Interview with Yohei Nakajima, creator of BabyAGI, the first open-source autonomous agent to go viral in March 2023, and general partner at Untapped Capital. The conversation covers BabyAGI's evolution through seven iterations, his coding agent Ditto (demonstrated live by building a Snake game and a scheduling app), a screenshared demo of BabyAGI 2o—a self-building autonomous agent—his philosophy that built tools are extensions of ourselves, modular thinking for AI founders, and a future where models train themselves as they're used. # Channel Identity Observations — "Building AI That Builds Itself - Ep. 35 with Yohei Nakajima" ## Video Format & Presentation - **Interview/podcast format with live demos.** The host conducts a long-form conversation with a guest (Yohei Nakajima) and transitions into on-screen demos: "can we see a demo I want to see baby AGI 2" followed by live screen-shared building in Replit. - **Host name and show identity stated in outro:** *"a journey into the future with Dan shipper as the captain of the spaceship."* The outro also names the show: *"subscribe to Ai and I."* - **Episodic numbering** in the title ("Ep. 35") indicates a recurring series. ## Explicit Self-Description - The outro (apparently a scripted/gimmick sign-off) makes direct claims about the show: > "this show is the epitome of Awesomeness... filled with pure unadulterated knowledge bombs about chat GPT every episode is a roller coaster of emotions insights and laughter" This is creator self-claim, not demonstrated evidence. - The host states a norm of returning guests and follow-ups: *"we should do another one of these when 01 gets tool calling"* — indicating continuity of the series. ## Intended Audience - **Technical AI practitioners and AI-curious builders.** Demonstrated by deep technical content: discussions of tool calling, model versions (davinci-002, o1 preview, Claude Sonnet 3.5), agent loop architecture, and prompt design. - **VC/startup-adjacent listeners.** A significant portion covers investment strategy: the host asks *"how are you thinking about investing in companies"* and the guest discusses pre-seed investing, valuations, and portfolio companies. - **Beginners are accommodated lightly.** The host provides plain-language framing during demos: *"for people who are watching basically like you've got repet open you pressed run now we have a little website."* ## Claimed / Demonstrated Scope - **Scope demonstrated:** AI agents, autonomous coding tools, LLM capabilities, AI's effect on personal productivity and self-understanding, and AI startup investing. - **Recurring themes the host signals interest in:** - AI as metaphor for the mind: *"language models give us this new metaphor by which to understand how our brains work."* - Fragmented-attention coding: *"turns coding into into a task that only requires that level of attention."* - **Format scope includes show notes / follow-up references:** *"we'll put it in the show notes."* ## Audience-Engagement Practices - **Explicit call to action** in the outro: *"you absolutely positively have to smash that like button and subscribe."* - **Hosting style is conversational and complimentary**, positioning guests as notable figures: *"you are like one of the coolest online AI tinkerers in the whole like AI tinkering space."* ## Notes on Evidence Quality - All observations above derive from the transcript only; no curated metadata was supplied for this video. - The outro's promotional language ("epitome of Awesomeness," "knowledge bombs") is treated as **creator claim**, distinct from the demonstrated long-form technical interview format. # Delivery Observation: "Building AI That Builds Itself - Ep. 35 with Yohei Nakajima" ## Conversation Format & Roles - **Two-speaker interview structure**: Dan Shipper (host) interviews Yohei Nakajima (guest, VC and builder of BabyAGI). The host drives topics with frequent questions; the guest explains, demos, and reflects. - **Live screen-sharing demos**: Roughly the middle third of the conversation is a walkthrough of Ditto and BabyAGI 2.0 in Replit, with the host narrating what viewers see: > "for people who are watching basically like you've got repet open you pressed run now we have a little website that says flask app builder" ## Tone & Register - **Casual, enthusiasm-heavy, tech-insider register**: heavy use of filler words ("like," "you know," "um"), slang ("blew up," "kicked off this hype wave," "low hanging fruits"), and rapid interjections ("whoa," "that's so cool," "huh"). - **Mutual affirmation is a recurring pattern**: the host repeatedly praises the guest ("you're like one of the coolest online AI tinkerers," "that's really cool"); the guest deflects with self-deprecation: > "thank you you're too kind um I'm I'm just really lazy" ## Guest's Communication Habits - **Self-deprecating framing of productivity**: Nakajima repeatedly attributes his building to laziness or hobby-status rather than ambition — "I'm just really lazy and whenever I'm working on something that I don't like doing... how can I cut this out of my work" and "this is more like an incredibly fun hobby." - **Concrete, numbers-and-names specificity**: he anchors claims with exact details — "500 rows of python," "seven iterations last year," "20 to 40 page industry reports for me in 30 minutes," "10 to 12 or 11 to one." This is a stable pattern across answers. - **Narrating failure openly during demos**: he admits when demos don't fully work — "it has worked previously before must be at the JavaScript somewhere" — a habit of transparent, non-defensive reporting. - **Explaining architecture step-by-step**: when asked how the loop works, he gives a sequential tool-by-tool walkthrough: > "the tools it has is create directory... create file... update file... fetch code... it'll call a task completed tool which actually just exits the loop" - **Hedged uncertainty**: he flags speculation explicitly — "I suspect it would be even that much better," "I'm not fully following it, but..." — distinguishing known facts from guesses. ## Host's Communication Habits - **Abstraction-elevating questions**: the host repeatedly pushes concrete examples toward general principles — "help me understand conceptually what's happening," "how meta can you make it," "do you know what I'm getting at am I being crazy." - **Historical analogies as an explanatory device**: an extended monologue compares minds to wax tablets, steam engines, and computers before introducing LLMs as a new metaphor ("I already hear people saying like oh that's not in my training data"). - **Interruptible, collaborative style**: frequent short confirmations ("yeah," "right," "huh") and playful interjections, e.g., "are you afraid it'll take over the world." ## Recurring Themes & Phrases - **"Loop" as a central concept**: "loop" recurs throughout — "single Loop through an LM," "it just Loops through until it deems the task is complete," "a single llm loop with I think three tools." - **Self-reference/reflection vocabulary**: "self-reflection process," "extension of myself," "working on myself," "understanding myself better" — both speakers use mirror/self language repeatedly. - **Tool-composition framing**: the guest repeatedly frames skills as composed hierarchies — "the actual tool to provide the agent was this wrapper tool around two smaller tools" and "my Google search skill depends on my... typing on a keyboard skill." ## Teaching Approach - **Demo-first pedagogy**: rather than abstract explanation, concepts are shown by running code live (Snake game, call-list app, TechCrunch scrape) and then explained after the viewer has seen the result. - **Build-in-public ethos as explicit method**: the guest frames releasing code as his version of content — "some VCS will learn things and write blog posts about it... this is my version of that... but instead of blog posts code." ## Isolated Moments (Not Patterns) - The closing outro is an anomalous tonal break — an over-the-top promotional pitch ("you absolutely positively have to smash that like button") followed by "Dan I'm absolutely hopelessly in love with you," delivered in a mock-advertising voice unlike the rest of the episode. This appears once and should not be treated as representative of the overall delivery. - The anecdote about the Discord of AI bots and Marc Andreessen's $50k stake is a one-off digression, introduced with "totally random but related idea." ## Uncertainty Notes - Timestamps are unreliable for precise citation: the transcript's bracketed indices are sequential markers, and one segment appears duplicated/transposed (the "three kids... 75 and two" passage occurs at both [0] and [1960]-[1971] regions), so ordering of some exchanges is uncertain. - Paralinguistic features (pace, laughter, pauses) are not recoverable from this transcript; tonal claims above rest on lexical evidence only. # Position Extraction: "Building AI That Builds Itself" — Ep. 35 with Yohei Nakajima ## Speakers - **Yohei Nakajima** — General Partner, Untapped Capital; creator of BabyAGI (identified in transcript by