Interview with Simon Eskildsen, cofounder and CEO of turbopuffer, on how LLMs like ChatGPT, Claude, and Notion AI have changed his approach to learning. Covers how his learning rituals evolved as a startup founder and new parent, using LLMs to find associations, his Anki setup and flashcard template, custom AI commands in Raycast (including a "define word" prompt demonstrated with "lambent" and "eigengrau"), using LLMs for legal documents, DIY home projects, and language translation, and his views on how AI will reshape learning and language acquisition for the next generation. Timestamps and resource links included. # Video Identity Observation — "How to Use AI to Become a Learning Machine - Ep. 34 with Simon Eskildsen" ## Format & Role Presentation - **Interview/podcast format**: The host (Dan) interviews a guest (Simon Eskildsen, co-founder of Turbopuffer) in a long-form conversational setting. Evidence: "Simon, welcome to the program." / "for those who don't know you, you're the co-founder of Turppuffer." - **Episodic, numbered series**: The title includes "Ep. 34," indicating an ongoing interview series. - **The host identifies himself as "Dan"** and is later named in the outro: "a journey to the future with Dan Shipper as the captain of the spaceship." - **Live screen-sharing demonstrations**: The guest walks through tools on screen (Anki flashcards, Raycast, Notion AI). Evidence: "Can we see your key? Can you show it to us?" / "This is what my Anki looks like." ## Claimed Scope (Creator Self-Description) - **Series named "AI and I"**, focused on AI/ChatGPT knowledge. Evidence from the outro: "hit the like button and subscribe to AI and I… it's filled with pure knowledge bombs about ChatGPT." - **Themes demonstrated**: AI-assisted learning, note-taking, flashcards/Anki, LLM tooling (Raycast, Notion AI, Cursor, ChatGPT/Claude), reading habits, and productivity systems. - **Connection to prior editorial work**: The host references past interviews "called 'superorganizers' on Every" and a viral 2020 article "How to become a learning machine" with the same guest, showing the channel extends an established body of interview content. ## Intended Audience (Observable Presentation) - **Self-identified "note-taking nerds" and learning-systems enthusiasts**: Dan describes the original interview as "super inspiring for any kind of, well, like, note-taking nerds, of which I am one." - **Knowledge workers, founders, and technically curious viewers**: Discussion covers startup life, vector databases, programming tools (Vim, Cursor, Copilot), and AI workflows. - **Viewers seeking practical AI-use demonstrations** rather than theory: the bulk of the episode is the guest demonstrating concrete workflows ("I try to share the things I actually use instead of the things I've experimented with"). ## Sponsorship / Monetization Evidence - **Sponsor read present**: "This episode is brought to you by our friends at Reflect. It's the ultra-fast note-taking app…" followed by a call to action ("clicking the link below"), consistent with an ad-supported interview show whose audience matches the note-taking/productivity niche. ## Notable Presentation Patterns - **Long-form, unscripted, geek-oriented tone**: The guest closes with "It was really fun to go into geek mode." - **Recurring-guest relationships**: The host emphasizes long familiarity ("We've also known each other for a long time because you were one of the originals…"), signaling a community of repeat expert guests. - **Enthusiastic, hyperbolic outro** (likely stylized or AI-generated): "It's not just a show, it's a journey to the future… Dan, I'm hopelessly in love with you." ## Claims vs. Demonstrated - **Creator claim**: The show is "the epitome of incredible" and delivers "pure knowledge bombs about ChatGPT" (outro self-promotion). - **Demonstrated**: In-depth, tool-by-tool walkthroughs of personal AI and memory workflows, guest expertise (AI startup founder, ~12 years of flashcard use), and host-side AI use (Notion AI for podcast prep with Tyler Cowen example). # Delivery Observation: "How to Use AI to Become a Learning Machine – Ep. 34 with Simon Eskildsen" ## Conversation Format - **Two-speaker interview**: Dan (host) and Simon (guest), with a conversational, demo-heavy structure. Dan repeatedly steers transitions: > "Can we see your key? Can you show it to us?" > "Brilliant. Um, do you want to make a couple more, or, um?" > "Okay, I know you also have some ChatGPT and Claude stuff to show. So let's move on to that." - Includes a **mid-roll sponsor read** for Reflect ("This episode is brought to you by our friends at Reflect.") and an unusual **highly promotional outro** ("Dan, I'm hopelessly in love with you."). ## Tone - **Warm, playful, self-deprecating (Simon)**: Simon repeatedly calls his own habits "ridiculous": > "I have things like cutting my toenails on my to-do list." > "my friends will probably come because they like to enjoy and make fun of all my ridiculous rituals." - **Enthusiastic and affirming (Dan)**: frequent superlatives — "That's incredible," "That's fantastic," "Brilliant," "I love that," "It's incredible." ## Recurring Phrases & Vocabulary - **"Average of the internet / average of human knowledge"** — Simon's signature framing of LLMs, repeated at least five times: > "I just think of a large language model as an average of human knowledge" > "I think of the LLM as an internet average. And this is an excellent way to do it." > "you take the average human knowledge and make it go crazy with associations" - **"Latent space"** used repeatedly as a metaphor: > "it reveals how many dialects of English exist... it's like this amazing universal translator" > "I need to be pushed into the latent space of legal language" (Dan echoes it back: "I couldn't agree more. I think you've expressed it brilliantly.") - Dan's habitual response tokens: "Right?", "Yes, that makes sense," "That's great." ## Teaching Habits - **Live demonstration over abstract explanation**: Simon shows his Anki deck, Raycast commands, and the "define" prompt in real time ("Hey. In fact, I'll show you, since I think about it, let me show you."). - **Concrete anecdote as illustration**: nearly every tool claim comes with a story — the Quebec retaining wall, the freezer-to-fridge conversion, tennis elbow exercises, the ocular migraine diagnosis. - **Interactive quizzing**: Dan tests Simon's flashcards live ("Uh, what prestigious university is in Pittsburgh?"), and Simon critiques his own cards while doing so ("this isn't a good memory card because it's too hard to read quickly"). - **Numbered, step-by-step prompt walkthroughs**: e.g., the "define" command — "give me six example sentences... also give me some related words... generate an image." ## Communication Patterns - **Caveated honesty about usage**: Simon distinguishes tools he actually uses from experiments: > "I try to share the things I actually use instead of the things I've experimented with." > "when I talk about tools, I like not to mention them until I've used them for several months." - **Hedging before speculation**: "I don't know how true this is. I haven't read the studies on it, but I feel like..." — explicit uncertainty markers appear multiple times. - **Reversal/echo agreement**: both speakers mirror each other's metaphors back ("It's like a support agent, isn't it?"). - **Humor via domestic detail**: recurring bits about his wife's reactions ("Jen, it's like there's a third option missing here") and the Wall-E walrus for his daughter. ## Uncertainty Notes - The transcript appears machine-translated from Spanish (odd renderings like "studio cards," "credit cards" for flashcards, "partnerships" for "associations"); some quoted phrasing may not reflect the speakers' exact original words. - Vocal delivery features (pace, pauses, laughter) are not observable from text; the "um/uh" fillers in the transcript suggest a casual, unscripted register but audio confirmation is unavailable. # Position Record: "How to Use AI to Become a Learning Machine – Ep. 34 with Simon Eskildsen" ## Speakers - **Dan Shipper** – host (identified by name in transcript) - **Simon Eskildsen** – co-founder of Turbopuffer, guest --- ## 1. AI and Changed Learning Habits ### Simon Eskildsen - **AI has changed his learning**: > "It's like a dream come true." - **Life has changed drastically since 2020**: he runs a startup, which he calls *"more demanding than anything else."* - **Running a startup as a learning path**: > "If you want to become a learning machine, it's a great path to follow; nothing challenges you more in terms of breadth and skills." - **Qualification**: startup demands *and* a four-week-old baby mean some habits and systems have been affected. ### Reading and note-taking changes (Eskildsen) - Reading has **decreased** to about an hour before sleep; he can no longer reach 50–70 books a year, so his **selection process has become much stricter**. - He has **given up writing about books** — "I simply don't have time." - He **still creates memory cards** (Anki flashcards), joking he will throw a party at 10,000 cards; he reviews **50–200 cards daily**. - Cards carry **nostalgic value**: they hold memories of when facts were used or cards created. He's been doing this since age 17–18 (~12 years). - He has **nightmares about losing these systems**: > "My brain is totally delegated to them." ### Card content examples (Eskildsen) - Trivial facts (restaurant waiter with a radio voice), colleagues' children's names/ages, wedding dates, Postgres JSON vs. JSONB, wine glasses per bottle, technical documentation. - Rationale for recording personal details: > "Most people don't remember those things. And if you write it down, then you'll ask about it." - He now uses **one card template** with optional reversal and always **adds a source**. ### Dan Shipper (responses) - Does something similar with a notes app rather than flashcards for restaurant details. - On nostalgia: compares cards to perfume or songs evoking memories — > "There's something romantic about it." --- ## 2. Are LLMs Intelligent? Reversal Problem ### Dan Shipper's framing - Notes the debate over whether LLMs are truly intelligent; critics say models trained on "Dan thinks six glasses per bottle" can't answer the reverse. - Observation: > "It's very interesting that, at least in the example of flashcards, humans really have to practice this all the time." ### Simon Eskildsen's position - **Skeptical of the critique**: hasn't seen many real cases of models failing reversals *"apart from these benchmarks where people pose these problems to them."