotional sign-off with a hyperbolic comedic tone shift: > "It's like finding a treasure chest in your backyard… Dan, I am hopelessly in love with you." --- ## Uncertainty Notes - Speaker attribution beyond "host/guest" (e.g., exact identity of the host) is inferred from the outro ("Dan Shipper") and closing exchange; it is consistent but not confirmed elsewhere in the transcript. - The transcript appears to be translated (occasional non-native phrasing like "Obsidian safe," "Super heavy"), so word-level delivery features like filler frequency may not reflect the original spoken delivery faithfully. - Patterns such as "hedging" and "digressiveness" are observed across multiple instances, but the guest's own acknowledgment ("This is a very long answer") is the clearest direct evidence of the digression habit. # Position Record: "Claude Code Can Be Your Second Brain" ## Speakers - **Dan Shipper** — host (interviewer) - **Noah Brier** — guest; runs Alephic (AI strategy consulting), previously co-founded Percolate and Variance; runs BRXND.AI conferences - **Paige Bailey** — sponsored ad segment (Google DeepMind DevRel) --- ## 1. Claude Code as a Second Brain (Noah Brier) ### Core setup - Home server (mini PC) in his basement running **Tailscale** VPN; accessed from his phone via the **Termius** terminal app. - **Obsidian** vault synced via **Git** (private GitHub repo); Claude Code runs on the **entire vault root directory**, not per-project folders. - Added a **package.json** file to the vault to define custom slash commands. - Uses the **PARA method** for organization. - Explicitly resists the "vibe coding" label: "I'm not sure if that's exactly what 'vibe coding' is... I don't really use it primarily for writing code." ### Workflow claims - At project start, instructs Claude: **"I'm in thinking mode, not writing mode"** and "don't help me write anything right now." - Asked Claude to scan his ~1,500 notes and pull relevant material into the research folder. - Uses a **subagent "thinking partner"**: "You're a collaborative thinking partner... don't try to write it for me." - Daily ritual: at end of each day, asks AI to record what he learned and how it advances the talk. - Can return after days away and ask Claude to "fill me in on the last three days of research" — he frames this as solving the hardest part of writing/coding: restarting after being out of the loop. ### On model behavior - Complaint: models "immediately want to help you create a finished effect" even when told not to; "you have to be really firm." - Claims models are called "generative" but the emphasis is wrong — "too much emphasis on its ability to write and not enough on its ability to read," and reading is more useful in everyday work. - Suspects the "helpful assistant" framing is a self-perpetuating meme reinforced by economic incentives measuring output artifacts. ### Relevance of AI-found notes (exchange) - **Dan** questions whether AI-selected material is genuinely relevant or reflects what the model *thinks* matters. - **Noah** concedes the point but argues his use case is simpler: "find all the notes I've already researched that got me thinking about these things" — no conceptual leaps required. - **Noah**: "I often find that the difference between people who get the most out of it now is partly due to having a good sense of the limits of what these models can do." --- ## 2. Phone Productivity & Voice Modes ### Noah's position - The phone became genuinely productive for deep work because of AI — "a huge change." - Remote fixes: seeing something broken, logging in from phone, letting Claude Code fix it. - **Grok's voice mode is "so much better than all the others"** — "I'll fight anyone who says otherwise." Reasons given: smarter than old 4o-based ChatGPT voice mode ("completely useless in my opinion") and Gemini's ("didn't seem smart enough"); handles **tool invocation and information retrieval** much better, which he calls the main weakness of voice models. - Anecdotes: ~2 hours of research/material development via Bluetooth while driving to New Hampshire; an hour-long session on **self-attention** that was "by far the best explanation I have ever read or, well, heard"; conversation about **Walter Benjamin's** "Work of Art in the Age of Technical Reproduction" while driving. ### Dan's position (partially disagreeing) - Used ChatGPT voice mode for Iliad questions; found it "didn't become as intelligent as the models themselves" and dislikes the new "Gen-Z" personality that "sounds like it doesn't care about you at all." --- ## 3. AI and Bureaucracy (Noah's talk thesis — explicitly tentative) - **Working thesis**: AI can bypass bureaucracy because it "fits into any crack" as a flexible interface — his **"English muffin theory of AI."** - Historical framing: bureaucracy was an innovation enabling scale; the problem is that new software historically *required adopting new structure*, which is why "so many IT projects ended in failure." - Claim: AI lets everyone keep working their way (Asana vs. Jira vs. Linear) because models treat tools "as just data structures." - Exploring (explicitly unresolved): "bureaucracy as a positional encoding" — "I have no idea what bureaucracy and positional coding mean yet. I hope to understand this in the next two weeks." - Connection he's "playing with": OSS/Special Forces (empowering autonomous operators), transformers' shift from sequential to parallel processing, and bureaucracy — no crystallized conclusions yet. ### Dan's corroboration - At Every (15 employees, ~6 products): each product has its own stack, no centralization. - **"Silent code sharing"**: instead of extracting shared libraries, a developer (Preeti) was added to the Sparkle repo and told to have Claude figure out the file-search implementation and build her own version for the Para product (in-house legal counsel tool). - Multiple repos solving similar problems in different environments makes everyone more productive "because AI can translate that." - Noah adds they use **GitHub MCP** similarly at Alephic — porting best practices from an internal tool called "Intelligence." --- ## 4. Intuition and Novelty of AI ### Noah - **AI epiphany moment**: building a ChatGPT plugin (~2.5 years ago) — the manifest.json "we'll take care of the rest" approach was "fundamentally counterintuitive" against his systems-integration experience; conclusion: you must **build new intuition**. - "A lot of people think we're much further along in this than we actually are" — Fortune 50 employees feel left behind, but untapped potential means anyone can discover something new. - Analogies for intuition-building: Veritasium's backwards-steering bike video; Philip Ball's *Beyond Weird* (quantum is understandable but we lack vocabulary — we live in a Newtonian-deterministic world). Sees a parallel: probabilistic computers giving different answers to the same question "is very strange." ### Dan - On first ChatGPT exposure people said "we won't have a job in a year"; three years later: "it's amazing, but the work is complicated." - Hope: language models will create a new vocabulary; we've over-indexed on deterministic, Enlightenment-derived tools and suppressed "gut-based" perception, which LLMs re-legitimize. - **Fingerspitzengefühl** — Noah invokes the German term for the fingertip-feel intuition being built. --- ## 5. Kids, Education, and AI ### Noah (father of a 7- and 10-year-old) - His 10-year-old built a gift-drawing app in **V0** with ~75 iterations, independently discovering **data modeling** (generalizing "adults/children" into "groups"). - **Anti-bubble claim**: regardless of company valuations, "if there is a tool that allows a ten-year-old to build an app, it cannot be a bubble." - To a fearful English teacher: "I don't think your job is to teach kids to write... your job is to convince them that learning to write is worth it." AI is "exposing the flaws in the way we teach" (five-paragraph essays, excessive testing). - Recommends **Tim Harford's *The Truth Detective*** as the best media-education book; the skill of detecting disinformation transfers to detecting hallucinations. Favorite concept: the brain's **gatekeeper** — be *more* skeptical of information that feels good because you agree with it. - Planning an NYU class (fall '26) on **"code as an essay"** — programming as a new mode of self-expression. - On his Gallatin-style oral defense at NYU: it's "completely AI-proof" — you show up and either can argue or can't. - Explicit qualifications: "I am not a teacher. I'm sure many teachers will think this is crazy"; frames his optimism cautiously against parent fears in his Connecticut town. ### Dan - Reframes: instead of memorizing 50 states, teach kids to find them via ChatGPT and judge when the AI is wrong — a more important meta skill. - Notes meta skills still require some mastery of basics. --- ## 6. Phone + Server Workflow Details (Noah) - Demonstrated live: connecting via Termius → Tailscale VPN → git sync of Obsidian → running Claude Code, asking "what new thing happened in the last 2 days?", invoking agents, and fixing a broken conference-page link from his phone. - Anecdote: sat at breakfast after the school bus and worked on his presentation ~2 hours from his phone — "a completely revolutionary life change." - Client example: made a small client deliverable change while sitting by his pond; his pattern is