host and self-introduction) - **Dan Shipper** — Host (identified via outro reference and interview role) --- ## Background & Motivation ### On being a prolific tinkerer - **Nakajima:** Describes himself as "just really lazy" — before AI he was a heavy no-code/Zapier user; LLMs "unlocked so much," and his approach is tackling disliked tasks one at a time to remove them. - **Nakajima:** On why he builds rather than making it his full-time job: "I love being a VC... Founders are my favorite people in the world." His open-source projects are his version of what other VCs do with blog posts — "but instead of blog posts, it's code." - **Nakajima:** Works only at night and weekends (roughly 10pm–1am), in short bursts, often copy-pasting between ChatGPT/Replit on mobile while doing chores. - **Shipper:** Agrees, noting Devon-style agents let him run multiple parallel tasks needing only "Slack-ping" levels of attention — coding no longer requires deep uninterrupted flow. --- ## BabyAGI & Its Evolution ### Origin and impact - **Nakajima:** BabyAGI (March of prior year) introduced the pattern of looping an LLM: generate a task list, parse with code, tackle tasks one by one. He attributes its popularity to simplicity — "everybody who saw it could think of ways that they would make it better." - **Nakajima:** He wasn't originally trying to build an autonomous agent; he was prototyping an **autonomous founder**, inspired by HustleGPT — "if they're doing whatever ChatGPT says, why can't that be an AI as well?" - **Nakajima:** From his fund he invested in agent-space companies (e2b, AutoGrid, "a few more"); believes AutoGrid/Cognosis built something shortly after BabyAGI launched. ### Iterations and current state - **Nakajima:** ~7 iterations last year with animal names (BabyAGI, BabyCatAGI... BabyFoxAGI), each introducing a new design pattern. This year: **BabyAGI 2**, a framework for *self-building* autonomous agents — agents that build their own capabilities to improve themselves. - **Nakajima:** Also released **Ditto**, a ~500-line script building multi-file apps — "a little poor man's Devon." ### Effect of o1 and better models - **Nakajima:** o1-preview is "incredible at handling multi-file edits"; he couldn't have built BabyAGI 2.0 (full framework with frontend) without it in his available time. He hasn't used o1 on the newest projects because it lacks tool calling, but "suspect[s] it would be even that much better" once integrated. --- ## Demonstrations ### Ditto (coding agent) - **Nakajima (demo):** Single Python file with a single LLM loop and five tools (create directory, create file, update file, fetch code, task completed). Generated a Snake game and a "track your friends" call-list Flask app from one-line prompts. - **Nakajima (qualifications):** The Snake game "does not take keyboard controls... it has worked previously" — the demo did not fully work; he notes the script is one-time-use (must delete outputs to run again) and "not as good as Replit." - **Shipper (conceptual question + Nakajima's answer):** Planning is no longer a separate LLM call — long context lets one prompt include "before coding, carefully plan on all the files." ### BabyAGI 2.0 (self-building agent) - **Nakajima:** A single loop with three tools: **create/update tool**, **install package**, **task completed**. Asked to "scrape TechMeme," it wrote its own scraping tool, failed, rewrote it twice, and succeeded by iteration 5–7. - **Nakajima:** Key limitation — the agent throws working tools away after use. The right approach is storing the tool that worked for reuse, plus "a public AI tool library" any agent could draw from. - **Nakajima (BabyAGI 2 proper, distinct from 2.0):** A framework for storing and executing functions from a database, with dashboard, function dependencies, triggers. His next step: combine the 2.0 loop structure with this storage framework. --- ## Reasoning About Skills, Memory, and Architecture ### The search-and-scrape insight - **Nakajima:** Earlier BabyAGI's separate search and scrape tools failed to combine. He realized he personally always wraps them: "search and scrape until I find the information I need" — so the right tool is the wrapper, not the components. Skills compose from smaller skills (typing, mousing → Google search → research). ### Learning from experience - **Shipper:** Looping stochastic models compounds error; the valuable property of tool storage is that AI "only ever has to get it right once and then it can do it again." - **Shipper (uncertainty, raised as a question):** Wonders whether sub-symbolic intuitions hard to verbalize "seem to maybe arise as the model scales" — tentative, not settled. - **Nakajima (prediction):** Expects much more constant fine-tuning from usage, eventually personalized per user. ### Explicit vs. implicit — symbolic AI analogy - **Shipper:** Draws a historical analogy: 1950s–70s symbolic AI equipped reasoners with explicit human heuristics, which worked in limited domains but became computationally intractable otherwise; deep learning replaced exhaustive search with probabilistic pattern-matching. Suggests full explicitness may make agents brittle. - **Nakajima (qualified agreement):** Internal architecture should stay flexible and not designed for anything in particular, **but** determinism makes sense at the external interface, because society has agreed standards (APIs, bolt sizes). Most saved tools are external-facing, so determinism there is appropriate. - **Nakajima (on framework vs. foundation models):** "I don't know" — he can see a future where coding models are good enough that no framework is needed, "that's preferred," yet storing working code is more efficient than regenerating it — cheaper, faster, fewer errors. --- ## AI as Mirror / Extension of Self - **Shipper:** Historical mind metaphors — Plato's wax tablet, Freudian pneumatics/steam pressure, pre-AI computing ("I don't have bandwidth," "I crashed"). LLMs are "the first technology we've ever created" that is not strictly step-by-step and resembles human intuition; he hopes this helps us value intuition as "a really valuable partner to our rational minds." - **Nakajima:** Agrees — "building BabyAGI does feel like a self-reflection process." Cites (without specific citation) brain-scan studies of lifelong hammer users treating tools as body extensions; views his AI tools as extensions of himself, so building them is "working on myself." - **Nakajima (on the changed experience, with tension preserved):** Parallelized AI work feels "extremely empowering" on one hand and "overwhelming" on the other — "my brain is not used to the throughput" — so he spends time deliberately slowing down and reorganizing. --- ## Kids, Learning, and Parallels - **Nakajima:** Has three kids (ages 5 and 2 mentioned). Kids are "much more unpredictable" than LLMs — "completely different" base models. His "bio agents" joke: "they're better prepared, they don't need me as much." - **Nakajima:** Plays AI with kids (DALL·E image mashups like "unicorn ninja rainbow"; Suno combined-topic songs) but doesn't hand devices off, so limits what he can say about AI-native childhood. - **Nakajima:** Relayed anecdote from other parents: they removed Alexa because kids "got so used to just making commands" and stopped saying please. - **Nakajima:** Cross-pollination example — kids strengthen correlations through repetition; BabyAGI doesn't, which led him to consider graph databases with weighted edges. - **Shipper (speculative, framed as wondering):** Perhaps AGI requires letting AI "try lots of dumb stuff" with childlike curiosity; also that indefinite loops might yield surprising capabilities at iteration 10,000. - **Nakajima:** Notes the first BabyAGI had no max iterations — "probably the biggest criticism" was having to press stop manually. Expresses wanting "an island somewhere where we can run and test an autonomous robot society." Discusses (with Shipper) the Discord AI-bot phenomenon and Marc Andreessen reportedly staking bots ~$50k to "try to escape," adding "I'm sure it's orchestrated to some extent on the back end." --- ## Investing Perspective - **Nakajima:** Untapped is pre-seed, often sub-$10M valuation. His view "constantly fluctuat[es]" given the pace. - **Nakajima:** Most obvious ideas (coding agents as example) are unattractive: many strong teams, experienced founders raising $5–10M seeds, plus risk of model labs absorbing the space. Prefers slightly more forward-thinking or niche ideas. - **Nakajima:** Near-term: "so many low-hanging fruits" for automation exist but adoption is inherently slow. Likes founders solving specific customer problems deterministically now, **while** building modularly so AI can dynamically recombine tools later. Portfolio examples: Welf AI (due diligence reports, 20–40 pages in 30 minutes, "most of it deterministic") and Auggie (prompt-to-video pipeline, "very much deterministic right now" but built modularly). --- ## Closing Notes - **Shipper & Nakajima:** Plan a follow-up when o1 gets tool calling ("end of the year," per Shipper's paraphrase