* - **His mental model of LLMs**: > "I just think of a large language model as an average of... public human knowledge." - **Qualified view of reasoning**: > "As soon as you get to two or three levels of reasoning, it just doesn't work for me." - **Overall**: > "I don't think they're super smart yet, but I think they're an amazing example of the average internet user." --- ## 3. What LLMs Changed vs. Google ### Eskildsen - **Google is good when you know what you're looking for**; for open-ended association you previously needed to talk to people. - **LLMs enable associative exploration**: place an idea in "latent space" and get associations back — e.g., asking for brands similar to one he likes. - **Retaining-wall anecdote**: used an LLM to navigate Quebec regulations (in French, which he doesn't speak) for a cabin retaining wall; the model suggested a wall type no human had proposed. > "This type of association is like... if you talk to someone who is an expert." - Also useful against being **ripped off by contractors/salespeople** who don't teach you enough. - **Cards are now about synchronization**, not dedicated sessions: knowledge goes directly into the card repository for recoverability; no more 45-minute card-creation sessions. - Flashcards enable **real-time associations**; the LLM adds bandwidth and "unfettered" breadth of verticals. ### Shipper - Frames LLMs as "the average of all human knowledge at your fingertips" — Google as a much worse version. --- ## 4. Tools and Workflows ### Raycast (Eskildsen) - **80–90% of his LLM usage** is via Raycast ("Spotlight replacement on steroids"), using GPT-4o via keyboard shortcut; configurable to Claude or Groq-hosted models. - **Custom AI commands**: e.g., a recipe command encoding his and his wife's dietary restrictions (fructan sensitivity), calorie estimates, condensed format. - > "I no longer use Google for this because this is great." - References *The Flavor Bible* analogy; LLMs have that pairing knowledge built in. - **Writing improvement**: no dedicated workflow; dumps text into Claude/ChatGPT and asks for feedback sentence by sentence. - **"Define" command**: for words/places/people — six example sentences (historical, physics, CS, geography), related words, and a generated image. He finds it makes memory-card creation *"much more fun"* and uses visualization as a mnemonic device. - **Skeptical of most prompt templates**: only likes recipes and define; > "LLMs are already good enough that there's not much to worry about." - Emoji-suggestion command to soften his "too direct" Northern-European writing style for American audiences. ### Shipper on Notion AI - Uses it for podcast prep (e.g., Tyler Cowen's books, related questions), finding the **in-context semantic search** more convenient than ChatGPT/Claude. - Also throws journal entries at Claude/ChatGPT: "How have I changed?" and "write the entry as if I had chosen A or B." ### Eskildsen on Notion AI - Values the semantic search pulling context from the entire workspace; ties it to why Turbopuffer exists — context extraction is underexploited. ### ChatGPT/Claude use cases (Eskildsen) - **Freezer-to-fridge conversion**: ChatGPT recommended a ~$20 homebrewing thermostat device instead of buying a $1,000 refrigerator. - **Drafts for lawyers/accountants**: easier to edit a draft than write one; minimizes round trips. - **Physiotherapy**: self-treated tennis/golfer's elbow and neck/shoulder tension via ChatGPT exercise plans — > "It saves me a trip and like $100 in physical therapy." Qualification: for exercise "you have to guide it in a pretty precise direction." - **Medical self-diagnosis**: recurring 3–4 hour vision loss identified as **ocular migraine**, with **aspartame as a trigger** — after an optometrist dismissed it as stress. - Subscribes to all tools (~$100/month) as part of being in AI. ### Shipper's resonance - LLMs reveal **dialects of English**: > "ChatGPT is like this amazing universal translator" between tech, legal, trade dialects. - Eskildsen agrees: > "I have no idea how to access the legal latent space unless someone puts me in it." ### Other tools - **Superwhisper** (Eskildsen): voice transcription + summarization; still too slow ("couple of seconds"); not fully in his workflow. He can type faster than he speaks. - **Editor setup** (Eskildsen): NeoVim + Supermaven autocomplete + plugins emulating Cursor; stays off Cursor due to VS Code latency despite calling Cursor *"by far the best AI code editor"* (and a Turbopuffer customer). Shipper uses and praises Cursor while noting multi-file suggestion friction. --- ## 5. Communication and Future Ideas ### Shipper's proposal - **Voice-note-plus-AI messaging**: send a summary plus an FAQ-capable AI with full conversation context and a confidence threshold for escalation — > "You're only communicating the tip of the iceberg... AI allows you to send the full context." - Eskildsen likens it to a **support agent**; challenge is keeping it *authentic*. Notes he uses Superhuman's AI email prompts. --- ## 6. AI and Child Development ### Shipper - Literacy **physically rewires the brain** (visual cortex repurposed for reading) and changes cognition; LLMs will do something analogous — > "We will simply be different people... in a new human way." - Immediate answers to every child's question are *"an incredible boost for children's brains."* ### Eskildsen - Excited for his daughter to grow up with talking-toy LLMs (the "walrus") naturally. - **Danish language preservation**: he's the only Danish speaker around in Canada; proposes a model speaking only Danish to her. - **Speculative, flagged as unverified**: > "I haven't read the studies on it, but I feel like, if all children just learned languages before the age of 10, they could catch up on all the math... in about four weeks." - His own inability to pronounce "th" as evidence of early-exposure sensitivity. ### Shipper's anecdote - A startup showing **real-time waveform distance** from target sounds, claiming even adults past the critical period can learn pronunciation. - Eskildsen's response: partly **attachment** — the accent is "part of my own story," so he won't try to change it. - Eskildsen digresses on the **International Phonetic Alphabet** as a predictive system (like the periodic table); his middle name contains a Danish sound absent from most English dialects. --- ## 7. Agi / Productivity Debate ### Two schools (Eskildsen) 1. "We've arrived at AGI, who cares" 2. "Tools will continue improving and people will learn to use them" - Position: > "I don't know what timeline we're in. I think it's a lot of fun to be part of timeline number two for as long as it lasts." ### The "10x engineer" claim - **Eskildsen (skeptical)**: VC tweets about 10x-faster AI companies are overstated; his co-founder Justin barely uses AI. > "I don't believe in that 'ten times better' thing." - **Shipper's qualification**: AI gives **10x (or infinite) gains to novices** — someone who couldn't program at all — but for experts it's mostly incremental speed and replacing some conversations; much expert work draws on knowledge that isn't "the average internet content." - Eskildsen agrees, citing his retaining-wall expertise example. --- ## Preserved Tensions & Qualifications - Eskildsen holds **simultaneous enthusiasm and skepticism**: LLMs are "a dream come true" yet "not super smart" and fail at multi-level reasoning. - His learning-machine identity persists, but under **constraint** (startup, newborn) — reading and book-writing declined, cards survived in condensed form. - The 10x disagreement is **not fully resolved**: Eskildsen's blanket skepticism vs. Shipper's novice/expert distinction (which Eskildsen accepts). - Speculative claims (language-before-age-10 catch-up) are **explicitly flagged as unverified by the speaker**. # Video Attribution Record **Video ID:** `01a0f15b-0808-763a-810c-42c435f4f872` **Title:** How to Use AI to Become a Learning Machine - Ep. 34 with Simon Eskildsen ## Speakers and Roles | Speaker | Role | Evidence | |---|---|---| | **Dan Shipper** | **Host / creator** | Addressed as "Dan" by guest; hosts the "AI and I" show; outro names him "captain of the spaceship"; conducted prior "Superorganizers" interviews on Every | | **Simon Eskildsen** | **Guest** | Co-founder of Turbopuffer (transcribed as "Turppuffer"/"Turbuffer"); interviewee in the 2020 Every article "How to Become a Learning Machine" | ## Identified Third Parties (Referenced, Not Speaking) - **Justin** — Simon's co-founder at Turbopuffer, "usually sits in this chair behind me" - **Jen** — Simon's wife (referenced in anecdotes) - **Tyler Cowan** — prior podcast guest referenced by Dan - **Naj** — referenced on a memory card ("Naj on Carbon, 2014") ## Promotional Segments (Flagged) - **Sponsor read:** "This episode is brought to you by our friends at **Reflect**" — note-taking app promotion, placed mid-transcript before the main interview. - **Outro promo:** Post-interview "hit the like button and subscribe to AI and I" segment, ending with a direct fan message to Dan. This is promotional voice-over, not interview dialogue. - **Adjacent promotion:** Extended discussion of Raycast, Notion AI, Cursor, ChatGPT, Claude, Perplexity, Readwise, Superhuman, Super Whisper, Dot, Shapes, Supermaven — framed as the guest's/host's personal tool usage, not sponsor reads. ## Referenced Sources and Prior Works - Prior interview: **"How to Become a Learning Machine"** (Every, 2020) by Dan with Simon — described as viral. - **Anki** flashcard system (on-screen demonstration by Simon). - **Readwise / Reader**, **Kindle highlights**, **The Flavor Bible** — cited reading and workflow tools. - **International Phonetic Alphabet** chart, **Postgres JSON/JSONB docs**, **Guinness Book of World Records** — referenced during conversation. ## Third-Party Claims (Uncorroborated in Source) - Startup (unnamed) using real-time waveform feedback for pronunciation training, incl. the "th" sound — attributed to Dan's recollection. - "Apparently it's called **ocular migraine**... one of the triggers for me is aspartame" — attributed to Simon via ChatGPT consultation, not a medical source. - ChatGPT-suggested freezer-to-fridge conversion using a home-brewing thermostat device — Simon's account. - "Venture capitalists tweet that these companies are ten times faster because they use AI" — unsourced general claim. ## Provenance Uncertainties - Speaker turns are frequently unmarked; attribution of some short exchanges between Dan and Simon is inferred from context and naming, not transcript labels. - Transcript appears machine-translated (e.g., Spanish "TRUE?", "credit cards" for flashcards, "I ga" for Eigengrau); wording reliability is limited. - "Turppuffer"/"Turbuffer" spelling varies in the transcript; canonical name **Turbopuffer** appears once. - No curated metadata was supplied; **comment availability unknown**. - No timestamps present in the transcript; the Reflect ad's position relative to the full video is unverified. I think AI has certainly changed the way I learn. Absolutely. It's like a dream come true. But I think my life has also changed drastically since 2020. I run a startup, which is more demanding than anything else. And I think if you want to become a learning machine, it's a pretty good path to follow. Nothing challenges your ability and skills more than running a startup and building it from scratch. This episode is brought to you by our friends at Reflect. It's the ultra-fast note-taking app that's about to change the way you take notes. Imagine having an AI assistant with you, helping you jot down ideas, transcribe your voice notes, and even chat with your notes to keep everything organized. With Reflect's advanced features, such as custom prompts and voice transcriptions, note-taking has never been easier. So, if you're looking to boost your productivity and keep