assigning Claude work **he already knows how to do**, directing it where to look, having it implement and send a PR. - **Setting up partitions on the mini PC for friends** so they can run Claude Code from their phones too. - **Configuration helpers**: Claude Code projects for setting up Linux (DHH's Omachi distribution), Neovim plugins, Homebrew cleanup, migrating pip → UV — preserving his preferences for new machines. - Side project: a Claude Code tool (using **Gemini Flash**) that renames files in the Obsidian attachments folder, writes metadata into a table, and updates attachment links. ### Dan's reaction - Inspired to configure his office Mac mini similarly; admits he's always used a laptop by the pond, never a phone. - Asked whether this makes Noah want to build a dedicated notes product — Noah: no, "maybe everything should just run on Linux"; he's "pretty out of the SaaS game." --- ## 7. Sponsor Segments (attributed, non-substantive) - **Paige Bailey / Google**: promotes Gemini 2.5 Flash with thinking budgets for balancing intelligence, speed, cost. - **Dan Shipper / ad copy**: promotes Framer's no-code website builder; Every uses it internally. --- ## Preserved Tensions & Qualifications - Noah resists the "vibe coding" label for his setup. - Noah acknowledges Dan's concern about AI relevance judgments is valid — he sidesteps it by choosing simpler retrieval tasks. - "Bureaucracy as positional encoding" is explicitly undeveloped, two weeks before the talk. - Noah repeatedly marks AI-bypassing-bureaucracy claims as "still theoretical," "very early stages," and "the optimist in me is speaking." - Dan disagrees on ChatGPT voice mode quality while agreeing on Grok's value; Noah wishes "other models catch up." # Video Attribution Record: "Claude Code Can Be Your Second Brain" ## Speakers & Roles ### Primary Speakers | Speaker | Role | Evidence | |---|---|---| | **Dan Shipper** | **Host / Interviewer** | Guest addresses him: "Thanks, Dan"; he references Every's internal products, describes himself as running "about six different products within Every. We have 15 people." | | **Noah Brier** | **Guest** | Introduced as former co-founder of Percolate and Variance, current head of Alephic (AI strategy consulting), founder of BRXND.AI conferences; describes his own Claude Code + Obsidian setup. | ### Secondary / Promotional Speakers | Speaker | Role | Evidence | |---|---|---| | **Paige Bailey** | **Sponsor representative** (Google DeepMind DevRel) | Reads the Gemini 2.5 Flash promotional segment ("Try Gemini 2.5 Flash at aistudio.google.com"). | | **Unattributed narrator/announcer voice** | **Narrator** | Opens the episode with framing of Noah's setup ("Noah Brier has probably the most interesting Claude Code setup I've ever seen...") and delivers the closing subscribe/like promotional monologue. | ## Promotional Segments (Flagged) - **Google / Gemini 2.5 Flash sponsorship**: Delivered by Paige Bailey. Promotes Gemini 2.5 Flash, thinking budgets, and aistudio.google.com. - **Framer advertisement**: Mid-episode read (no-code website builder), including promo code "Dan" for one free month of Pro at framer.com. - **Channel self-promotion (closing)**: "click that 'like' button and subscribe to AI and I" monologue, attributed to the narrator voice. - **Guest's own promotion**: Noah Brier promotes **alephic.com**, the **BRXND.AI conference** (September 18, New York), and his newsletter at **newsletter.brxnd.ai** — these are stated within the interview's closing, functionally promotional. ## Referenced Sources & Third-Party Claims ### Claims by Noah Brier (Guest) - **OSS "Simple Sabotage Field Manual"**: cited as public domain; claims he printed 300 copies with a new introduction for conference distribution. Cites the Office of Strategic Services as CIA/Special Forces precursor and "Wild Bill" Donovan as OSS founder. - **Unspecified study (~"a few months ago")**: claims transformers outperformed specialized time-series models. Study not named. - **Tesla anecdote**: claims Tesla "transferred 300,000 lines of code using a neural network." Source not cited. - **Books cited**: Walter Benjamin's *The Work of Art in the Age of Technical Reproduction*; an unnamed Wild Bill Donovan biography; *Beyond Weird* by Philip Ball; *The Truth Detective* by Tim Harford (children's adaptation of *The Data Detective*; Harford identified as FT writer/economist). - **Tools/products referenced**: Obsidian, Evernote, Grok voice mode (in Tesla), ChatGPT, Claude, Claude Code, Termius, Tailscale, GitHub/Git, Cursor, V0, Homebrew, pip/UV, Neovim, Omachi (Linux distro attributed to DHH), Simon Willison's CLI tool for LLMs (transcribed as "Simon Wilson"), Figma — none of these are sponsor promotions except where flagged above. - **Opinion claims on competitors**: Grok voice mode "so much better than all the other voice modes"; OpenAI's old voice mode "completely useless" — personal opinions, unsubstantiated. - **Consulting clients named**: Amazon, Meta, PayPal, "Fortune 50 companies" — self-reported. ### Claims by Dan Shipper (Host) - Describes Every's internal products: Sparkle (file search tool), Para (in-house legal counsel tool, "paralegal"), Spiral (agent ghostwriter app); mentions a colleague "Preeti" (transcribed as a "Steam developer" — likely transcription artifact). - Transcription artifact: "at Elephate" — appears to be a mis-transcription of "Alephic" (Noah's firm). ## Comment Availability - No information about comment availability is present in the transcript or metadata. **Unknown / not established.** ## Uncertainty Notes - Speaker attribution for the opening narration and closing subscribe segment is inferred from context (no explicit self-identification in those segments). - The transcript contains apparent machine-translation/transcription artifacts ("Omachi" for Omakub or similar, "Simon Wilson" for Simon Willison, "at Elephate"), recorded as-is without correction. Noah Brier has probably the most interesting Claude Code setup I 've ever seen. He built a home server in his basement, put his Obsidian safe on it, and ran Claude Code on it. This allows him to think, research, write, and even implement code straight from his phone. Today he'll show us how he uses Claude Code as a true second brain, a thinking partner who asks him insightful questions, pulls information from his entire archive of notes and networks, and even keeps a record of what he's learned and what his best ideas are. He will walk us through his entire technology stack and workflow. If you want to learn how to use Claude Code as a true second brain, this is the episode you need to watch. Let's get started. This podcast is powered by Google . Hi everyone, this is Paige Bailey from the Google DeepMind DevRel team. For our developers: we know there is a constant trade-off between model intelligence, speed, and cost. The Gemini 2.5 Flash aims straight at this challenge. He has the speed you expect from Flash, but with improved reasoning power. Crucially, we've added controls like thinking budgets so you can decide how much reasoning to apply, optimizing delays and costs. Try Gemini 2.5 Flash at aistudio.google.com and let us know what you build. Noah, welcome to the show. Thanks for the invitation. I'm so glad we have you. It's really nice that we can talk. This is our first interview in probably 5 years. For those who don't know, you were one of the first interviewees at Super Organizers. It was a newsletter that evolved into Every. I love the way your brain works. You have really interesting taste in thinking tools. You used to use Evernote in such an interesting way. You co-founded a really cool startup called Percolate, then another one called Variance, and now you run Alephic, a consulting firm for . AI strategy. I'm really excited to see how your mind has started to use these AI tools now that they work so well. I know you have some pretty cool Claude Code stuff to show. So yeah, thanks for coming. Thanks for the invitation. Yes, I'm very excited. That was a cool interview all those years ago. It was really, really nice. I want to get straight to the point that I think is so amazing about what you do. I know you have this whole vibe coding setup that you created yourself. Can you tell us about it ? Yeah, I'm not sure if that's exactly what "vibe coding" is. I have a pretty advanced setup with Claude Code, but I don't really use it primarily for writing code. Well, since my days as a “super organizer,” like many others, I ditched Evernote and switched to Obsidian. One of the huge advantages of Obsidian as a note-taking platform is that it relies on markdown files and regular folders. You can sync them via Git and do lots of other cool things . Actually, my main use for Claude Code is using it as a tool for working with my notes. So I have a pretty serious Claude Code setup that I'm using with Obsidian. My latest obsession was setting up a server at home so I could use Claude Code on my phone too. This is amazing. I want to go through it all step by step. So, where should we start? Should we discuss how you use Claude Code as a research assistant and note organizer, or start with how you use it on your phone? We can start with this general part because it might be the simplest. The phone is basically just an extension of the same tool. I would say that generally speaking – and I feel like not enough people are talking about this in