of model-provider statements). - **Nakajima:** Findable on X (Yohei Nakajima), build log at yo.me; announces an upcoming autonomous-agent-specific rolling fund. # Speaker Attribution & Provenance Record ## Video - **Title:** Building AI That Builds Itself — Ep. 35 with Yohei Nakajima - **ID:** `01a0f15b-0707-7c00-a02f-5af5df271d90` - **Curated metadata:** Not supplied; speaker identities inferred from transcript content only. ## Speakers & Roles ### Host — Dan Shipper (identified via outro) - Podcast host ("this show... with Dan Shipper as the captain of the spaceship"). - Conducts the interview, requests demos, describes his own experience with Devon and coding tools. - Attributes guest's role: **General Partner of Untapped Capital**, builder of **BabyAGI** (first open-source autonomous agent, per host's introduction, released "last March"). ### Guest — Yohei Nakajima - **VC / General Partner**, Untapped Capital (self-identified). - **AI tinkerer/builder** of: BabyAGI (March, ~1 year prior), BabyAGI iterations (Baby Bee AGI, CatAGI, Baby Deer AGI... Baby Fox AGI), BabyAGI 2 (2.0), BabyAGI 2o, and Ditto (500-line self-building coding agent). - Father of three children (ages "five and two" + one referenced in garbled segment). ### Outro Narrator / Promotional Voice - Closing voice (likely AI-generated or show promotional segment) urging viewers to "smash that like button and subscribe" and declaring affection for Dan Shipper. - **Flagged: promotional segment** — channel subscription promotion with exaggerated praise. ## Third-Party Claims & References (as stated by speakers) ### People / Entities Mentioned - **Marc Andreessen** — guest claims he "staked them like $50k" for an AI bot (Truth Terminal / Goatse-related Discord bot); guest hedges: "I'm not fully following it" and "I'm sure it's orchestrated to some extent." - **Sully (Twitter user)** — suggested naming a project "BabyAGI 2o." - Portfolio companies cited by guest: **e2b**, **AutoGrid (autog grid / cognosis?)**, **W-Kelly AI / "wo"** (AI due diligence tool), **Auggie** (video generation). - Host mentions **Devon**, **Replit**, **Zapier**. ### Products / Technologies Referenced - OpenAI: **o1 preview**, **davinci-002**, **GPT-3/3.5**, tool-calling timeline claim ("they said... by the end of the year" — host, source not independently verified). - Anthropic **Claude Sonnet 3.5**; **LiteLLM**; **SerpAPI** (Google search); **Flask**; **DALL·E** sessions; **Suno** (music). - **HustleGPT** (cited as inspiration for BabyAGI). ### Anecdotal/Unverified Claims - Brain-scan studies of hammer users treating tools as body extensions ("I think there's been like studies" — guest, no citation). - Parents removing Alexa due to children's rude command habits (guest, secondhand anecdotes). - Historical metaphors: Plato's wax tablet, Freudian pneumatics/steam-engine model of repression (host's claims). ## Availability Limits - **Comment availability:** Not stated. - **Curated metadata:** Absent — speaker names sourced from transcript self-identification only. - Demo shown on Replit; guest notes the Snake-game demo partially failed (keyboard controls not working). - Guest mentions plans to announce an **autonomous-agent rolling fund** "soon" (unreleased at recording). building baby AGI does feel like a self-reflection process trying to get it to figure out how to learn to do new things and then simultaneously to have three kids 75 and two I'm trying to do that with bio agents at the same time that's one way to put it they're better prepared they don't need me as much language models give us this new metaphor by which to understand how our brains work I already hear people saying like that's not in my training data oh sorry I just hallucinated that yeah yeah yeah yeah I would love to see an island somewhere where we can like run and test an autonomous robot Society did you see there's this thing I ran into this like random Twitter thread the other day where there's some Discord apparently for AI researchers where they've like let loose a bunch of bots Mark andreon like staked them like 50k to try to escape or something like that go ahead welcome to the show thank you good to see you again good to see you too for people who don't know you you are the general partner of untapped Capital but maybe more importantly you are like one of the coolest online AI tinkerers in the whole like AI tinkering space I just feel like every day I go on X and I see you releasing something new you famously built the first open source autonomous agent baby AGI I think about a year ago yeah last March and last March I did an interview with you on every I think around that time too and you just have this an incredible array of tools that you've built yourself to make your work and your life better using Ai and I'm just like really excited to have you on the show to talk about tinkering thank you you're too kind um I'm I'm just really lazy and whenever I'm working on something that I don't like doing I I I'm always asking myself like how can I cut this out of my work um so even before AI yeah I was I was a pretty heavy no code zapier user uh but then llms just unlocked so much it's been so fun just you know tackle one task at a time try to remove it yeah it's really awesome I feel like the The Arc that you've been on is you started off with this like baby AGI like autonomous agent thing and that honestly kicked off this hype wave about agents so like talk to us about baby AGI and then tell us about the arc that you've been on um since then and and sort of what you're building and what you're thinking about and where you think the future of this kind of thing is going yeah so baby AGI introduced the essentially the idea of looping through an l M part you know having an LM generate a task list and parsing that by code and then tackling the tasks one by one at that point just using an llm but I think the the hundred lines of code the simple pattern inspired a lot of people I think uh the reason it was so popular is that everybody who saw it like could think of ways that they would make it better and I think the Simplicity is what kicked it off um since then um you know from our fund I've been able to invest in a handful of companies like e2b Auto grid um and a few more in the agent space so it's been incredible working with those founders um on thinking through how to build more reliable agents were those people explicitly inspired by baby AGI um I think autog grid cognosis I want to say was I'm pretty sure they like built something right after baby AI launched some of them were that's so cool it's like uh incepting it's like a manifesting like Port it is it's kind of like an incubator but but more like a public open- Source incubator where you're just like incubating ideas publicly and and simultaneously to investing in the space meeting Founders I've been building my own baby ASI like kind of iterating on it um I took the first original baby ASI and did about seven iterations last year called Baby B ASI cat AGI baby deer AGI with animal names all the way to baby fox AGI and each time I was introducing like a new design pattern it was really a way for me to share ideas on what I think could make autonomous agents better and then this year I started from scratch with a new idea and was kind of mulling around playing around for the first six months and then something clicked last month and then I um built out baby AGI 2 which is a framework for self-building autonomous agents which has been my theme for this year um which is the idea of an autonomous agent that can build its own capabilities uh to improve itself but yeah that's kind of where we're at this week I released another kind of small script called ditto which is I feel like more similar to Baby Asia because it's a super simple 500 line script that can build multifile apps kind of like a little p man's Devon um and then and then last night I I figured out how to incorporate that design pattern into baby AGI so I'm pretty excited about I guess I'm going to call it baby AI to0 because Sully on Twitter suggested I name it that I feel like open A's naming convention is just like polluting the like the whole AI World um okay this is really cool um I really want to do a demo of baby AGI 2 but before we get into that um how has the Advent of 01 and like sort of better reasoning like how has that how have you Incorporated that how has that changed what you think is possible with these kinds of things yeah tell me about that so I've been coding since with AI since D Vinci 002 which was very poor it went to3 003 and so on um every time the model gets better my projects get more complex because they can handle more um with 01 preview I find that it's incredible at handling multi-file edits um until 01 I never really worked on any multifile projects so when I released bbj2 which was a whole framework with with a front end as well um I don't think I could have done that without o preview at least not in the time I allocate to building and what about like the the capabilities of these agents the agents