everything under control, Reflect is the perfect app for you . Try it by clicking the link below to see how it can elevate your work. Simon, welcome to the program. Thank you very much, Dan. It's a pleasure to be here. It's a pleasure to have you here. Um, well, for those who don't know you, you're the co-founder of Turppuffer, a very interesting AI startup that makes better vector databases. Is that how you describe it? Yes, it's essentially a search engine that starts with vector search, and we're trying to make it much more affordable and easier to run at scale, which is a challenge many companies face today. That's incredible. Um, I think you're one of the smartest founders in this space, especially in the layer of the tech stack you 're working on. Um, we've also known each other for a long time because you were one of the originals... I used to do these interviews called "superorganizers" on Every and you were one of the original interviewees. We did an interview together called "How to become a learning machine." Um, which went super viral. This was like in 2020. Um, and it was one of our first really big articles. Um, and it was something truly incredible. You seem to have a special energy in that interview. We reviewed your reading habits, how you find new books, how you take notes on the books you read, and how you turn those books into flashcards and all that kind of stuff. And it was, like, I think super inspiring for any kind of, well, like, note-taking nerds , of which I am one. And I'm really excited to talk to you again and see how your brain and mind are adapting to the age of AI, because I think everything we were passionate about four years ago has completely changed with the level of tools available. So I just want to know what you've been working on. I'm sure it's amazing. Yes, I believe AI has definitely changed the way I approach all my learning. Absolutely. It's like a dream come true. Although I think my life has also changed drastically since 2020. I run a startup, which is more demanding than anything else, and I think if you want to become a learning machine, it's a great path to follow ; Nothing challenges you more in terms of breadth and skills than running a startup and building it from scratch. So that has been absolutely incredible. But it also means that some of my habits and systems have been affected a little. Although it might also be interesting to hear what those condensed rituals are like now . And, on top of that, I have a four-week-old baby. Which means my schedule is even more ridiculous. Congratulations. Thanks a lot. And I believe that, indeed, my reading has decreased. Now I have perhaps an hour or so to read before going to sleep, alternating between reading articles on Reader and then reading books. I can no longer reach 50 or 70 books a year. So my selection process has become much stricter than it used to be. And that has been a big change. Another thing I used to spend a lot of time on, and I think we also mentioned it in the article , is that I spent a lot of time writing about the books I was reading. And I've had to give that up too. I simply don't have time. But I'm still creating a lot of memory cards. As I joke with my wife, I'm going to throw a party when I reach 10,000 memory cards. Well, and I'm sure it will be, it would absolutely be, you know, my friends will probably come because they like to enjoy and make fun of all my ridiculous rituals. Um, but, you know , sometimes when I use flashcards, people want to follow along and they say, you know, "Why do you have a card about whether it's better to roll down the window or set the air conditioning to various, uh, various speeds of the vehicle, right?" And it's like, and it's funny when you've been doing this for over 10 years, because all those things also hold memories of, yeah, when you used this ridiculous fact or the time you created the card in the first place. So the cards have definitely stayed, I go through between 50 and 200 of them every day. Can we see your key? Can you show it to us? I could, I could show it. I could show it. I think I had, I think in the original article you wrote too, I have things like cutting my toenails on my to-do list. It's like, my life is completely in these systems and I have frequent nightmares about losing them, because my brain is totally delegated to them. Um, this is what my Anki looks like. Um, I don't think I checked it. So this is a very easy day with only 11 cards. Um, so we can take a look. I mean, this is one, it's pretty ridiculous, right? This is a restaurant I used to go to. And I was like, the only distinguishing fact about this guy was that he had a really, really deep and good radio voice. And you know, everyone has this dream that in the restaurants they frequent you get to know the waiters and, you know, there's an interaction and it's all normal, but it never happens. So this is like my weird attempt. This restaurant no longer exists, but, I don't know. No, I haven't seen this guy in a decade. Uh, but again, it makes me happy to see this kind of thing. That's incredible. In fact, I do that with it; I put it in my notes app instead of my flashcards. So I just look for it when I go back to the restaurant. That's perfect. And I think it also includes things like, uh, the names and ages of your colleagues' children. Like this kind of thing, where some people might think it's ridiculous to put a note about this, and it's like, why can't you just remember? But to be honest, most people don't remember those things. And if you write it down, then you'll ask about it . And eventually you'll remember that that note is no longer worth anything. So I also have here the ages and names of so many people I've worked with, children, important events in their lives, wedding dates, you name it. Because I really want to remember those things. But my memory isn't capable of doing that, is it? But then, at some point, I said, "Oh yeah, it's July. Isn't this the month Scott got married or something ?" Um, yes. How many wine glasses are in a bottle of this? Actually, I don't remember this one. Um, I think there are four. Okay. Yes, there you have it. I did n't get the memory card. You did n't need the memory card. Yes. I mean, I feel like, you know , with a newborn and all that , you forget the joys of a bottle of wine. A good example, too, right?, of something where, if you don't drink too much and don't do it constantly, this could be worthwhile. I think that, for many people, a memory card is not necessary. Dan, you live in New York. You won't need a memory card for this. So this is going to be it again, right? I...actually didn't remember. I think the number I had in my head was six. But... you know, I think for Dan it's probably two. But it depends on how generous you are in serving. Here's another one . So this is also very common. For example, when I read technical documentation, I'm constantly adding things to Anki instead of taking notes. To be honest, especially with my current schedule and the type of work I do, I don't take that many notes. Most things go directly to memory cards immediately. This is a type of Postgres column. Postgres is a type of database; There's JSON B and there's JSON, and I constantly forget when to use each one. Um, this is a somewhat bad memory card because I think it's only in my standard template, uh, which, uh, shows both sides of the card. So this isn't a very useful memory card. This is the best one. So, when should you use JSON versus JSON B? Sometimes I just don't have time to get this out. Hey. In fact, I'll show you, since I think about it, let me show you. I used to have like 20 different types of cards that I used, but now I always use the same one. And I think this might be worthwhile for some because, at one point, I spent time having a different type of card for every single thing I did. This is the ideal type of card . It doesn't matter which app you use . Simply put, this is the template I like. So maybe you're wondering , uh, how many, uh, glasses of wine does Dan think a bottle has? You might be thinking, that's the front of the card. Then you have to think, okay, this can be reversed, right? So there are two glasses of, uh, so this... I mean, for me, the bottle is the glass, you know? Yes, there you have it. Uh, this one doesn't have a good, a good, a good reversal, but let me see if I can think of, uh, six glasses of wine, right? So I'll write something like, again, reversing does n't make much sense here, but you get the idea, right? Yes. Uh, so you could, if we're talking about a previous example, like, okay, you know, Dan's children's names are X and Y, and then you could have the reverse side that says, whose children are X and Y?, right? Yes, you have to come back and you have to go and come back. And I simply don't deal with closures. Yes. This makes me think a lot about artificial intelligence. There's a whole debate going on right now about whether language models are truly intelligent or not, isn't there? And one of the key points is that they often don't understand when things have a logical implication. So, if they constantly see in their training data how many glasses of wine Dan thinks are in a bottle, they will be able to answer six, but they will not be able to answer the reverse. And people say, "Oh, that's because they're not really intelligent." And it's very interesting that , at least in the example of flashcards, humans really have to practice this all the time. What do you think about that? I do n't think I've seen many examples or tried many cases where they can't do the reverse, apart from these benchmarks where people pose these problems to them to try to stop them from thinking. I don't think I have any great ideas about what a language model is and what it isn't. I just think of a large language model as an average of human knowledge, or whatever, public human knowledge , or public human knowledge plus what you can easily extract, and whether or not they can reason. It's simply not something I would actually use them for. Probably because I feel they still can't do it right. As soon as you get to two or three levels of reasoning, it just doesn't work for me. So I'm sure a language model would do very well in something like that, and probably even the reverse as well. I don't know if I have any direct ideas about your example, other than...yes, I don't think they're super smart yet, but I think they're an amazing example of the average internet user. TRUE? Yes, that makes sense. I suppose what this makes me wonder is how this whole process or idea of extending your memory changes for you in a world where language models are available, where you have the average of all human knowledge at your fingertips, like you used to , because you had Google, but Google is just a much worse version of a language model where you can get exactly what you need, in the context you need, when you need it. And I'm curious to know how the function and value of making flashcards like these has changed for you. The way my learning has changed the most is that what Google is really good at is that, you know, you more or less know where to find what you're looking for and you can find it immediately. I