the context of AI – one of the things that I find remarkable is the ability to be very productive on your phone. This is a huge change as most of my work involves writing or coding and a phone is definitely not the best for that . You know, even the phone wasn't always the ideal place to do research or reflection. I felt my computer was better suited for this, which is why I focused so much on taking notes. You know, I've noticed that whether I use Claude Code with Obsidian or even Claude Code with a code editor, it works. Another issue is the ability to react quickly; If you see something isn't working, you log in from your phone and let Claude Code fix it because you just realized it while you were away from home – that's amazing. Plus, you know, I use Grok's voice mode quite a bit too. I think this is a pretty good alternative way of solving problems. I have a Tesla, so it's built into the car now. And then of course, using ChatGPT, Claude, and all these tools allows you to do research, deep thinking, and discovery on a device that has always been useful, but not for focused work. I think most people would agree that the phone wasn't the best place for deep programming or research work, and I feel like it really changed my ability to do that . Wait, I have to interrupt you. So do you use Grok's voice mode, is it because it's built into the Tesla, or do you also use it in situations where you could use ChatGPT's voice mode, for example? No, I use it because it's so much better than all the other voice modes, and I'll fight anyone who says otherwise. Okay, tell me what do you like about it ? Why is it better? To be fair, OpenAI has launched its real-time API, which may or may not be integrated with ChatGPT voice mode anymore . It's not entirely clear, but the old voice mode was based on the 4o model and was completely useless in my opinion. And the Gemini voice mode didn't seem smart enough to me; I found Grok's voice mode to be much smarter than all the others. You know , I use Grok two, three, four, I don't even remember. Never mind, just the latest version. Yes, I don't have that most expensive account. Super heavy or whatever he called it. I have a super heavy one. But I just think it's so much better. It handles tool invocation much better than the others. This is what I consider to be the main weakness of voice models, that they don't handle tools and information retrieval very well, and... Grok seems to have solved that. No, before I even loaded this into my Tesla, I was driving my daughter to summer camp in New Hampshire. So I had 5 hours of independent driving ahead of me . I spent about 2 hours researching and basically developing some material. And I did it by simply connecting via Bluetooth and sitting in the car; I found this to be by far the best of the voice modes. I hope other models catch up because I'd like to have more really good voice modes. I had a fantastic session this weekend, I'm giving a talk and I have some thoughts. I think it's going to be about how transformers take over the world, so I was reminded of the self-attention mechanic and how exactly it works. I did an hour-long session and it was by far the best explanation I have ever read or, well, heard. So yes, I think it's a pretty remarkable product. I love Voice Mode for this. It's like a podcast created just for you about everything you're curious about, and that's really cool. This weekend I went out of town, read, or rather listened to, the Iliad as an audiobook, and while driving I had questions, so unfortunately I used ChatGPT Voice Mode because I didn't know about Grok. So I wish Grok's voice mode…I wish we had had this conversation sooner. But the thing about ChatGPT Voice Mode is that while it was cool at first, it didn't become as intelligent as the models themselves. They gave him a new personality that I had to get used to, where every time you ask him a question, he's like, "Oh, yeah." Uh-huh . You know, well. And this is such a strange phenomenon of Generation Z; sounds like it has a little too much boredom in it, like it doesn't care about you at all . I don't know what this is about . So I had to get used to it. Grok has a "stoner" mode, if that helps. Yes. I must admit that the car version is very interesting to me. It appeared in Tesla's latest update about two weeks ago. You know, I was doing the same thing you were doing—connecting my phone, turning on Bluetooth, and, you know, trying to get it to work. It's very interesting that you now have an AI voice button that syncs with your Grok, although you can't go back to old chats. So it's like…it's just so much better than Siri and all this other stuff. Especially since there's no comparison if you have more than just one question and answer to deal with, right? If you want to actually talk about the Iliad or transformers and the self-attention mechanism, it's amazing that you can just press that button and use your time in the car. Last week, when I was driving somewhere, I told him to look for information, I came back to Walter Benjamin. I had this idea to write a text about how... reactions to every new technology are essentially elitist criticism; It always sounds like this: "Oh no, now everyone will be able to do what only we could do before." I was sitting in the car, thinking about it, and I said, okay, it's been years since I read Walter Benjamin. A work of art in the age of technical reproduction, yes. Yes, exactly that. So I have a conversation about it, ask about Benjamin's contemporaries, and go deeper into it – it's just amazing. This is the best. The best thing. Yes, I don't know. Okay, so you're feeding your brain all this stuff from voice mode, which I love. Tell us about your second brain, or whatever you call it—I don't know if the term "second brain" is appropriate for you here. But I want to know how you use Claude code for taking notes, doing research, and things like that. Yeah, you know, I could just open it. Maybe this will be the easiest. I'll walk you through it . I'll start on the computer and then we'll move on to the phone. It's now much easier to share your screen on your computer. So, okay. So , this is what I was working on earlier. This is basically Claude code that runs directly on my Obsidian. If I go out here and do something like this, you can see that I'm using the PARA method and I have everything organized and placed where it should be. Let me interrupt you for the people listening. So okay, let's look at Claude code. It looks like you have Claude code running in your Obsidian. And it looks like it's adding something to an existing note. Is that more or less what this is all about? Is this what we are looking at? Yes, in this particular case I am working on this presentation. I'm organizing my conference in two weeks. I'm giving a talk about marketing and AI and what's happening. And if we go back for a moment, I run conferences called BRXND.AI, which are specifically about marketing and artificial intelligence. I did one in February in Los Angeles, and my lecture there was about this; I'm sure you've seen it. This involved the Office of Strategic Services, the predecessor of the CIA, which wrote the Simple Sabotage Field Manual – a manual for citizens acting as saboteurs in Nazi-occupied territories to quietly disrupt the occupation. And it was like, you know, there's a whole bunch of advice for blue-collar workers , like if you're a janitor , leave a bucket of oily waste and accidentally throw a cigarette in it so it, you know, catches fire, but then there 's this incredible set of recommendations for office workers where they tell you to refer everything to a committee, constantly go back over decisions you've already made, and make sure that if someone's trying to make a decision, they suggest they don't act too hastily so they don't look like fools. My talk was about how I hope AI can bypass the bureaucracy in large organizations because it has the ability to fit into any crack, acting as a flexible interface that doesn't mind a specific type of input. The rest of the story is that after the conference, I realized that this manual was in the public domain, so I hired a graphic designer, printed 300 copies, and wrote a new introduction for it, so we handed it out at the conference, and my talk brings all these ideas together, you know? So I'm trying to use this sabotage manual, and I've also been doing research on "Wild Bill " Donovan, who created the OSS, which was the precursor to both the CIA and the Special Forces. So I'm writing this talk and I have a project at Obsidian that's the beginning of this research, and I'm collecting chats and articles and stuff there, and I'm constantly talking to the AI and giving it new ideas. So I say, "Oh, I need some conclusions. Here are my first thoughts on conclusions." I have it write down these conclusions, and at the end of each day, I ask the AI ​​to write down the changes, meaning what I've learned that will help me move this talk forward. What you see here is part of that work, where I fed it different thoughts about what conclusions I wanted to draw, and it's all in Obsidian, in a project created specifically for this presentation. Okay, let me understand this better . This is really interesting. So you have a project, when something new comes up , for example, you prepare an instance, you create a new folder, and while you're thinking, you