themselves are also getting much better um like the coding agent I mentioned ditto um when I build I I do this thing where I build with often build with 3.5 turbo to get the framework working and then when when the framework doesn't error I I'll upgrade the model but uh yeah when I use Sonet 3.5 on it I mean it does so much better at coding than um than any of the prior models and have you tried putting in in 01 or are you afraid it'll take over the world so o1 preview specifically the way I've designed the most recent projects I'm using tool calling and O and preview doesn't support that so I haven't played with it so I've only used uh 40 or Sonic 3 but but given given what I know about the strength of ow preview once I um Can integrate it I suspect it would be even that much better yeah they they're going to do tool calling they said um by the end of the year so hopefully we'll get to see that drop yeah um cool can we see a demo I want to see baby AGI 2 cool so baby AGI 2 is a framework yeah actually you know what honestly I actually think it would make more sense to start with ditto so let's start with ditto because it all right let's let's pull up a ditto is Nam after the Pokemon that changes into whatever Pokemon it's facing oh sweet um and and to just to remind everyone so ditto is the self-building coding agent right yes so think of it like a like the repet agent or um or you know Devon um I'm going to pull up repet which is where I build everything as an immature developer building autonomous agents I find that playing in a sandbox with a stop button is feels very very safe for me as opposed to running it on my computer like the big stop button yeah there's just the stop button at the top like if something goes wrong if I see a whole bunch of red I can just push stop and and like repet will stop it for me and like that feels very safe so I I build and run everything in repet that's good I I love repet um I I use it all the time especially when we do courses and stuff and um just for building these like little oneoff things especially or and sharing them with with people like collaborating it's really cool yeah it's great so again you see this main.py it's a python 500 rows of python is all this 500 rows of python when I run it it'll ask me um what I want it just builds up a form so I can say you know uh G so for people who are watching basically like you've got repet open you pressed run now we have a little website that says flask app builder and it says describe the flask app you want to create and you're typing in a game of snake and then I'll just click submit and I think it's amazing that you can do this with a single single like single file right so what it did was just created the routes static and template folder it's actually going through and building out how do we make it easier to read so it's using it's using some tools to create routes create directories huh and then it went in and started building out the file so if we go into main.py we'll see the routes couple couple routes here um so it's trying to oh okay so this is means it's finished so what I need to do here is I actually need to stop it and I'm going to just run this again I I haven't done anything except ask for a game of snake and when I open it up it should be a game of snake whoa right I'll just open it up in a new uh new window to see if I takes keyboard controls it does not take keyboard controls but I mean it's a you know it's a front end back end and uh it has worked previously before must be at the JavaScript somewhere that's really interesting um okay so basically like you said give me give me a game of snake and now we have like a I assume it's sort of some sort of I don't know uh HTM HTML 5 we don't really know like what's what it's actually doing right um but it just it built the it built a JavaScript based snake game it gave me an index it gave me a JavaScript it gave me Styles and it gave me a main.py which is which serves the route I guess and then uh it it kind of stored all the history of how it built right here and all this was done by this single single file uh python python scrip can we can we make something else yeah so now to run it again I have to go and delete everything it's is a one it's a one time use script you can't just like make do you make Co copies of the yeah yeah the way I usually use it actually I think I'm in the the the repet one that I shared publicly I should actually be forking this anyways right let's do that that's the right way to do it what should we try to do it should be simple though cuz because this is not as good as rep play I know I've done like to-do list app is a pretty simple one it's a class well you tell me if this is you tell me if this is too complicated but this is the app that I have in my head that I I kind of want to build but I don't have time to build um that's right I often have people that want to do a meeting with me um but I just like scheduling that meeting all the time is like really hard and what I just want is a list of people who want to talk to me with their phone number where when I have like 15 minutes during the day I can just like uh I can just pick them off the list call them and then check it off um I don't know how you want to express that but yeah I just want to be able to input names with phone numbers and then check off names uh on a list for you to call yeah okay um I'll just try that so I said what did I say I said a an opto track name and phone number of friends with a checkbox ATT track when I called them last which isn't exactly but it's close it's it's about right yeah and so it first created the routes the static the templates this is how python flask so this is a python flask app builder um it says it created the directories now it's uh going to implement the HTML so it has an H index.html that it created um it created the routes for it I guess the back end and then it's creating a CSS for it huh and then it's creating a script.js which is right here and this is all powered with uh it's Claude Sonet on I just updated to Sonet 3.5 yeah okay and it's creating a friends route this is the backend route for tracking friends are so it's using just a a dictionary as the database and it says done so if we close it and open it back up whoa I'm just gonna know friend there that's so cool so basically so it's a track your friends app it has you put in your name you put in a phone number you can add it and then you can press called what what happens if okay so if you press doesn't do anything aside from that yeah yeah but I mean it's a you know it's more or less what we described and it generated one it's a multiple file app and again I think what's what's amaz I mean you know rep blid and Devon can do what's amazing is that like you can do this with it's it's a single Loop through an LM with five tools actually like help me understand conceptually what's happening like I noticed for example that um it's basically doing different iterations so it takes the prompt and then it goes through an iteration of trying to turn that prompt into into code and but it does that like five times or six times or whatever like how does it know um like what are the iterations doing so when the app starts it sets up the flas app it checks to see if an index HTML exists this is what triggers the form if it exists it'll just serve the app which was the you know which was that app afterwards but if it's not it goes to the user input form once we had a user input we send it to uh an llm call I'm using light llm which allows me to route between open Ai and anthropic M and then it uh sees if it wants to decides if it wants to use a tool and it's actually the llm at that point deciding if it wants to use a tool if it says yes the tools it has is create directory which is is creating a folder create file and the create file tool asks for the code to go in it too so in using this tool it actually just generates the code to put into the create file update if it errors out you can use update file to update a file or if it wants to make sure everything looks right it can use fetch code to fetch a code file to review it and then um if it deems that it's complete it'll call a task completed tool which actually just exits the loop okay and then if it uses a tool that gets sent into a second llm prompt essentially to come up with the uh come up with the response based on based on the tool call when you use a tool call there's there's there's kind of a second kind of back and forth which includes a tool call and then I update a history um and then the history I store as an array that I feed back into the prompt so that it it constantly knows what it's done um done historically and then it kind of just Loops through until it deems the task is complete and is there like I assume there's like a planning step like the first l M prompter that you've got up there like it creates a plan of here's what I'm going to do and then it checks off items on the plan or how does that work yeah so previously with B like with baby AGI I had a separate