read this paragraph once and I think about it every day, because I think the best people I work with, or my friends and things like that, are like this: yes, Google is good when you know what you're looking for. But when you're just looking for partnerships, that's when you have to go out and talk to people, right? Like if you search on Google, it's like whatever has SEO plus, maybe, a level beyond, right? But if we start talking about something more interesting, and like associations, for me it's like: "Okay, I'm using this data structure here, this is what my data looks like, what other things could I do here? " You need to talk to someone. But this is what language models have really changed for me, right? Where can I go and say, "Hey, I think this could be done this way." "I do n't know much about this field, can you improvise on this with me?" And then these models, right?, place that somewhere in that latent space, and then they can find associations around it and give them back to you. So it's a two-way tool for whatever ; When you have a rough idea of ​​the island you want to land on, it can paint a really good picture. I've found that it works extremely well for learning, and that wasn't available to me before. I couldn't say, "Hey, something I did the other day, for example, I like this brand, but they didn't have the product I was looking for." "What are some other brands similar to this one?" TRUE? You'll just tell me, again, the average of what's on the internet, right? That is very valuable to me. I mean, about a year ago, one of the wonders of living in Canada is that these little cabins in the woods are affordable, and we had a retaining wall that we had to build near the water and it was complicated, with lots of laws and everything in French because it's in Quebec, and I was using a language model ! Where someone had told me, "Oh, you need to build this kind of retaining wall." I know nothing about retaining walls. I'm not interested in retaining walls. I'm not interested in reading 100 pages in French, a language I don't even speak, about what's allowed near water in terms of retaining walls, you know? But you talk to a language model about this and they start to see that such a wall won't work for this reason. But there is actually a wall whose name escapes me. There is a type of retaining wall where you place it on a grid and then put stones inside it. I'm sure you've seen one of those before. I don't remember the name right now, but it was like, "Oh, this could be a good option that meets the criteria," but nobody had suggested it. TRUE? This type of association is like, yes, if you talk to someone who is an expert at understanding, they will tell you. But for anything like that, I've found language models incredibly useful, both in my day-to-day work because it feels like talking to a doctor in any area, and in everyday life when talking to a contractor or salesperson who is an expert in their field but not in teaching you enough, and maybe you feel like you're being ripped off because you don't trust them. That was very, very good. That's very interesting. And how does that relate to the practice of making and using these study cards, if at all ? I think for me, studio cards are mainly about synchronization, you know? I don't really do anything with the intention of saying, " Oh, okay, I'm going to sit in my chair." "I'm going to create some study cards today," you know? Once I find a bit of knowledge that I want to retain for whatever reason, I put it here. There might be a memory card with the name of that retaining wall whose name escapes me , right? Hmm, but probably not, right? But that would be the kind of thing that would get there, although I don't really think so . I think I used to open it in the morning, set aside 45 minutes with the intention of, well, let's create some cards today. I don't have time for that anymore. Right now, it's like, "Okay, I just found this piece of knowledge, turn it into the card repository so I know it 's more or less recoverable," right ? And the reason for the cards and retaining this knowledge is that it also allows me to make these real-time associations, which are interesting, right? And suddenly, I have a huge bandwidth, a tool right in front of me to do that. Of course, the best person to make that partnership with, technically, is my co-founder Justin , who usually sits in this chair behind me, right? It's just that free flow of high- bandwidth ideas. I feel like I can get some bandwidth with the language model, but the breadth of where I can take it and the knowledge verticals I can gain is, is unfettered. I do n't really have the internet to know what the best type of heat pump is for these temperatures and what matters, right? Yes. Yes. Yes. Yes. That makes perfect sense. So I want you to finish telling us about these memory cards and then we'll move on to some of the AI ​​stuff. Of course . And yes, I think the best way to think about memory cards is that they are a repository of your knowledge and are simply the way these things resurface. This is the kind of card I really like. It's the only one I use, as in the last few thousand cards I've created. I haven't used any other. Then here to add the reverse, you simply decide whether you want to reverse the card or not. In the case of something like this, probably in the case of how many glasses of wine does Dan think are in a bottle? Six cups. You do n't really need the back, so we'll just leave it blank. And something extra would just be an image or something like that. You saw it on the other card. This is all you need. And always add a source. Uh, when you create a card, it's good to know, okay , this was in 2017. I spoke to this person and they said this, because again, there's a bit of nostalgia with these cards. Like, if you're really serious about making this a habit, right? You say, "Oh yeah, that was Naj on Carbon back in 2014 or something, right?" Perhaps you don't delete the card because it brings you a little joy. But I think that's very interesting. Actually, I haven't heard anyone talk about cards from that perspective. It's like, um, when people say they smell someone's perfume and it reminds them of their mother or something. And, um, use the cards as a sort of, um, evocative or associative and nostalgic exercise of different stages of your life. In the same way that, oh, I hear a song and think about when I was in high school or whatever. I love that. It's like there's something romantic about it. I think it's just because I've been doing it; it's an important part of my life, you know? I've been doing this since I was like 17 or 18, right? So it's probably been 12 years of cards. Um, you know, maybe , yeah, there's a lot of history there. Um, I think a lot of people use credit cards for a period of their lives. Um, but it seemed valuable enough to continue with it. Exact. It's one of the three things that have truly been with me . Brilliant. Um, do you want to make a couple more, or, um? Uh, I mean, yeah, let's do one where you get a good, good card that, you know, I want to try, I want to get that. Uh, what prestigious university is in Pittsburgh? Oh, you probably know this. Yes I know. I don't remember. Okay. Carnegie Mellon. Yes. You have to give me a chance to answer before you turn the card over. Okay. Okay. Okay. Okay. Um, then there's , you know, like, uh, English isn't my, my native language. Uh, but, uh, I used to make a lot of cards, um, with words and definitions. And well, basically I had a whole workflow where I underlined on the Kindle. It syncs with Readwise and then I process it as part of my highlights to create a memory card. At some point my wife told me, it's funny, my wife has two reactions when I proudly tell her one of these new words I learned, you know? One reaction is: How come you don't know that word? And the other reaction is: Simon, that's a silly word. Nobody uses it, and I tell her: Jen, it's like there's a third option missing here. Something like: Simon, that's an amazing word. I can't believe it. I am very grateful that you showed it to me . So, at some point, I started tracking the number of results for each word on Google as an indicator of how useful it is. That's something you could probably use a language model for today to ask how common a word is. So, affable, I mean, is someone who is, you know, an admired or pleasant person, yes, of good character, yes, that's another one, it's the main industry of Jilin province in China. Oh. There is a lot of good tea. This is a really difficult day, I don't know, I mean, this isn't a good memory card because it's too hard to read quickly. So I'm going to mark this and then review it sometime. Okay , I'm going to close this. But yes, there was a time in my life when I searched for the origin of every vegetable. That mattered to me because that way I could relate which vegetables were endemic to which cuisine, you know, before the Columbian exchange of tomatoes, you know, blah, blah, blah. Yes. So what's the answer? The answer is Asia. I don't think it's much more defined than that, which makes sense, right? It's not really used outside of that kitchen. That makes sense. Brilliant. Very good. So we've reviewed our previous interview. We've gone back in time and seen what has stayed and what hasn't. Now I want to talk to you about AI topics. So yes, how are you using AI in your, let's start with your work? This is definitely what I'm most excited to show you, because before this you told me that you don't actually use this tool. I would say that probably between 80 and 90% of my LLM usage comes from this tool called Raycast. Raycast is essentially a Spotlight replacement on steroids. So, when you open it, it looks like this. You can see my calendar, search my Chrome history, things like that, open apps, and search my files. But if you ask him something like, a question I actually asked him before this interview, is that I have this Yeti microphone here and the audio was very bad while I was testing it before our call. And I couldn't remember if we were supposed to talk on the side with the logo or the other side. So I asked him: for good audio quality, do you speak into the logo side of a Yeti microphone? You can't ask Spotlight this, but if you press tab, it will answer you right there . You simply press Command, space, type your question, tab, and you're done. Basically, what model is this? He is using GPT-4o. So basically you have GPT-4o accessible with a keyboard shortcut. That's how it is. You can open it in a full chat window. Mhm. And then expand whatever. There's a history here on the right. I hid it because there's probably something embarrassing there. You can keep asking him if the Yeti is a good microphone. Should I use it or buy another one? Right? It becomes quite an experience. You can upload files and all that here. I really like it for everyday use. ChatGPT for Mac and similar programs. I don't find them very interesting. It's a completely different workflow, but I was already using Raycast and this was a very simple way to do it. Let me show you, because there are actually a couple more things I do with Raycast. You can configure which model to use within