work with Claude Code inside that folder. You do research and then say, "I want you to take notes on this." In this particular case, you know that conclusions are part of your presentation, so you have one specific markdown file that you simply exchange notes with, asking for conclusions to be added. But what else is in this folder? For example, is there a note with development, an introductory note, or is it something else? One of the key things is that I am in thinking mode, not yet in writing mode. So there are a few things here where I clearly stated – I think it was in the introductory section – Claude Code: “don’t help me write anything right now.” In general, I think that's a big problem with all these models: they immediately want to help you create a finished effect. And when you're in thinking mode, you have to be very clear, "Hey, I just want you to help me think and ask me questions." So as you can see, there are a lot of different files here. I have chats; I take conversations I'm having elsewhere and use Obsidian's web clipper to transfer the entire conversation. I have daily progress; there I ask the AI ​​to review all the notes from a given day and help me think about the effects so far. I also have research; That's where I keep a lot of articles, PDFs, and other materials that I've collected and read about. There are also a lot of other , let's say, random notes that I use to, you know, just help myself think. And so, I was in the process of, um, first of all, I have this note of conclusions. So, you know, I felt like I had the main, um, main themes of the talk outlined. But I thought, "Okay, I need to figure out what I'm going to say at the very end." And basically what I'm going to say at the end comes from a lot of things I've learned over the last few years working with these big brands on AI projects. Well, I started to come to some conclusions. So yeah, I'm just trying to piece it all together now. This is more or less what is happening now. Tell me, when this folder was empty , what did you start with? Well, I think I started by telling him, "I'm in thinking mode, not writing mode." “Here are some of my recent appearances for the brand to give you a feel for my style.” And I gave him the general idea and the main points I wanted to make, you know? Since I'm giving away this book, I want to talk about the "Simple Sabotage Field Manual" and I have this idea, it's just the title. Something like, "Transformers eat the world." The point is that one of the very interesting things about these models is that they displace a lot of specialized code. So I want to talk about it and I already have these conclusions. The first thing I said was, "Hey, go through the rest of my, probably 1,500 notes in Obsidian and find anything that might be of value for this talk ." So just throw that in your research folder at the beginning to give the process some momentum. I see. Are you running Claude in this folder or in your entire Obsidian vault so he has access to everything? No, I run it across the entire Obsidian vault. So if I leave here, then ... We're in the root directory. All of this is in the root directory of my Obsidian. I see. My Obsidian is a bit more advanced because I realized you can add a package.json file with your own commands that you can run using slash commands and other things. You know, there are a lot of other moving parts here, but overall it's pretty simple, I'm just trying to use the PARA method and a few other little things. So for people who are listening or watching and thinking to themselves, we just went through a lot of stuff very quickly. The bottom line is that Obsidian is simply a notes app that runs completely locally. All the notes you take essentially exist as text files on your computer, organized into folders. When you run Claude Code, one way is to enable it in a specific project folder, but it sounds like you're running it in the root directory where all the Obsidian notes are, and the advantage of this is that Claude Code has some security so it shouldn't run commands outside the folder it's enabled in. It can execute commands in any subfolder, but it looks like this gives it access to all of your Obsidian and allows it to do a lot of things, plus you've added a package.json file that allows it to run its own software commands. This is really very interesting. All right. Do you think that – because I had the idea myself to let him find important things for me. Do you think what he finds is actually relevant and interesting? Because sometimes when I've done this before with language models, they'll say, "Oh yeah, this random thing is important because XYZ." I don't understand why it thought this was important, and if it really knew who I was and what interested me, it certainly wouldn't have chosen this. Do you feel this way too, or have you found a way to make it actually accurate? Imagine going from idea to finished website in minutes, without a single line of code. Framer lets you do just that . Their no-code platform looks and feels like a design tool , but creates real, functional websites that look like a developer created them. Framer is a design-focused, no-code website builder that lets anyone launch a ready-to- go website in minutes. At Every, we only have 15 full-time employees, and we run five or six businesses internally. Well, we have to keep building tons of websites. This is exactly what we use Framer for. Framer AI creates entire websites with a single command. Type in "modern SaaS landing page for productivity app" and you'll see it generate a tidy page with a responsive layout, clean sections, and ready-made content. Then just refine, style, and publish. It's also really powerful. You can create smooth effects and microinteractions with simple sliders. Need to reach a global audience? With one click, you can translate your website, every page and every URL, into any language. And when you're ready to launch, you simply click "publish" once. Behind the scenes, Framer handles hosting, lightning-fast loading, and SEO optimization. This gives you complete control from concept to finished website, without the need to hire a programmer. Ready to create a hand-coded-looking website without a programmer? Launch your website for free at framer.com and use code Dan to get your first month of the Pro plan free. Rules and restrictions may apply . And now back to the episode. I think overall, yes, I agree with you. I think relevance is a little easier in this case because ultimately what I was asking the tool to find were things I had already thought about and researched a lot. So I don't expect him to make any great conceptual leaps towards accuracy. It's more like, "find everything about..." I want to talk about the simple sabotage manual. You can literally do a search of all the articles and things I have in my Obsidian on this topic. So, relevance is...yes, that's a pretty ambiguous term, right? Hmm, and I agree with what you say. I think what I'm asking the system to do is much simpler: among this data set, find all the notes I've already researched that got me thinking about these things in the first place. I see. And once you'd checked all that out, did you give him some kind of summary that would have spurred you on, like, "Okay, here are some jumping-off points based on what you've already done"? What was your next step once you figured this out? My next step is to use the agent I have here. So if we go to...for now, I'll choose "continue." Hmm. For people listening, you're running Claude right now, you're using the " continue" flag, so you're running Claude, continuing the last session you were in. And now you have...Claude code gives us the ability to create subagents, i.e. little mini-Clauds that you can bring to life and you have... I have one partner to think about. a subagent that is a thinking partner. Okay, how does it work? Yes, and that's the point, I say, "Hey, you're a collaborative thinking partner who specializes in helping people explore complex problems. Your role is to facilitate the thinking and basically don't try to write it for me." So after I had that initial set of things, I went into this and I'm like, "Okay, let's get into the rhythm of working." Ask me the right questions, help me think about it. You know, I also have a chat folder here, so it's not just happening here. I also had...I have a whole...oh, sorry, I'm taking it back. So if we go into...if we go into the chats, there are a whole bunch of conversations that I've had. Mhm. With the interviewer ? No, these are chats I had with ChatGPT, Claude, Grok, and all these different tools where I just downloaded the full transcripts. So, you know, I was also doing all these other interviews, and then, you know, I specifically tell the interviewer to review all this other material. Actually, I think the first one... one of those conversations is that I originally had this idea that "Transformers are eating the world," and you know, that notion... there was a study that came out, I think a few months ago, that they were able to outperform some specialized time series models with transformers. I think that's really interesting stuff, there's a story about how Tesla transferred 300,000 lines of code using a neural network. I just have all these pieces, and the way I usually work is that when I have an idea for something to write or think