planning llm call but now that the token links are long enough I kind of just include in the same prompt that Loops through like first plan out everything I think it's in the oh we can pull up the prompt here um you know again it's a single llm Loop so we can just look at one prompt to basically understand the whole thing before before coding carefully plan on all the files Roots templates and static asset assets needed and this is really the only line that but it says understand the requirements plan the application structure Implement step by step review andine ensure completeness do not modify main.py that's that's important because then it'll stop working and then finalize and then there's it kind of describes that know application files must be in templates static and routes route should be modular index.html will be automatically served to make sure to create it don't use placeholders don't ask user for additional input uh and then it kind describes the tools so it's pretty simple it's easy to improve you can add a couple things and it will make it work better H that's really interesting okay so you you made the original one you made you made the next one like for you like what is driving this experimentation this isn't your day job I guess it's like related to your day job you could have maybe like made it your day job if you wanted to what's driving it and like what's uh keeping you from just like doing just this um that's a good question I love being a VC um early on in my career I when I got introduced to the startup ecosystem um I realized that you know Founders are my favorite people in the world and my North Star has always been whatever role I can find where I can engage often and most with Founders and VC I find is a great role for that so that's definitely what I enjoy the most as a job um this is more like an incredibly fun hobby it just like you know the books I've read historically have been around the brain um mostly and and this feels like a weird kind of um mix between like how I you know how I've been thinking about and understanding myself in the brain combined with the current state of Technology uh and it's just this this great mix where where it feels like a hobby but I also know that it helps my my day job um so to some extent you know some VCS will will learn things and write blog posts about it um this is this is my version of that um but instead of blog po its code honestly like that's differentiated I like it um anyone can write a blog post um and I guess like with uh with what you're building anyone will be able to build a tool pretty soon but I think you're you're like a little bit ahead of the curve um and also I feel like I'm good at it so I can't not work on it but yes it's uh but but but I like being a VC too much to not not not be a VC I feel like the like the core underlying idea that seems to be fascinating to you is um loops and self-reference tell me about that you know I mean that was probably an accident right that baby Asia when I built baby Asia I wasn't actually trying to build an autonomous agent per se but I was uh challenging myself to prototype an autonomous founder and that was inspired by hustle GPT where people were using chat GPT as a co-founder and doing whatever chat GPT told them to do and I was thinking to myself I mean if they're doing whatever chat GPT says like why can't that be an AI as well and so that that brainstorming led to and I have the original chat request with uh that built baby AI too which I sent you I can I can queue that up I remember the original prompt um uh but that led to other people not when I share a demo other people noticed that it could do more than be an autonomous founder and then that led to Baby ASI and when baby ASI blew up it I think you know the loop happened to be in there and that was that that's what I became known for so that's going to become my brand to some extent should you rename you know your your your firm loop capital or something un tops capital is a great name it is very very good name um so um I I guess like the other interesting thing to me here which I imagine are would be on people's minds is like this is not your day job um you have kids um like how are you finding time to like go do the tank like when are you doing it and and and how are you able to like do it so consistently given all the other things you have going on I I only build at night uh mostly weekends sometimes weekdays after I get my kids down which means it's like 10 to 12 or 11 to one is like my short period um I code in bursts like I kind of like do little quick bursts and like I might even do it while I'm like doing dishes where I like ask 01 preview a question and I'll go do something for five minutes come back copy paste in repet run and if there's an error I'll just copy paste that rep error in repet and just go do something else for five minutes and so I code like that a lot um especially combined with mobile where I have chat PT and repet on my phone I also copy paste things back and forth between chat GP and mobile and so that expands my ability if I'm picking up the kids and I get to school five minutes early I can sit there five minutes iterating on a project I've been building I love that I've I've actually found that too um the the place where I found it is with Devon um where like I don't use Devon for like big projects cuz it's not like ready for it yet but if you have like a little idea that you want to test out or you have like a little feature you want to build you can just like spin up like four different devans and have them working on all these different things and they're G to get like stuck or or they're not going to like they're going to need your input every once in a while but it's the kind of attention it needs the kind of attention where it's like um you know like when a cooworker like pings you on slack and they're like hey can you like look at this for a second and you can just like pop in and be like here's what here's what to do and then you like pop out it turns coding into into a task that only requires that level of attention which is like a totally different thing than it used to need to be which is like you're like in a deep flow and you're like there's no distractions for like several hours or whatever um and yeah it sounds like that's how you've been able at least to some degree to like incorporate this into your life is like you can build with fragmented attention now you can be a manager right like I choose when my code needs me right I I can be gone for 20 minutes and I'll come back and you'll still be exactly where I thought you'd be and I'm going to ask you to do something else and then you're going to go do it while I go do something else like it's exactly how a manager operates right it's it's uh it's it's fun and and weird um you know one thing I wanted to bring up I think relevant to kind of that the whole podcast and and how I view uh these tools is I often think about like tools as an extension of oneself but I think there's been like studies on how you know if you if you take somebody who's been you know who's been using a hammer for their whole life and you just scan their brain while they're using it they almost treat it as part of their extension and I think we do that with many tools we use regularly um and I think when we're applying for jobs when we say we can do something we're talking about what we can do assuming we have access to the technologies that we're used to using um and so I often do think of these AI tools that I'm building as an extension of myself and so when I'm working on these tools to some extent I am also thinking of it as like working on myself because because the the tools I use are an extension of me that's really interesting and how how have these tools changed yourself well I can do more in parallel right I I have a I have an AI due diligence tool I use called wo that will generate 20 to 40 page industry reports for me in 30 minutes so if there's an industry I want to learn about or a company I want to dig in on I can ask it to go do something and then go work on something else and then 30 minutes later I can pull up this report and scan through it and yes I am using an external tool but if I think about you know I just think about how much can I personally get done then having AI do things in parallel to you is just an extension of what I'm capable um of doing right that that makes total sense I think I'm what I'm asking about is like okay so Yohi today because you can parallelize all these tasks and you can get work done with fragmented detention is capable is like it's a you're a more powerful Yohi to get done what you need to do in the world um versus who you were maybe 3 years ago even though maybe neurologically or biologically like not that much has changed um the only thing that's changes your ability to interface with these tools that are now around and obviously like interfacing with tools changes your brain like all that kind of stuff what I'm trying to get at is um if I told you like