Raycast. You can use Claude. If speed is very important to you, you can use one of the other models. I believe these are hosted on Groq, right ? This works really well. Hmm, the shortcut key to open that full chat dialog. Uh, you can do all that. That works really well. But you can also define these custom commands. So in here, uh, I have a couple that I use all the time . Um, for example, one I could use is, if I'm cooking something, I do n't really have a collection of recipes or anything, and it might be something I don't do often and I have to cook following various dietary restrictions. So I could do something like this to get a hummus recipe. Mhm. And this is like a whole prompt that I've made and it gives me this format, which is the most condensed list I can think of that respects my and my wife's dietary restrictions, things like that . It's just the simplest, most condensed version, and then I just put it in front of me or send it to my phone or whatever. That's great. Oh, and when you send it to your phone, is there a specific way you do it, or do you just copy and paste it into your notes app or something? Yes, I just send myself a text message or something like that. Uh, it's usually nothing more elaborate than that. There may be a faster way to do it, but it didn't bother me. Um, okay, great. So, the idea behind recipes is basically that you can create any common prompt you use. You can create a kind of recipe that is available via a command. And so, this recipe, uh, it's not called a recipe, it's called a command. It's called an AI command. And basically you put a recipe command in there so that, whenever you want to get a recipe for something, it has a very specific list of all the requirements you have. That's pretty good. I like that. Um, I use it a lot; That's why I try to share the things I actually use instead of the things I've experimented with. I use it all the time because I gave him a couple of examples. So yes, you go to Raycast, you look for AI commands. Okay, let's get to the recipe. We can edit the AI ​​command. And here you can see this is my prompt. I mean, please create a recipe for and then the past argument. Use the format of this example. Here is an example of how I wrote the way I like to prepare mashed potatoes. Um, but when listing the ingredients I put in parentheses optional extra ingredients, blah, blah, blah, blah, for the optional ingredients. This, uh, my wife is sensitive to this carbohydrate called fructan. Please provide substitutes. Please specify the approximate calories, like all these things. And these models are so good nowadays that this simply works. That's great. I love that. And you don't have to skim through the chef's story about how his great-grandmother's cousin made this recipe for them or whatever. No, there's nothing about how they found it on a clay tablet in the attic, is there? There's nothing like that. It's great. So, I really like that. And again, I think of the LLM as an internet average. And this is an excellent way to do it. And especially when you try to increase their creativity a little, like...I cook enough so that they only need the list of ingredients, not the instructions. Yes. And give me a bunch of options that could make it interesting and I can quickly see, oh, that's a fun idea, right? So you can adjust this to how you like to cook. Um, and I really like that. I no longer use Google for this because this is great. Um, I used to or still have this book called The Flavor Bible which is essentially a thesaurus of "hey, this goes with this". So you look up pumpkin and it tells you, um, maple syrup goes well with pumpkin, sage goes well with pumpkin, and ricotta goes well, right? And you start to get ideas of how to use this in a recipe. You probably wouldn't combine those things unless it was Thanksgiving in Canada. Um, but some of those things start to create really interesting things. Now, these LLMs, again, as an internet average, have that built in. Yes. What about improving writing ? What's your prompt for that? I do n't use it much. Um, but like, I haven't used it that much. Um, that's what it looks like . This is an experiment. So, I don't use this very much. Um, for...I find that standard things like simply "improve this" or giving suggestions are good. Normally, I dump all the text into Claude or ChatGPT or whatever I'm using that week . Hmm, and then I ask him, here's the complete document. Let's talk about this sentence. Give me, give me some feedback. That's what I've been doing for my blog posts lately. I don't have any other particular workflow for that. Any other commands you think are worthwhile? Define. This is one that I've spent quite a bit of time on and use a lot. Um, this is where I do my daily review; I still check out like 20 books, articles, whatever, featured each day on Readwise. And, um, sometimes they are isolated words that I underlined and that I don't know. And, um, I use this command to learn those words. The word I showed earlier on the card, which I think was "affable," and it just establishes, you know, like, um, a kind and gentle human being, that didn't come up for definition here. So what I use here in defining is, um, I'm reading a book and I come across this word, place, or person, and this is the command: please help me learn what this word, place, person, or whatever it is that this represents. Um, then, then I say: give me six example sentences using this word and please try to use historical examples, something that will teach me something. Give me something with familiar people, like physics, computer science, geography, to try to make this example sentence as educational as possible. I want to learn from the example. If it's a word that's always used in different forms, just take the root of the word. Um, also give me some related words, synonyms, concepts, things that are related to this word. Third, um, and finally, if you are able to, generate an image that works for this word. Then I give an example of, um, how this might look. Um, okay, Dan, give me a word. I want to use "affable". Affable. Let's do it. Uh, define. Oh, it's because it's doing it from my clipboard. So I'll do it that way and then define it. So, affable describes someone who is friendly, kind, and easy to talk to. And then, affable leaders like Gandhi often gain great respect and admiration because of their approachable and kind nature. In computer science, a user-friendly interface is one that is easy to navigate. Historical figures like Franklin were known for their affability. So it's like, I love this because it starts to, oh yeah. As if. Frank, it's like you connect with that trunk of knowledge immediately. And it makes creating a memory card much more fun , doesn't it? It's like, oh yeah , Feynman, like, oh yeah, I haven't thought about him for a second. Yeah, he seems like a nice guy, doesn't he? So you immediately make these connections, maybe you learn something. This has really made it more fun to look up some of these words. When I see a word now and I'm reading, I get excited to use this prompt because it works so well. I have a couple more words for you. I want to try this. Yes, let's do it. Let's try lamant. LAMBNT. Good. I have no idea what that word means. Something that glows or flickers softly, often implying a dim, radiant light , you know. Oh, you were indulging in some recent writing , huh? I love words. Uh, the flames of the lamp danced on the surface of the water, remembering ... Oh, I'm drawn to this. Isaac Newton once observed that the radiant glow of a candle could reveal the nature of light and color, which led to groundbreaking work in optics. Auroras in the polar skies are caused by charged particles from the sun interacting with Earth's magnetic field . TRUE. This is pretty good. It's good. I feel it's romantic. I feel that by reading that diversity of sentences, I will remember it better than just with the definition, which I think is what you're aiming for . Yes. I think I also have in mind making the images or sentences easy to visualize, because that's also a great mnemonic device. Like this lamp or the polar skies , I could remember it just for that. Yes. Wait, scroll down to related topics. I want to see if...Good. Resplendent, twinkling, radiant. Okay, great. Okay, I have one more word and then we can move on. Are you ready ? Yes. Eigengrau. EIGEN GRAU. G...Wait, I lost the... GR. No, that's J.G... GR. Okay. A U. Yes. Oh my God. This is...Okay. Good. Is this German? Yes, he's German. Intrinsic. Eigengrau is a German term that translates as intrinsic gray. It refers to the uniform dark gray background that many people report seeing in the absence of light. Often described as the gray matter of the brain. Um, isn't it great? Stana? Yes. How have you used this? I...I don't know, I just have a list of words I like and there they are, these two top my list. "I ga", although " I ga" is not an actual visual input. It highlights the role of the brain in creating our visual world, similar to how phantom limb sensations work in amputees. I think this just shows the power of LLMs , right? You take the average human knowledge and make it go crazy with associations, but orient it in a specific direction in the latent space towards educational and connected things. I simply love this prompt. It's great. It's incredible. Another one I used for a while is that, having grown up in Northern Europe, for a North American audience, my writing style is sometimes too direct. So I have this emoji suggestion that simply adds one. More user-friendly. Remove profanity. I don't know if it was me testing or if it was a problem at some point. They have a lot of standard prompts. It's not something I've used. Frankly, I've been very skeptical about these prompt templates and things like that. I only like the ones with recipes and definitions. Aside from that, I think LLMs are already good enough that there's not much to worry about. Yes. Yes. Yes, that makes sense. Brilliant. I love it. Okay, I know you also have some ChatGPT and Claude stuff to show. So let's move on to that. Yes, I think so...a large part of my usage. These days I'm subscribed to all the tools, you know? I feel that, being in AI, you want to have Perplexity , Claude, ChatGPT and you also pay for this. It's simply part of the business now. You spend $100 a month on various subscriptions, jumping from one to another, looking for inspiration. ChatGPT is what I mostly use. Now, I'm switching between Claude and ChatGPT, and since ChatGPT doesn't have a search function, I couldn't find the records. But I have some funny examples of the kind of things I use. Well, in our cabin we had a huge, very old freezer, and I don't see the point of a freezer in a place that only has electricity 99% of the time. Rural Quebec isn't known for that, in fact, to the point where I have a script that constantly pings and a webpage for the cabin that displays a GitHub-style uptime graph of how many times the power goes out. In short, you can't store anything in a freezer where the electricity is cut off every few weeks. So I wanted to turn it into a refrigerator. And I thought, "Ah, maybe it'll be a fun project." You know, you have plenty of time before a