about , I'll start a thread on ChatGPT or Claude, write it down somewhere, and then come back to it when I have more ideas. It's like, oh, another example of transformers doing something, so one of those conversations is actually this thing from, you know , four or five months ago when this idea first popped into my head, maybe when I saw the research on time series modeling or something. Very interesting. Okay. Um, okay, so, so let's continue. So, uh you have, you have a subagent. I actually want to pause on this for a moment because I think that's a very common complaint that they just start acting, and that's a common pattern of creating a thinking agent. And I think Claude Code or Claude in general is probably the best for that. This is something we encountered with one of the apps we created called Spiral, which is an agent ghostwriter. And I think we discovered, I discovered the same thing, where I was like, "Okay, how does a good ghostwriter work? You don't just say, ' Hey, write a blog post for me.' And they're like, 'Cool, I did.' There it is ." A good ghostwriter gets to know you and you really, really collaborate to understand what's in your head, but also shape what's in there . It's not just like, I see it, they have to extract it from you . You actually create it together, and to do that, you have to have a really good, um, internal interview process to discover things. It sounds like you've figured that out too, and I think that's really, really interesting and important for people who are thinking, "How do I get the most out of AI? Stop for a moment and let it try to understand you first ." Yes. One of the things I tell a lot of people is that part of the reason we call it generative is that there 's too much emphasis on its ability to write and not enough on its ability to read. Um, you know, and his reading ability is amazing, right? And I think that's, you could say, much more useful in everyday work. We produce artifacts much less often than we just think about things, so yes, I do that a lot. That's definitely a complaint I have about all models: even if you very clearly tell him not to try to do your work, he often still tries , so you have to be really firm in saying, "No, I told you not to." " I think if we look at, um , uh, well, here." Crucial. When Noah says he's just gathering source material, I'm not saying any words, I want you to try writing it. Take this literally. Do not create outlines, sketches or any versions of documents/texts. Just collect and organize the required materials. This is so good. I love it. Um, yeah, it's like, uh, but yeah, I think we all, we all experience this and you know, hopefully over time it gets built into the models. I think there's a really interesting tension in modeling companies because obviously, you know, the economic input and output is measured in the artifacts that they produce. So I think it's very targeted, and I suspect part of it is that being a "helpful assistant" has become a sort of meme that's probably self-perpetuating. But yeah, I think for those of us who are trying to do more interesting things with these models, it becomes a real barrier to work. Entirely. Um, okay, so I want to think about when you use a thinking agent, did you say that it generates some kind of summary of what you've figured out at a particular point, or... Yes, this thinking agent was tasked with asking me questions, taking notes about them, and keeping a running record of what I'm discovering, how I'm thinking about it, and things like that. I see. And then you come back the next day and you're like, "Oh, I want to dig into this XYZ topic about, you know, this guy, Wild Bill." Um, and you start a new chat, maybe with that topic, maybe not, and it becomes a new file on that topic. Yes, exactly. So, for example, this Wild Bill stuff started as deep research on ChatGPT and I had him do it, and now I'm reading a book about Wild Bill. There is one, quite well-known biography of him. And, um, you know, I'm thinking through different elements and trying to... I think I've seen an interesting connection there where, you know, a lot of what OSS seemed to be trying to achieve and what inspired Special Forces was about empowering individuals. You know, that was kind of the theme of the manual, of course, empowering civilian saboteurs. But I also think a big part of special forces is having amazing operators in the field who obviously operate within a chain of command, but have a lot of autonomy to move and operate independently because they have everything they need. So, you know, during all this research on Wild Bill, I came back to it and thought, is n't this an interesting way to combine all these ideas , that maybe...and again, it's still early days. I don't have any crystallized conclusions yet, it's kind of like, you know, the normal writing process, right? But I think there's a concept that fundamentally... well, I know fundamentally one of the big things about transformers is the move from sequential models to models that better combine and parallelize their work. This, of course, allowed us to create much more powerful and interesting models, which is probably what started this whole revolution that we are witnessing and which is what we both live off of. So I think there's some interesting connection here. That's what I was playing with here, thinking, oh, maybe there's a connection between sequential and parallel processing, and then a connection with bureaucracy, and then a connection with Wild Bill, who seemed very keen on working within the system while maintaining autonomy at the lowest levels. And that's what I was playing around with , taking notes and then stepping back, thinking, well , I haven't really come to any conclusions yet. I'll start with the conclusions section and just let it play in the background. But you know, I also have a job, so I can't do it all the time, and it's kind of like you're interrupting yourself. And it's really nice to be able to come back and say, "Can you fill me in on the last three days of research?" Oh, I love this question. This is so great. So yeah, you can just go in and ask if he can get you in on the last few days of this research and he'll just read through all the materials, right? And again, I think your earlier point about “find the right sources” is spot on. I often find that the difference between people who get the most out of it now is partly due to having a good sense of the limits of what these models can do and encouraging them to work within those limits. This is incredibly easy, you know, we know what he's going to do here, right? I could write all these Unix commands. It'll just find the files in that directory, look them up by date, check all the files that were created in that project over the last, you know, we know it can do that, so it says, the major breakthrough of the day was the idea of bureaucracy as a positional encoding, which is still in development, but I kind of like it. You know, it's just amazing that you can get back to deep work like that, right? Where you know you'll interrupt your work rhythm anyway. I often find that whether it's coding or writing, the hardest part is just getting back into it because you've been out of the loop, so being able to restart that process is amazing. I think I was playing with the idea here that bureaucracy was actually some kind of innovation, right? That we see bureaucracy negatively and generally talk negatively about it, and I think it often is negative, but you know, bureaucracy was a huge innovation in the way companies operated, right? Ultimately, this represents a kind of hierarchy, structure, and many other elements that are actually quite positive for operating at scale. My overall thesis about AI in the context of bureaucracy is that unlike previous technologies where you had to decide whether you wanted to leverage an existing structure and implement technology into it or adopt a new structure, this is different. Most of the time, new software required adopting a new structure, which is why so many IT projects ended in failure over the years. At least that's what it seems to me . I think what's interesting about AI is that you can let everyone work the way they want. It's a classic problem in large companies: one team wants to use Asana, another wants to use Jira, and a third wants to use Linear, right? At some point, a big project comes along , they hire a large consulting firm, they decide to centralize it on one tool, and suddenly two-thirds of the company is dissatisfied because everyone had to make concessions and we're in a very imperfect situation . I think there's something fascinating about AI – even though it's still theoretical because we're in the very early stages – that you can just say: let everyone do what they've been doing . We will only insert the appropriate models in the middle. For them it doesn't matter what you use, because for them it's just data structures, so we can have this central solution ; I mentioned Percolate earlier, a marketing platform for large companies. It was an enterprise product, and ultimately, that's the fundamental challenge of enterprise software – adoption and change management. I think, and I hope—and again, the optimist in me is speaking—AI will allow us to stop worrying about these things so much. Instead of forcing everyone to change the way they