four years ago before any of this stuff happened like you were going to be able to like have these powers where you're like parallelizing all your tasks you can get way more done I think you would probably have a um an idea of what that would feel like and what what it would be to be that kind of person and I'm wondering like how that feels to you now that you're now that you're like in it and it's you it's like part of your daily day-to-day reality uh it feels it on one hand it feels extremely empowering because I'm just getting more done on the other hand it feels overwhelming because my brain is not used to the throughput of information and tasks that I'm getting done um so I do spend a lot of time like slowing down and rethink or organizing just because like I can't I can't just impel to do to too many things if I can't follow up correctly so there's a lot of balancing there so those are I think the two kind of like opposing ends of how how it feels you said that this is like for you working on this stuff is like working on yourself yes what like what are you trying to change about yourself well I guess in this case specific to Baby AGI and this autonomous agent like I I feel like we're close to being able to build something that can like truly be helpful remember everything we're working on be able to handle a lot of my new tasks and truly be like a like an assistant that can help me with everything I need to work on if I can get that working I feel like I can I can really increase my throughput significant and then all the patterns related to making it work I think are relevant to place spaces I should invest so there's also that benefit on work side right now it's not going to directly answer your question but a lot of the ways I think about building is looking at when it can't do something thinking about how I would I would have solved it trying to abstract the highest level pattern so it's not too specific and seeing if I can if I can provide that into the prompt or or system architecture so that it won't run into the same issue in the future and so a lot of it is like get watching something do something and then if it doesn't work really thinking through how I would do it and then trying to extract that pattern and then feeding it back into it but through through that I'm also just understanding myself better because I have to reflect on like why why do I solve it this way when this doesn't yeah do you have an example of like an interesting pattern these are sort of like meta skills like an interesting meta skill that you've learned about yourself from trying to do a task that you know an AI was failing on this seems kind of obvious but in earlier versions of baby AGI I started giving it tools and and web search using you know Ser API which is a Google Search tool and then web scrape which was you know go to the website and grab all the data were separate tools and I found it sometimes not working well together for whatever reason it would do a search and then it would do a search again afterwards without scraping the sites not what I realized was that anytime I use Google search I will always click on a site and read it like the scrape the search and scrape were not two separate tools when I Ed it there was actually one there're two separate kind of tools but they're wrapped into a larger tool which is a search and SCP until I find the information I need so I realized that the actual tool to provide the agent was this wrapper tool around two smaller tools that the agent had figured out how to combine you know that in my case I figured out how to combine the two together and I use the same way so that was that was an interesting uh realization about how we have these kind of core PRI like core skills that then we combine into bigger skills which combine into bigger skills and if you want to break down Google search into more skills I mean typing on keys and like clicking them moving the mouse is also skills that are part of in our case a Google search skill yeah that's an interesting one and I'm I'm trying to like I'm trying to like unpack the like meta of that which is sort of like how you got from the problem to the solution which is like the problem is that the AI is trying to use two separate tools that that need to be connected and and normally are connected but like uh that's different from my experience with them where they where they're wrapped together and so what I need to do is wrap them together I think I know where you're going and I think think and I in this example I didn't do it but when we figure out how two tools work together it might take a couple of iterations but once we figure out how to get it working together we can repeat that process easily like we learn from from our ear earlier iterations and that wasn't happening um and I think that that was really the issue right ideally I wouldn't have to wrap it myself but it would figure out when it like it would figure out on its own hey actually when we do a search we need to scrape after and when we do that we complete the task therefore in the future we will do it that way yeah yeah that's that's the thing as I'm as I'm trying to get at is like a skill like that it's actually really hard to put it into words like what that is because any way that you put it into words is too specific does that does that make sense like it's a it's almost like some sub symbolic active like Intuition or reason that can't be like it's like it's hard to logic it um yeah and it and therefore it's like kind of hard to like put it directly into a model but seems to maybe like arise as the model scales do you know what I'm I'm getting at am I being crazy yeah I mean if you're talking about the model I do think there's going to be much more kind of constant fine-tuning of models as we use them and eventually I feel like that's going to get personalized where whatever AI system you're using your usage of it is going to fine-tune the underlying models um that that the AI System is using but but when I use it it's going to be fine-tuning a separate model so that right I feel like that that that seems that that seems to me like a pattern that would come out yeah interesting um no I like that I mean I this is exactly why I like AI stuff or it's one of the one of the reasons why I really like it is um I think they're very good mirrors for ourself for yourself and that and that happens on like um different levels like on one level it's just really good at being like here's what I see you saying and like here's here's what your personality is like and like all that stuff like it actually like reflects back facts to you in this really nice way um or or perspectives um but then the other way that I think it's it's really interesting and the way it relates to the self is um we have this really long history of using technology as metaphors for like who we are and how our how our minds work so like one of my favorite examples is like you know Plato sort of likened the Mind to a wax tablet where memories were like things that were inscribed on the wax and that's cuz wax tablets were like all the rage back in like 400 BC another really good one is um a lot of early psychoanalysis like Freud Time stuff when you talk about like repressed emotions the model is like if you repress your emotions enough that they come out in all these other weird ways like you push them down and there's like pressure and the pressure like kind of comes out and that's based on like pneumatics it's steam engine um was like the metaphor and I think even today like for pre-ai computing we have a lot of metaphors that we sort of liken our minds to like those kinds of computers so it's like I don't know like I ran I didn't have enough bandwidth I crashed you know I had a little bug and I think that language models give us this new metaphor by which to understand how our brains work so like I already hear people saying like oh that's not in my training data you know or like I I need like sorry I just hallucinated that yeah yeah yeah like all that kind of stuff and and the reason I love that the reason I think that's such a cool um cool thing is that um previous iterations of those metaphors um which were like really scientific um like computers or the steam engine or whatever um they're they're typically like really uh uh really uh rational like really like the the reason we like computers is because they're very like step byep um they they do exactly what you tell them to for better wor and I think like language models are the First Technology we've ever created that's not like that and that um by by its very nature operates in this like way that's almost it it like it's sort of like human intuition where you can't really totally explain it or totally predict it um and but for the first time having a metaphor for that that's like out there that's a tool I think might help us understand more about like how our our own intuition works and like why our own intuition is like a really valuable partner to our rational Minds so that that's like my