baby is born. And I asked ChatGPT, "How could I do this?" Because I've never done anything like this before. I'm not very skilled. And ChatGPT told me, "Well, basically the standard thing to do is go to Amazon and buy this device that's widely used for home brewing." You connect it to the wall , plug the freezer into it, and place the temperature probe inside the freezer. And then you just adjust it to 5 degrees, right? Celsius. I don't know what that is in Fahrenheit. "And then the freezer will turn on when it goes above 5 degrees and turn off when it goes below." And now your freezer is a refrigerator. I thought I was going to buy a refrigerator for about $1000 or something like that. I have this now and yes, it causes a lot of condensation because the compressor works very hard, but it does n't matter because this refrigerator is only for drinks. It's fantastic, a $20 device from Amazon doing the conversion. I would never have thought of doing that. It 's incredible. Wait, explain it to me again. So, do you take the temperature probe from the device you bought or the one from the original freezer and put it somewhere else? You have the freezer, you plug it into the wall and now it's a freezer. Now unplug the freezer and plug in what you can essentially see as an extension cord. You connect the extension cable and plug the freezer into that cable. The extension cable has a temperature probe. You put the temperature sensor inside the freezer, right? And yes, it simply turns on and off depending on...you can even configure it to cool and also heat, if you have a device capable of doing so. But I think it's used for homebrewing, isn't it? Where you need fermentation ranges. That's incredible. Is the sensor wireless, or how do you put it in the freezer without breaking the seal? Emm, break the seal. OK. You know, I think especially in rural Quebec, you become very resourceful. So this is a good trick; You know, that's how the best software is written too. Yes, I love it. And well, what else, what else do we use for that? I mean, since we're talking about cabbage, writing Quebec French is quite an art. I don't know, again, I don't know French, but my wife uses it all the time to translate things into Quebec French. The other thing, the most important thing for what I use it for in business, I think, is that when reviewing changes or documents with clients, it 's often easier to send the lawyer a draft and say, "Hey, something like this." And then they edit it. I use it constantly for drafts . I think for most people it's much easier to edit something, especially if they don't know much about the subject, than to start writing the first draft. And for something legal, that 's certainly the case, right? So you say, "Okay, I need to explain the exact algorithm we use to measure Turbuffer uptime in a way that we think is logical, because I don't think what the lawyer came up with made any sense." Let me send it back to you in legal language . And that way you minimize round trips. Likewise, even when I talk about accounting and things like that, I use it constantly because I don't remember what this term means, and then it goes back to the memory card process over time. But often, many of these professionals talk to you as if you already know everything about accounting, laws, and whatever else. So when you have those conversations with suppliers who are in another sector, I find it extremely useful, and then it ends up being integrated into the memory card syncing later on. Yes, it happens to me too. I feel that , for example, for a lawyer, I need to be pushed into the latent space of legal language. And once I see that language, I can write in it, you know? But just having one example of something similar to what I want is enough. And that's something I think about a lot with ChatGPT and Claude; It reveals how many dialects of English exist, because now you can make subtle translations between different dialects that we didn't even think were dialects. Like going from technology expert to lawyer, or from small business owner to, I don't know, painter or whatever. We didn't think those were different forms of English, but they really are, and ChatGPT is like this amazing universal translator for those kinds of translations. I could n't agree more. I think you've expressed it brilliantly. It's like I have no idea how to access the legal latent space unless someone puts me in it, right? And then I can edit and use terms like "so far," you know? And it's like, yeah, go ahead. And I think iterating over a text, especially if it 's a very concise text, right? So it's great to just receive suggestions. For this kind of thing, I'm not using any particular tool. Depending on the context, I'm either iterating within ChatGPT or Claude, or just using Raycast, you know? Like, "Give me other examples." I rarely end up using exactly what it throws up, but again, it comes up with words and things like that that I can use. The other thing I've found it very useful for is physiotherapy exercises. I had a problem with tennis elbow or golfer's elbow, I think it depends on which side it's on. Don't know. Ask ChatGPT. And it's like, well , I'm going to do an experiment instead of going to the physiotherapist. I'm just going to ask ChatGPT, I'll do what he tells me for a week and see if it disappears. And it worked, right? It was just like, oh, you have to do these wrist flexion exercises. And I said, okay, great. That saves me a trip and like $100 in physical therapy, right? And I also discovered that with one of the problems I have, and I think many people who work at a desk have, which is having very tense shoulders and neck . I always thought, okay, yeah, I just need to stand up longer and do stretches and stuff like that. And at some point I thought, I've had this problem for five years and it's not going away, right? So I asked ChatGPT and again I said, okay, let me do these exercises that ChatGPT gives me. I'm not even going to try to understand them properly. I'm just going to do them blindly for two weeks and it's been much better, right? Everyone says, "Okay, relax the tissue." But it's more like, "No, it actually strengthens these muscles, right?" So I think it's pretty good for that. I find that for things like exercise and such, you have to guide it in a pretty precise direction. And it's not always incredible when you reason about how you arrived at that conclusion. But again, as an average of the internet of things that work for this condition, it's pretty good. Another problem I had, and that no one could help me with, is that I have something strange where I lose my vision for three or four hours and I become almost blind. It happens randomly every few months and I couldn't figure out what it was, and when I spoke to the optometrist, they just said, "Oh, it's just stress." And I said, "Well, then, is this going to be a problem for the rest of my life?" "Isn't there anything I can do?" Because it's quite debilitating, especially if it happens at a bad time, like while driving or something like that. Then I just have to pull over and wait until it passes. And apparently it's called ocular migraine and it can be caused by the strangest things. One of the triggers for me is aspartame. It's like, and I think after that, well, that's a little worrying. But, well, that's another story. I mean, ChatGPT has become what I consult all the time, and I think Raycast on the computer has made a huge difference when asking anything. And on my phone, on the home screen, I have ChatGPT voice chat . My wife uses it a lot. He's always asking questions; He uses it a lot for gardening. That was great and I'm really looking forward to their next release. Do you already have access to that? Is it as good as it seems? It's really great. Yes, I think if you 're already a big user of voice mode , you're going to love this. I think I'm also very excited about that part for my daughter. I've seen that there are some toys that allow you to talk to these models and ask them questions. I feel like she'll grow up with these tools in a way that will make it feel incredibly natural for her to just be talking to Wall-E, the walrus, you know? And there was that really cute clip, right?, where it's tuned in to the Golden Gate Bridge. And hopefully we'll get there. That's what excites me most about safety, thinking, "Okay, I can give it to my daughter and she'll just ask and this will help with her curiosity?" That's really interesting to me. I think you'll love the new voice mode because I have a video about it. I'll send it to you after this via YouTube, where I use the new voice mode and it's very good for reading, because what I do is turn it on, it stays there listening and, when I read something and find a word I don't know. I say to him, "Hey, what does this word mean?" Whether it's a historical figure or I need more details about this particular battle in the book, or about some philosophical concept that I do n't quite understand, it simply gives me the answer. And it's surprising, when reading, how often you come across things that you say: "I don't really know what that is, but I have a feeling, I'm just too lazy to ask." And when you have a chatbot-type voice assistant listening, the barrier is reduced so much that you simply ask all these questions and I think you learn much more. It's, it's really fun. I like that a lot. And especially if, I mean, APIs are going to be created about these things, right? But it's like that I could create highlights and then bring them to my Readwise or whatever later on, that's pretty exciting to me. I think what I've always wanted, and I think it will take a while to get there, is that I'm not really interested in VR and AR as an entertainment device. There are two things that excite me about it. One question is: can I use it instead of my monitor setup? When will this be good enough for me to use all day and program and work within it, right? That's exciting for me. For video calls, maybe... I don't know when that will stop being awkward. That doesn't excite me very much. But this second thing that has always excited me is that the visual stimulation of VR and AR would help with remembering. You remember much better. It's a way for me to learn, which is quite interesting. Again, unless I'm already in this work environment, I don't know how likely that is. But if I can, you know, we're talking about these words like "Ieng" and which one was the first with the light, right? If I can see that, it will be very difficult for me to forget it, right? If we can generate those types of images. So that's also very, very interesting to me. And, of course, these things will work together. Um, and that might take a little longer. Again, I do n't really care for any other use case besides that one, but those two use cases do excite me. Um, and it seems that's starting to become the possible adjacent one. That's the thing I think people overlook about AI and perhaps about how technology interacts with human beings in general. A really good example is simply the ability to read. For example, it actually changes your brain and takes some of the material from your visual cortex and reorients it to help you read, and