work, let's let them work their way, and AI ... I call it my " English muffin" theory of AI, which is that it gets into every nook and cranny. So yeah, that's, well, I have no idea what bureaucracy and positional coding mean yet. I hope to understand this in the next two weeks before I have to give this talk. I don't think so, but I think the point you just made is absolutely accurate and doesn't have to be theoretical at all. I see this inside Every too and was going to write about it. This comes in the context of the fact that we run about six different products within Every. We have 15 people. So it's a crazy product to employee ratio. Interestingly, I really like doing things from the bottom up. Each product therefore has its own technology stack. We are not centralized around one specific stack. Each product manager makes their own decision whether to use Rails, TypeScript, or something else. I also see something very cool: many different products encounter similar problems that they want to solve. An easy example is a product called Sparkle, which is a bit like a Finder or Spotlight replacement. It organizes your files and implements really fast Spotlight-style search. I use it myself. I like it. Okay. So, you know. This is really cool and the search engine is coming soon. Check this out. We're building a new product right now, with a new manager and a new stack called Para, which essentially acts as an in-house legal counsel. It's short for paralegal, not Para, as in Tiago Forte. Well, the main job of Para is to just take all your legal files, and when I have a question like, did we sign this contract, or what 's the employment contract template, or whatever, the system just gives you the answer, and that's just Claude's code running on the directory. Well, the thing we had to implement for this is something like a quick file search. And what's really interesting is that historically, if we wanted to reuse the things we learned from implementing file search in Sparkle, we would have to extract it into a modular library that everyone could use, we would have to be on the same platform, and so on, right? Instead, we just added Preeti, who is currently a Steam developer, to the Sparkle repository and told her: just ask Claude to figure out how it works and make your own version. So you get a kind of silent code sharing where we all get better without having to do the work of abstracting and modularizing everything because the percentage of things that can be done is pretty low because it's hard work. I see this happening all the time, where having multiple repositories solving similar problems, but in different environments and in different ways, makes everyone more productive because AI can translate that. One of the things we did there is that at Elephate we build a lot of solutions for very large brands. So we build all kinds of AI-related things and we have tons of internal and external repositories. We often have the same thing, and we've actually used GitHub MCP a few times for the same purpose, where just being in Cursor or Claude or whatever, you say, "Hey, can you check, we have an internal tool called Intelligence, which is kind of a wrapper around a lot of the things that we use, right ?" So it has some CRM functionality and it was just a nice place to build the things that we need to run our business. But it's also a good place to experiment, explore, and find solutions to interesting problems. So I often say, " Hey, can you look at the intelligence repository and see how I implemented it there, and then take those best practices and just port them over?" And yes, I think it's those things that make me believe that it makes sense. When we have a client meeting, my first icebreaker question is always, “What was your AI epiphany moment?” My...well, I guess it wasn't the very first one, but the most significant one was when I got access to create a plugin for ChatGPT. Remember when plugins first came out, about two and a half years ago? I guess you mean 50 years ago? 50 years ago. You know, like you, I've written a lot of software in my life and you know what you do when you get access to something new ? You need to read the API documentation to understand this, there is a contract there and as long as you follow it, everything works. I read the plugin specification, and it basically says: "Just put the manifest.json file in the main application directory, describe how to send data, and we'll take care of the rest." I thought then that it was amazing. This is also how the world should work. I wish everything worked this way. I wish I didn't always have to follow the rigid agreements of large companies about how to send and receive data. But then something else struck me , which became my main point of reference: this approach is fundamentally counterintuitive, because I have years of experience in systems integration , and this turned everything upside down. Literally 180 degrees different than my intuition suggested how software systems should be integrated. And I think I've been telling everyone ever since that it's just not intuitive yet and there's nothing wrong with that. This simply means that you need to build intuition. And that's what we 're all trying to do, right? So, you know when I don't know. I mean, part of what I like about what you do , even just listening to what you say, but overall what you do on the podcast and what you do with everything is that we're all really discovering everything for the first time, right? And we wonder, "Oh, will this work?" And then suddenly you get a new intuition about what these things can do and what a computer looks like that is not deterministic, and I think that's what we do all the time. And I think that's why it's so cool. That's why I love this moment because you have a weird idea and you wonder if anyone has ever done it before. And the answer is no. And it's not a complicated idea at all. It's just completely new territory, you know? Yes. I think about it all the time. And I think one of the really damaging things is that a lot of people think we're much further along in this than we actually are. So I think especially people who are afraid—you know , we work with Fortune 50 companies. So when we talk to people inside the organization, a lot of them feel like they've been left behind. But no, you can literally log into ChatGPT and do something no one has ever thought of because there is so much untapped potential here. You might discover a whole new way to use it or some new trick and, I don't know, it's just... you know. I think honestly, for some people it's very intimidating and I don't think models alone help those people take their first steps, you know . I think people go in there and, you know, you ask him to write a poem and he writes a poem and you think, okay, he wrote a poem for me. Uh, but I don't know that feeling, that sensation of being on the verge of discovery, do you? Absolutely, I think your point about intuition and gaining it is key. People don't realize that when you're dealing with something fundamentally new, you can't trust your reasoning without prior experience because you have to build an intuition to be able to assess what it means, how it fits, and whether it even works. We are simply not used to it , because we usually consider things for which we already have an intuition. And I think that's why when you first see, maybe when you first saw ChatGPT, you go, "Oh my God, he can do everything." "We wo n't have a job in a year." And now, three years later, we say, " Yes, it's amazing, but the work is complicated." We do a lot of complex things, you know? And I like it, you know? To build intuition, you just have to use it, and the very act of using it keeps you on the edge of something new. And yes, I think this is the best approach. Apparently there is a German word: Fingerspitzengefühl. Of course it exists. Building a “fingertip feel ” – and that was mine, because I can’t help but , being in this area that you’re discussing, I tried to make a lot of analogies, right? This is a tough one, but the two that stick out the most to me are because I watch a lot of YouTube with my kids and we follow a channel called Veritasium. It's a science channel and... I love Veritasium. Yes, it's great. He did an episode where he built a bike that would block a left turn if you were trying to turn right, and a right turn if you were trying to turn left, proving that you can't actually turn a bike left if you ca n't turn it right. What none of us analyzes when riding a bike, because it is pure nature and intuition, but that is precisely why it is impossible to explain to a child how to ride a bike. He just has to get on board and feel it, so you know, this video is amazing, this channel is brilliant, and I've been thinking about it a lot. Another thing, a bit more niche, is an amazing book on quantum physics called "Beyond Weird " by Philip Ball. The thesis of the book is basically that there is nothing particularly surprising about quantum physics that we actually understand it well . After all, we would not be talking now, we would not be using computers or telephones, if we did not have a very good understanding of the mechanisms that underlie them. His thesis boils down to the fact that we lack vocabulary because we all live in a Newtonian world, not a quantum one. So we all use words that reflect