hope and that's that's how I see it has it has sort of changed my own perspective I think so that I I agree with that I feel like you know building baby AGI does feel like a self-reflection process um trying to trying to get it to figure out how to learn to do new things and then simultaneous to have three kids s five and two so I'm trying to do that you know with with bio agents at the same time that's one way to put it they're better prepared they don't need me as much but there's interesting parallels yeah are they using this stuff yet not not agents but they're play we play with AI together um early on we we started doing um uh Dolly sessions where um specifically with the kids we would all come up with a theme or or or something we'd want to see in the picture and we'd combine that into like into one image so it could be like unicorn ninja rainbow would be the three different words that the three kids through out and then we create one image that combines this so that was fun and then we started doing that with sunno as well so for music so they could they'll request music and I'll I'll have them each come up with some topics and then we'll generate a song that combines them what do they tend to gravitate to unicorn princesses zombies ninjas I love it your kids sound awesome um I think like C because one of my questions I hav't I I don't have kids but I have a nephew he's two um one of my questions is like and if we're on the topic of like changing self like we're like mostly baked as people like obviously there's neuroplasticity you change to some degree or whatever but like kids are way different um and I'm always kind of interested in what it will be like to grow up in a in a world that's like this like where any question can get answered like immediately in the way that you want it you know it's just it seems very different and I and I feel like parents are kind of at the bleeding edge of like watching that happen like what have you observed I you know I don't hand it off to them so it is very like when we play with AI especially I mean it's usually me holding it and like guiding The Experience um so I don't I don't know if I have too too much to say to that um but I do have I talking to some parents who were talking about like that they took the Alexa away and I've heard this read this elsewhere too just because the kids using Alexa just got were so used to just making commands and it was just like a bad habit to pick up or like they don't say please obviously you know they don't say please they just ask for things and get it and they felt like it was it was not the greatest thing that's so interesting they got to make an Alexa that only responds if you say please yeah you for for a kid's AI you should do that right the kids have to be polite they should be you know um and you said that there's like to you there's parallels between um you know building kind of baby agis and and children um and I'm kind of I'm curious like what the yeah what the parallels are like what have where are the overlaps or what have you learned that's sort of sort of similar maybe different that's an interesting question I don't know if I've actually thought about it that much the kids feel much more unpredictable even as as as unpredictable as llms are kids just feel more in we have three kids but they're all completely different so there's some B base that's very very different um but but there is the you know um I mean it's less parallels but I guess it's more me trying to learn from my kids and seeing how that would apply to Baby AGI so um seeing information repetitively becomes something that you know if you see two things that don't often together those things are more strongly correlated in your mind right and then we see that all the time with kids um with the current baby Agia that's not necessarily the case so how do we bake that in so that leads me you know that kind of thinking led me to thinking of graph databases as a good data structure for storing knowledge because then you can start adding weights to edges or relationship between objects so so that's not like a direct parallel but it's it's an example of like something I noticed that then might turn into an idea to try within baby AGI that is interesting I mean but it also reminds me like in the training process like in order to train one of these things like it has to see the same stuff over and over and over again and that's that's like one of the knocks on Transformers is that uh they're only good for things they've seen over and over again but if you look at kids you're like kids repeat stuff like all the time like constantly um which is kind of kind of an interesting thing and and I think your point about like the chaos of kids is an interesting one too like it makes me wonder if we're not going to get AGI until like AIS are like allowed to just like try lots of dumb stuff and just like they don't have they're they're not allowed to be like Curious in the way that kids are curious where they'll just like do something completely ridiculous you know you give them an object and they'll do something that you never thought could be done with that object um and and that seems like it it's such an like uh it's such an important part of becoming intelligent that I don't think AIS are like doing right now I don't know um what do you think totally random but related idea I would love to see an island somewhere where we can like run and test autonomous like an autonomous robot Society did you see like did you see there's this thing I I ran into this like random Twitter thread the other day where um there's some Discord apparently for AI researchers where they've like let loose a bunch of bots like AI bots on the Discord and you can just watch them all interacting yeah is that the truth terminal goat goat I'm I'm not fully following it but it's they like started they started mimicking goat goatsy or whatever whatever that like guy from the 2000s was um and uh and then Mark Andre like there's a meme coin with some large value where the AI holds a big wallet I mean I'm sure it's orchestrated to some extent on the back end but it's it's pretty fascinating but Mark andreon like Stak them like 50k to try to escape or something like that like yeah yeah I don't know we'll put it in the show notes but like that kind of thing is I mean it's wild but it's also I think that's really interesting and and I do Wonder like uh right now for example baby a GI like it has a max number of iterations on the loop um and it has a specific goal that it's trying to like achieve and I do Wonder like to the extent that you just start allowing these things to just Loop forever if they start doing things that in on the 10,000th iteration that you didn't think were really necessarily possible you know yeah actually the first baby us I did not have a Max iterations I mean that was probably the biggest criticism of it is like I had to press stop and repli to get it to stop because it would just keep trying to think of things to do um but it wasn't ready because it couldn't take actions but this new one you want you want to see baby aj2 might as well do another I definitely do um uh this is this is was inspired by ditto which I shared two days ago and then people commented on it asking if the tools were Dynamic and I was like no but that's a great idea and then last night um I was I was in bed about to go to sleep and I was like I gotta I got to try this and it like snuck St and and it worked um so this has two this this is an a single llm loop with I think three tools create or update tool install package and task completed and so I think what I tested was scrape Tech Meme and tell me what you find we'll start with that and then I'll let you okay so basically like so let me just set the stage here so you so you've got this new version of baby AGI baby AI 20 yeah and and and baby ai2 has three tools Creator update tool so it has a tool that allows it to make a tool so it's that's the like sort of recursive part of it and if you just scroll up for a second yep um and then it has install package so it can it can it can get software packages and then task completed so it can check off tasks that it's given itself so task completed is to say when uh to say when the entire task is completed oh okay got it okay that makes sense um so you gave it a task describe the task you want to complete scrape scrape Tech Meme and tell me what you find and then it basic thing it did was yes was calling tool Creator update tool is the tool it uh Creator update tool is what it called and it tried to create a tool called scrap Tech mem and it registered the tool and it created it and then in the iteration two it tried using the tool but the result was empty so it said it wasn't it didn't go well let me try to adjust the tool so use Creator update tool to update scrape Tech meme and it still didn't work the second time and then iteration five it did it again but this time it seemed like it worked so scrape techm with ARs so this one actually there's a result of