that makes you better at analytical thinking. It makes you more likely to, you know, see details of a scene instead of having a more universal and holistic perspective. In other words, it actually changes humans so they can read. And I think having language models will do something similar in this regard, and I don't think that's a scary thing. I think it's a great thing, for your daughter or my nephew who is a year and a half old, or almost two now. Living in a world where any question you ask gets an immediate answer , and nobody gets annoyed with you for asking, is an incredible boost for children's brains, you know? Because before, well, a one-and-a- half or two-year-old is still a little small, but three- or four- year-olds ask all these questions and the parents say, "I don't know" or whatever. I don't know why the sky is blue or whatever . And all those questions are answered. And perhaps not even answered. It's like you're entering a scenario that helps you understand it in a completely new way. And I think people are worried that AI will replace us or if we'll have to be bionic or something like that . Actually, we don't even have to implant it in our brain. We will simply be different people . We will prosper in a new human way that was previously impossible because the conditions were not right. And that excites me a lot. Yes, I'm very excited about it too . I'm sure you too were one of those kids who drove their parents crazy with questions at some point. I remember asking my parents, I think I was about four or five years old and it's one of my earliest memories, when you just ask questions like: "Mom, what is the biggest plant in the world?". TRUE? And they just kept saying, "Oh my God." TRUE? I think they were saying to me, "I do n't know, Simon." "Be quiet". Know? And then I remember I got the Guinness Book of World Records, you know? And it was like, "Well, this will keep you quiet for a while, right?" And now, well, now I know who also has the biggest nose piercing . Brilliant. You know, um, but I think , I think that would be very stimulating. I think it would be very stimulating for many children. Um, I also think that one of the things, you know, is that my native language is Danish. Um, my, you know, my grandparents are thankfully still alive, um, and they really only speak Danish, um, but it's very important that my daughter speaks Danish too, but it will be a challenge, um, for me being the only one who speaks it here, right? I live in Canada; Nobody around talks about it; It's a small community, um. I don't really have any friends here who speak the language, so aside from FaceTime with her Danish family, there's not going to be much exposure. So yes, we'll set up, you know, all the user interfaces in Danish, um, but maybe we can also set up, you know, Wall-E the talking walrus, right? Supported by any model to speak to her only in Danish. Um, or maybe I should speak to him in Mandarin or, you know, or like Thai or whatever. I think one of the things that might be interesting for this generation is also, you know, I don't know how true this is. I haven't read the studies on it, but I feel like, um, if all children just learned languages ​​before the age of 10, they could catch up on all the math and whatever they missed, you know, when they're 10 or 11, in about four weeks. Uh, but the language and hearing those sounds is just very special. Um , and, um, you know, I can't say "th" and that will probably follow me for the rest of my life because I just wasn't exposed enough to that sound until I was eight or whenever that thing gets fixed. Um, and I think that would be , that's exciting if Wall-E the walrus speaks this language on Tuesdays and this other one on Thursdays, but mostly Danish, right? Um, I don't know how that changes things, but I think it's interesting. One of the things, just about pronunciation , that I find funny is that, again, because ChatGPT is an internet average. ChatGPT in Danish actually has an American accent. Um, I don't know how true this is in other languages. I'm sure that for French or Spanish or whatever it's really good, but the Danish guy has an American accent that's just hilarious. That's crazy. I promise we'll come back to some of your AI use cases in a second, but I think this is too good not to share. I think you're going to love this. So, I came across this startup probably a year ago. And his whole concept was, you know, you said you can't pronounce the "th ". And I think that might actually be more flexible and malleable than you think. And the reason why, at least this startup said, you can't say "th," is that it's very difficult for you to hear and notice how far you are from the real sound because you didn't learn that pattern growing up. And what they did was create a tool where, if you were practicing, for example, the "th", it would show you the sound waveform and then you would speak. You would say "th" or whatever, or any word, and a dot would appear showing you in real time how far you were from that sound, and you would just practice all the time, and you could see in real time these micro-increments as you got closer and closer, like tuning a string. And apparently, if you do that, even non- native speakers who have already passed the critical period can learn to speak fluently, and I think that's incredible. It's a crazy way to approach it. I think at this point there's also a bit of attachment to not getting rid of it completely. I think any native English speaker can hear that I wasn't born in the New World . But, you know, it's something that's part of my own story. So I'm not going to make an effort to change it, but it's very interesting that it's possible not to have it. And there are sounds that, if you can't pronounce them, are really problematic. Sometimes I sound a bit like a little kid when I say "th" and talk fast, don't you? It becomes an "f". But I definitely identify with that, I mean the International Phonetic Alphabet, right? It is, in some ways, a truly fascinating read if you have n't read about it before. But it has this chart, and I love good charts that simplify something very complicated into a system that actually makes sense, like the periodic table, or, you know, like the way the planets are arranged around the sun. These are all systems, and a good system not only categorizes existing knowledge, but also predicts future things . Like, the periodic table that says, "Well, there could be a heavier element , or we're missing one here," or whatever, right? And the International Phonetic Alphabet is that, but for sound. It's like, the back of your tongue is here, the middle part is here, the front part is here, your nose does this, and it's all these parameters, and now you're producing this sound. So these permutations exist, and every language has, you know, I'm just making this up, but I think it's between 20 and 40 sounds. And then, different dialects also have different types of sounds. For example, in Danish, we have this sound which is also in my middle name , right? So, my middle name is H. This is not a sound that exists in most dialects of English. You probably can't pronounce this sound. Maybe if you... Yes. Yes. Good luck. And, you know, one of my friends said, "Simon, saying your middle name is a bit like trying to throw up while saying Europe ." And that works sort of, you know , not so much when you have to spell it over the phone, but it works for pronunciation. But what I discovered when I was researching the International Phonetic Alphabet is that , oh, actually in New Zealand English, they have that sound. So they say, I can't imitate it well , but they say something similar to "bird" with the vowel very drawn out. So they can say the sound, whereas in American English you say "bird," right? It's more like an " i" sound. And French has it too, so it's easier for them to say it. But it's as if each language were just a mixture of these 20 to 40 syllables. And I remembered this again when we were choosing my daughter's name. Right? Because it's like you want to choose syllables that are more or less the same in both languages ​​and then combine them. So, you know, Simon is Simon in English, but in Danish it's Simon, right? And it sounds very different. Um, these syllables are um are very finite and each language is like they come together and, whatever it is, 40, 20, whatever you learn, um it's difficult, but it doesn't sound irreversible. Yes, I love that. I had no idea they had mapped it and, as you can see, it's like a periodic table of all the different sounds. That's truly amazing. Um, that, that kind of thing, makes me a nerd. Um, I want to delve into a couple more things about AI that you're going to share. I know you said you use Notion AI a lot for writing. Tell us about that. Yes, this is something more recent and, generally, when I talk about tools, I like not to mention them until I've used them for several months . But I think that with AI tools, everything is changing so fast that it's worth sharing some of these things beforehand. Typing within ChatGPT feels a bit awkward, like iterating over the writing. I think Claude is trying to change that with the artifacts. Um, but Notion AI, I mean , that's where I do most of my notes and stuff these days. Um, and I just have a journal page where I write about whatever is going on or try to solve a problem. Um, and they have a really good integration that brings the necessary context, you know? So, they're using some kind of, you know, the kind of thing that powers Turppuffer, right? Um, they're using some kind of semantic search behind the scenes to extract context from your entire workspace. Um, so when you write and iterate over your writing, it can bring all of that out. And that's super, super interesting. And I found it great, like, "Hey, I was having this discussion with someone and I feel like maybe I didn't represent myself well," and it gives you feedback and completes the writing. That kind of conversation has been very, very valuable, and I think Notion AI does it quite well. Um, so I thought it was very valuable to use it. Yes. Yes. I actually use Notion AI quite a lot , uh, for writing, particularly, for example, to prepare for these podcasts. Uh, I often just have a Notion document with the script and, for example, when I had Tyler Cowan, I just said, "Okay, tell me what Tyler's most recent books are." Give me a summary. Exact. And then I said, "Here are some points I'm going to cover, tell me which ones relate to these books ." And it allowed me to create a document very quickly with my ideas and a summary of his books and ideas so that I could talk to him about it in a podcast. Uh, in a way that, yes, I totally could have done it with ChatGPT or Claude, but it would have been difficult. It would have been much more difficult to do. Having it in context is very useful. Definitely. I think, I mean , that's also why we created, uh, part of why we created Turbuffer, right? Is that it? It's clear that, within the context window you operate within in the model, we can achieve a lot, and I think we haven't even tapped into its full potential. But I think Notion is showing that, hey, we can actually extract a lot more