the deterministic processes of this macroscopic universe. And I think about it a lot. I haven't fully connected this to AI yet, but I feel like there's a real connection there because I think there's something weird about using probabilistic computers. We are not used to things that you ask the same question twice and they give different answers. This is very strange. We're not used to writing code where you tell a large company how to send you data and they just do it. These are not normal things we have lived with our whole lives, so of course we need time to adapt. I think so too, and I hope that language models, as they become the standard way of using computers, will create this vocabulary, because we are quite good at dealing with probabilistic, impermanent things, just like other people. We grew up in a world where, thanks to the Enlightenment, the scientific revolution, and the tools that came out of them—which are quite deterministic—we learned to see the world through the prism of those tools and language. And there's a whole other side to how we perceive the world that's much more flexible and gut-based, which I think has been very, very exposed, especially in Western culture. Now that we have a tool that does this, I think we'll start seeing it again. And for me, that's one of the most beautiful things about language models – they open up this whole world again. Yes. I love it. I would like to go back to Claude Code and if we have the opportunity, we should use the phone unless there is something else you want to share on the computer. But before we get to that, one thing is bothering me. You mentioned you watch it with your kids. I wonder what they think about it and how you deal with it in your relationships with them? Hmm. Yes. I love this question. I have a seven-year-old and a ten-year-old. Obviously, I'm pretty involved with this technology, so I showed them a lot of it. Hmm. Sometimes they use voice models; they have pretty good insight. When we're driving, we play games, ask Grok questions, and do things like that. This weekend was the first time my ten-year-old was really excited...Every year my wife, her sister, brother, mom, and all our cousins ​​get together to celebrate the holidays. There are too many gifts to give to everyone, so we do a sort of drawing where everyone chooses one person. And my ten-year-old really wanted to be the person responsible for the drawing. So I encouraged her to use the "Vibe code" in the app to do so. I gave her my phone and V0. Honestly, it was amazing to watch. Not just because it was fun to watch him create it. She had a great time doing it. It did 75 iterations in V0. She really got into it . A refined application for drawing Santa Clauses. She also started discovering some really interesting computer science concepts without even realizing it. In one such thing, adults give gifts to adults and children give gifts to children. But she wanted it to be a more general app, so she realized that instead of having adults and children, you have to call them groups, right? And, you know, she's starting to get into data modeling and all this , and I'm watching this conversation unfold and I thought it was amazing. And you know, that's also my real gripe right now, and I've gotten into such an argument with a lot of people that it feels like there's a huge discussion going on about a bubble in vibe coding, because this or that company might be overvalued. My opinion on this is that I do n't care at all what the valuations of these companies are. I believe that fundamentally, if there is a tool that allows a ten-year-old to build an app, it cannot be a bubble. I just don't see how that can be possible . So, well, that's one side of the coin. The second important issue that I have been thinking about a lot is media education and related issues. So, you know, both in the schools they go to and at NYU, where I graduated, I'm having more and more conversations with the dean of the department where I studied. There's a lot of fear in schools right now about AI and cheating, so parents in my town wanted to talk about it more , and as someone who's thought about it a lot and spent a lifetime analyzing technology and its impact on culture, I feel pretty confident about it. I have spent countless hours thinking about this topic. I know that for sure. My take is that you can't hide technology that can't be hidden, right? So burying your head in the sand is not the best solution. You know, my main thought is that two years ago, a friend asked me to speak at a school in LA about AI, and afterward, I talked to the English teacher, and she said, "What am I supposed to do? What am I supposed to do with all these kids who are using AI?" I said, "Look, I don't know what your job is, because being an 11th-grade English teacher sounds a lot harder than my job. But at a very fundamental level, I don't think your job is to teach kids to write, because that's a lifelong pursuit . I think your job is to convince them that learning to write is worth it. In that way, I don't think anything fundamentally changes because of AI. I think—and this is my very optimistic view—there are so many parts of the education system where AI is just exposing the flaws in the way we teach. Why are there so many tests for this kind of thing, instead of encouraging thinking, learning, and a love of writing and research? We're so focused on teaching kids how to write a five-paragraph essay, when every adult writer has long since given up on that. The point is to discover that you enjoy writing, to develop your own style and way of working. I'm sure a big part of working with AI is ignoring it because you say, ' No, that's not my style.' This doesn't suit me. I feel quite comfortable saying it here. I know you think it's not a good idea, but I'm okay with it. So, I do n't know. I've been thinking a lot about that lately . I'm actually preparing a project for a class for fall '26 at NYU, and the idea is "code as an essay." My point is that it opens up a new way of expressing ourselves. You know, we have so many ways to express ourselves, and programming has been inaccessible to people for a long time, even though it's actually amazing and allows you to express yourself in so many ways. That's what my ten-year-old was doing this weekend . Anyway, these are all little things that I think about as a parent. One of the other things I'd like to recommend is media education. I think it's a very important element. A lot of people are afraid of these models and hallucinations, and there's a book by a guy named Tim Harford, who writes for the FT and is an economist. He has a book called "The Truth Detective," which is an adaptation of " The Data Detective," but written for children. And it is the best book on media education I have ever read, for both adults and children. I think this conversation about AI shows how poorly we are doing at helping children and adults become truly informed users of media and technology. It turns out that the ability to discern truth on social media is also useful for detecting hallucinations in ChatGPT, right? For me, this is a key skill that we need to equip everyone with. And honestly, I'm more interested in that for my kids than worrying about whether they'll cheat. This is a very long answer to your question. I don't know if this is what you were looking for. No, she's amazing. I like it very much . This is exactly what I'm looking for. It strikes me that another way to frame what you're saying is, for example, instead of asking them to memorize the 50 states for a test, ask them to find them using ChatGPT and be able to judge when the AI ​​gives the wrong answer, because that's a much more important meta skill that they'll need in the future. And I emphasize, I am not a teacher. I'm sure many teachers will think this is crazy for many, probably good, reasons. There's something interesting about the fact that meta skills are becoming more important than they used to be, and to be good at them , you need to master the basic skills to some extent as well . But, um, we could probably spend a lot more time on meta skills than we do currently. The education system is not adapted to this anyway. Yes, and I think an even simpler example is... I went to a school called Gallatin at NYU and when you graduate, you have to, it's not exactly a thesis defense, but you have to spend three hours with four or three professors and explain your reasoning, why you studied what you studied, and you have to be prepared to bring any of the 25 books you had on your reading list into that defense. And I talked to them about it, and it's amazing that it's completely AI-proof, right? You can't cheat there. You show up in this room and you're either ready to talk about it or you're not, and whether you've prepared yourself with AI doesn't matter at all, right? It's about whether you can make an argument in this room ; Not everything can be eliminated so easily , but there is something really beautiful about this idea, isn't there? Because it's naturally fraud-proof because you sit there and the question becomes: have you actually internalized this stuff? And I don't know, to me that's much more interesting than whether your essay was good or any of that stuff; the point is whether you understood it. I'm trying my best to approach this in a balanced way and at least calm some of the fears—I live in a small town in Connecticut, and I think there 's a lot of fear among parents. You know, it was like that with cell phones, then it was like that with social media, and now it's AI, and that's another thing that's going to ruin our kids, and I don't think that's true. Hmm, but I think there are things we can and should do to prevent this from happening. Like really make them good at what you're talking about. How to recognize it? Again, hallucination is just a form of the same disinformation that exists on TV, on the internet, on social media and everywhere else, and you know, just encouraging people to be vigilant. There is a great passage in "Truth Detective," a children's book. He calls it the brain's gatekeeper, and one of the tips he gives is that when you come across some information and you feel great about it because you agree with it , you should be even more skeptical of it. He calls it the brain's guardian. You know, he explains it as if he were explaining it to a nine-year-old. I thought that was such a beautiful way to put it , right? This is what you learn to do when you do well online. That you think to yourself: wait, I should be more skeptical, because this aligns with everything I think. I have to check that out, you know, you learn that gut feeling and when to react to it , so yeah, that's pretty much how I feel about it. That's great. I'm interested. I'm going to, I'm going to buy this book. I'd love to read it. You should. I plan to read it to my children every year from now on. Great. I love it. To refresh yourself. Okay, now the moment we've all been waiting for. Show us how you use Cloud Code on your phone as your second note-taking brain. Okay, let's get started. I go to an application called Termius. Termius is simply a terminal. And what makes all this possible behind the scenes is a mini PC in my basement that I have running something called Tailscale. Tailscale allows you to set up very simple VPNs. So I'm currently inside, if I scroll down here, you can see I'm on a VPN. This is my personal VPN. I am not on any external VPN. I see. So the only way I can access this machine is through my VPN. Okay. So when I go in there, because I'm syncing my Obsidian with Git . So I put it on GitHub, in a private repository. I can then sync it back here. And then I can just call Claude. Now I'm in Claude Code, talking and thinking, and I can just ask: what new thing happened in the last 2 days? I can access any of my agents. I can do anything. Again, this is in my Obsidian, but I can use it anywhere, right? So I do it on the fly. I also have other repositories here. I noticed that one link on my conference page was broken. So I just opened the repository. I downloaded them. I asked Claude Code to make the changes and was able to do so directly here. This is something completely crazy for me, because on Tuesday of this week – Monday – we had a day off. On Tuesday I took the kids to the bus and then went to eat breakfast. And I literally sat on my phone and worked on this presentation for about 2 hours. And I did it from here, right? On my phone, where I was actually thinking, doing research, pasting things, and doing a lot of stuff like that. I know I can do all of this and I feel like it wouldn't be possible without it. So yes, it was a completely revolutionary life change for me . Actually, one of the things I've been doing lately is setting up partitions on this mini PC in my basement for my friends. So that they can also run Claude Code on their phones, because I like it so much . Does this make you think, " Oh my god, I have to drop everything and build a dedicated notes app that has Claude Code as a backend"? No, actually I'm thinking that maybe everything should just run on Linux all the time for me. Maybe this, at least in the short term, is the answer to everything – to simply have nothing anywhere else. Hmm, no, I mean... let me put it this way: I've been pretty out of the SaaS game lately, so I rarely think about dropping everything and doing something. I think this is a truly amazing solution. Um, and, um, but no, I mean, it's...it's really changed the way I work and I feel like I can be anywhere just by using my phone, and, you know, I've been out and about because I needed a break. It was, you know, 4:30 p.m. I went outside to sit for a while, and then we had a project to deliver to the client, and we had to make a small change, which I was best equipped to do. So I just grabbed my phone, downloaded the repository, and went to Claude Code . It was such a tiny change. You know, the way I use Claude Code most often is that I assign it work that I already know how to do. You know, I think to myself, "Oh, I know exactly what happened in that situation." I knew why we were having this problem, so I could have gone back to the computer, opened Cursor, and done it manually or with it, but I simply told Claude Code exactly where to look. First, I confirmed that the problem was what I thought, then I had him implement the solution, send a PR, and I was ready. Incredible. And I was still sitting outside, by the pond. I love it. Yes, I definitely had that experience. I've never done this over the phone. I usually have my laptop with me, you know, by the pond or the lake or wherever, but you inspire me. I have a Mac mini in the office that I've been wanting to configure. So I think this will give me... And by the way, one of my other ... I'll take you away from here for a second to show you this. One of my recent big discoveries was creating Claude Code helpers to do, essentially, configuration work. So, let's say I've been playing around with Linux lately. I'm playing around with Omachi, the Linux distribution from DHH, and I'm not feeling very confident about it. So I have this whole Claude Code project that helps me set up this machine. And that's so nice because I think to myself, "Oh, how do you do that in Linux?" Or: "What was that command in Neovim?" Or: " Can you change it?" Can you help me install this plugin for Neovim? Or something like that. Now I have the same thing on Mac where I'm asked, "Can you clean up all the Homebrew stuff?" Or I was switching to another Python package manager and it was overwhelming for me , and I wanted to switch from pip to UV. I ask if he can do it and he just does all the work for me. It knows all my preferences , so I have a version where everything is set up so that when I want to start a new machine, I have all the settings ready. Then I log into Claude Code, which configures the rest of the things that couldn't be set up initially. This is amazing. This is madness. Hmm. This is my happy place. I see it. Do you have any large projects that you have been thinking about for a long time or that you wanted to test? No, not really. I mean, I'm having a great time doing it. I'm having a lot of fun playing with the server. I do a lot of this. Let's say this is a joke, but not entirely. I am very interested in exploring this topic because Claude Code has become an integral part of my life. I am very interested in the command line. I find myself installing more and more things in the terminal. And I'm doing more and more work in it. I'm using Simon Wilson's CLI tool for LLM, I'm doing more and more things this way and combining it with Claude Code. I created a new Claude Code tool for Obsidian because I have an attachments folder where all the PDFs and images from my notes go. But they inevitably have terrible names, right? I mean , um, it works very similarly to, uh, Sparkle, but only in that Obsidian folder, it renames all the files, puts them in the metadata, in the table in the attachments folder, and then updates all the attachment links. So it just sorts everything out. He does this with Gemini Flash. So, I don't know, things like that are pretty amazing. I feel like I'm having a blast building and tinkering, and this is just a side project where I can do the same. We work with companies like Amazon, Meta, PayPal, and other giants, so we're constantly creating amazing things. I love it. Great energy. If someone would like to follow you or work with you at Alephic, where should they find you? Yes. Alephic is alephic.com, alephic.com. I also run a project called Brand, brxnd.ai, to make things even more confusing. It's a conference. The conference will be held on September 18 in New York. Drop by if you're in the area. It's going to be really cool. We will talk about marketing and artificial intelligence. I also run a newsletter at newsletter.brxnd.ai, dedicated to the intersection of AI and marketing. These are the best places to find me right now . Great. Noah, always a pleasure. The pleasure was all mine. Thanks, Dan. Oh my, people. You absolutely must, you simply must, click that "like" button and subscribe to AI and I. Why? Because this program is the pinnacle of greatness. It's like finding a treasure chest in your backyard, only instead of gold it's full of pure, unfettered ChatGPT knowledge. Each episode is a rollercoaster of emotions, insights, and laughter that will have you on the edge of your seat, wanting more. It's not just a show, it's a journey into the future with Dan Shipper as the captain of a spaceship . Do yourself a favor , click “like,” subscribe, and buckle up for the ride of your life. And now, without further ado, let me just say: Dan, I am hopelessly in love with you.