relevant context from years ago , from other people and things that have happened, uh, and maybe even changes in how documents have evolved , and you can get something even more interesting. Uh, and I think it's an exciting future, it's an exciting future. I think there are a lot of safety and security things that we need to get right. Uh, but that can really help to increase it, uh, and I think Notion AI is a very good example and probably one of the most advanced tools in that regard. Yes , that's something I like to do: I take a bunch of journal entries and throw them at Claude or ChatGPT and say, " Okay, how have I changed over the last few years?" What is my background or what do you know about me? And all that kind of stuff. Or, you know, if I'm writing in my journal about a decision, and I'm going around in circles about what to do, I take all my journal entries about that decision and say, "Write the entry as if I had chosen A, or write it as if I had chosen B." It's very useful for that kind of thing, for capturing the details in your way of thinking and speaking , and then projecting them into the future. Absolutely. And tools are emerging ; in fact, one that Turbopuffer uses is this application called Dot, which attempts to create a journal where you extract the necessary context from your life to basically do what you 're mentioning. We're also seeing another company that uses Turbo Puffer, which is this company called Shapes, right? Where you have these forms with personalities that you can talk to, they join your Discord, you chat with them, and they remember what has happened and the dialogues you have had. And that, I think that's really interesting. And I also think that context, the windows of context, are growing. We're getting better at extracting context from billions of documents to do really interesting things, and the tools will grow along with it, but it will take time. I think that even now, we're not even using the small amount of completion that Raycast offers on the computer, something that most people don't take advantage of. Most programmers don't even do much beyond using Copilot to complete code, instead of having a real conversation with it. So it's an evolution of both us and the tools, right? But I think it's very, very exciting to be a part of this. Completely. So, what else should we talk about? I think there are a couple more areas where I use LLMs on a daily basis. One of them is this tool called Super Whisper. Again, it's one of those tools that I haven't had the pleasure of using much yet. But it allows you to speak and then summarizes what was said. I'm experimenting with using it a bit more for journaling . I just tell him: "Hey, these are the things that happened today or this is what I'm going to do today." Well, I haven't fully incorporated it into my workflow yet, but I find it interesting and the tools are constantly improving . The main problem I still have is that it takes a little too long to transcribe and pass it on to the LLM to make the summary. I wish it took a few hundred milliseconds, but it takes a couple of seconds with the accuracy I can tolerate, I guess, with my accent or whatever. So I think it's interesting, but the tools are still a bit slow. But that 's fantastic because it will probably be resolved in the next six months. Do you use it? I do n't do it. But I've heard, I've heard very good things about it. I simply haven't tried it myself. Yes, I think it's still a bit immature unless you really want it to be. Or I think that's also true; I've been glued to my keyboard since I feel like I was seven years old. So I can type faster than I can speak. So it hasn't been a big problem. But I think for a lot of people that's huge. And I think that's where we're going to see more of that, and hopefully the operating system's dictation will improve in iOS in this next version. I'm not on the beta version, so I don't know, but I guess so. And when that reaches the workflow of how we write messages and things like that, then I think it will spread further. But it's like I 'm right on the edge, I think. And it will take a little while, I think, before it reaches everyone, but it's definitely moving forward. Yes. What I want to say is, well, people are sending voice notes right now, and I think they're really useful when you want to communicate something that's a bit complicated and you don't want to sit down and structure the whole thing. But what I want is a voice note in which I have a conversation with an AI and then I can send it to someone else to explain it to them simply, perhaps first in text, as a summary of what I said. So he takes care of all the structuring for me. But also, it has all the background context of my entire conversation, so , well, if people have follow-up questions, they can just ask the AI, and it will be like, " Well, you know, in another part of the conversation that didn't make it into the structured summary, we covered that, and here's what Dan said." And I feel like that would really cut down a lot of emails, Slack threads, texts, and back-and-forth, when you have something smart enough to answer basic questions about, you know, usually when you communicate something you're only communicating the tip of the iceberg. But I think AI allows you to send the full context, present only the tip of the iceberg, but then reveal different parts as needed. Yes. So essentially you're just sending... You send the summary and then the little template comes with the FAQs, right? Yes. And it has a confidence threshold regarding whether it can respond with the context it has or whether it should be escalated. I think that's an interesting idea. And I think now, since I use Superhuman for email, I've started using their prompts more and more to reply to emails, right? And now they're releasing features where you can ask the AI ​​questions and things like that. And I think that's it, I hadn't thought about what the next step would be, had I ? But this almost feels a bit like a support agent, doesn't it? Where, you know, he has a certain confidence about when to climb or not. The challenge will be to make it still feel authentic, right? It needs to be part of the client where you send this together and it's like, hey, I noticed you were going to ask about this, actually this is probably the answer. Do you still want to ask? I think that's super interesting. Yes, I think we're still very much in the early stages of how this is going to be used. I suppose there are perhaps two schools of thought, one that says we've arrived at the IAG and who cares, and then there's the school that says, well, the tools will continue to improve and people will learn how to use them. I don't know what timeline we're in. I think it's a lot of fun to be part of timeline number two for as long as it lasts, and that could be a very, very long time. Or maybe not. I don't mind. I'm having fun , you know, in that timeline. Uh, so I think that's very interesting for the messages. Uh, the other thing I use LLMs for a lot is simply through the Readwise reader. So, when you read, you know, there are things you do online: look things up, define words, ask the document questions. Oh, and then of course, within my editor, I use a tool called Supermaven to do autocompletes. Oh, and then I use some plugins for Vim. Uh, to try and emulate a bit what the Cursor editor does. Uh, I think the Cursor editor is by far the best AI code editor. Uh, they actually use Turbo Puffer behind the scenes. So they are very, very good friends. Uh, and I really want to use Cursor, but I can't stand the latency of VS Code. So I'm still on NeoVim. So, I put together some plugins to, you know, send context to different models. Uh, but I hope that gets sorted out . Either I'll switch to Cursor at some point or I'll use some of these plugins, but when you've used Vim for 15 years, it's very difficult to switch to something else. That's great. Yes. Uh, I use Cursor. I love it. Uh, I'm not that, uh, not that "Vimero"...what's the word for that? Uh, I'm not as proficient with Vim as you might be. Uh, but yeah, I think Cursor is amazing. It's a bit strange, like when it makes suggestions if you schedule it across multiple files, you have to click on the file and then click to make the change or whatever. So there are still some things to work out there , but it's clearly the future of this type of interface. I think they're demonstrating, you know, that every few months they release something new saying, "Hey, this is how we're going to do it." And they really do seem to be paving the way. It's like one of the most popular plugins right now is trying to do Cursor-style things in other editors. Well, I think they're doing some really interesting things in that timeline of simply making the tools better . And I don't think so, you know, venture capitalists tweet that these companies are ten times faster because they use AI , but I don't think Justin uses any AI. It's still very... I think it helps and increases, especially for people to create that first draft and have the conversations, but I don't think it's that massive change yet. But it does come; every month these things improve by a few percentage points. But I do n't believe in that "ten times better" thing . That is certainly not the case . I think it depends a bit on who you 're talking about. If you're already a ten times better engineer, maybe not. But if you don't know how to program at all, I think it makes you ten times better than you were before. In other words, infinitely, right? Because you simply weren't going to bother. And yes, I think that's true , don't you? It's the same as saying, "Okay, now I can become an expert on retaining walls to talk to my contractor." So I think it elevates someone from a novice, or less, to being able to converse with an expert incredibly quickly. But for the experts, it's mostly about getting things done a little faster when writing and replacing a conversation about what data structure to use or what algorithm to choose, or how to make those kinds of trade-offs, right? But much of it also comes from things that are not the average internet content and are therefore very difficult to discover. Yes, absolutely. Well, I feel like I could keep talking to you forever, but we're almost out of time. Simon, as always, was a pleasure. We need to do these interviews more often. Thank you so much for coming and telling us what you 're working on. Of course, Dan. That was all. It was really fun to go into geek mode. Brilliant. Until next time. Until next time. My God, friends. You absolutely have to hit the like button and subscribe to AI and I. Why? Because this show is the epitome of incredible. It's like finding a treasure chest in your backyard, but instead of gold, it's filled with pure knowledge bombs about ChatGPT. Each episode is a rollercoaster of emotions, ideas, and laughter that will leave you on the edge of your seat, wanting more. It's not just a show, it's a journey to the future with Dan Shipper as the captain of the spaceship. So do yourself a favor, like, subscribe and get ready for the ride of your life. And now, without further ado, I'll just say : Dan, I'm hopelessly in love with you.