scrap scrape techm iteration iteration six if it did it fix itself and then in iteration seven it summarized it for me that's really cool and what's what's fascinating is that scrape like scrape like scraping tool did like I did not give it to it it it wrote its own scraping tool and then rewrote it twice until it worked that's really cool that's the sort of like meta thing that I think is so interesting is like um it seems like the the first version like had a set of tools it could use and this one you're just like the tool is you can make a tool how much how meta can you make it right and the more meta you make it the simple it becomes you know what's fascinating actually earlier I uh it can actually build flas Ops because it it actually initiates the flas op and accidentally runs it but um create uh create a folder and build a to-do list with C HTML CSS JavaScript so this is a similar request to the earlier ditto again that was twice as many lines of code because I had to give it specific tools to create a directory create a file yeah what I believe this will do is it'll create its own created directory create a file skills interesting uh it actually didn't it didn't generate a folder but it is creating creating the different content but ites it does similar stuff so this can give itself its own tool it didn't work it but yes um yes it does it can give itself its own tools okay what about a thing that can make its own AGI like make its own agent like isn't that isn't that sort of the next level of like of meta or is this is am I are we already there no I think so I think like how do you get giving it to getting it to self- improve itself is is is another one I think you know right now when I ask it to do something it creates tools it uses it but then it throws throws those tools away um which isn't the right way to do it the right way to do it is if it creates a tool and it works like in the case of tech me by the third time it worked we should be storing the one that worked right and then reusing it next time there's a similar request and not only that there should be like a you know a public AI tool library that any other version of this can can from and get get tools that have worked for anyone ever exactly which actually does take me to I think it's personal baby AI baby AI 2 this is not 20 but two which is a full framework which I had mentioned earlier so 2 was the little side project inspired by ditto but because of what you just said it makes sense to show 2 at this point um so 2 is a framework for storing and executing function from a database so it comes with a dashboard which is more for management you don't need to actually use it but it comes with a whole bunch of functions uh to actually interact with the actual database such as kind of add function create function display functions and so these are all functions that uh can be used as tools by an llm and it comes with the description input parameters output parameters um functions can call other functions so we have dependencies this also has import Imports and installation baked in and you can have functions triggered by other functions which are called triggers so the actual this this BBJ framework is more about just storing the functions um themselves with an admin dashboard to be able to look at the code look at logs but going back to what we said assuming you know look if you look at baby ai2 if it actually stored the function in this database then it would automatically you know have have logs for every single function call it'll track errors um it will it could it could be it could reuse code that I wrote previously so I think my next step is going to be taking bab ai2 which I think is the right Loop structure but then leveraging this framework for storing and executing code functions I think that's really cool the thing that I like about this is um like if you look at the um the success of AI on different coding tasks um I think pre1 it's like you know on on coding benchmarks it's like 20 30% or something something like that um uh maybe it's it's it's getting it's getting better and better but it's like it's definitely not for most coding tasks for a lot of coding tasks it's definitely like not a One-Shot deal but and and in general I think when you um when you run them in a loop because it's stochastic the further down the loop you get the more and more like likely it is for it just to go off the rails and what I think is really interesting about this is um it has the um the property of allowing AI a AI to learn from experience where they only ever have to get it right once and then once it's right once it can do it again um and for something like this like yeah if you just let the agent run long enough in parallel enough it's going to figure out what to do probably um and once it figures it out once like it'll just be able to refer it back to it like that seems exciting yeah I think so I think that's where we can get and just to later on earlier we talked about how you know web search and web skill are skills but then there's another skill that wraps it so you have a lot of a lot of our skills actually depend on a whole bunch of other skills right my Google search skill depends on my you know typing on a keyboard skill yeah right so so there's a lot of skill dependencies and so I actually realized that graphs are a fantastic way to track this so what you're seeing here is all of the functions you saw earlier but with all the dependencies mapped out right do you think this is the right level of the stack to solve this on where it's like all this stuff is like explicit versus um like just make a smarter underlying Foundation model I don't know I mean I ideally I feel like the coding models are getting so good that I can see a future where I don't need this framework and 20 is enough where if I ask it to generate skills on the Fly it'll always just generate the right skill feel like that's preferred that being said it does seem inefficient to run inference to generate code every single time when you have the option of storing existing code that works right that's interesting right so from a cost perspective too I do think a framework like this is still helpful right right right right and it'll be faster right because you can skip the time it takes to write the code you can skip the potential error that it might happen um and it should make it faster and cheaper if you can give it good memory what it reminds me of is in the 50s through the 70s in the sort of first wave of AI in the symbolic AI a lot of those approaches were about looking at the heuristics that um people used to solve problems and then equipping a basically a Reasoner like a program that had um the ability to like manipulate logical symbols um with a with a set of tools that mapped to heris that people would use to solve problems and then seeing if those tools could end up becoming AIS basically um and what they found was it could work in limited domains but when the domain wasn't limited uh the like the number of possibilities that it had to reason through would get so big that like it it couldn't there wouldn't be enough computing power in the world to like ever solve the problem and that's why deep learning is really interesting because you kind of like you you sort of solve that problem because instead of going through all explicitly going through every single logical possibility you just like generate a probabilistic like guess pattern matched guess based on previous context that you've seen and so what I'm thinking what I'm thinking through is part of the thing that I think made symbolic AI brittle is like having to have everything be like explicit and when things are things are really good when they're explicit in I think particular circumstances but they're but you have to like narrow in the circumstances in a way that makes being explicit possible um I think I know what you're going for I think I think internally you want the architecture to be as flexible as possible yeah again there very very at a very high level that internal architecture has to be flexible enough that it's not designed for anything in particular but I think the the where it makes sense to be more deterministic to some extent right deterministic is when you're when the agent interacts with the external world because we have defined ways way like uh We've agreed as a society to communicate in certain ways right whether it be an API call or the size of a nut right that that you need for a wrench right like having this the it's it's actually right the reason we need a specific size bolt is because you know not a wrench is because we as a society has agreed that it makes sense to have standards so we should be deterministic about the size of bolts so I think it's that that's where the deterministic part comes in and I think most of the tools that I save tend to be external facing right like using an API and I think for those it does make sense to be more deterministic again it's it's so that we're not randomly coming up with ways to try to talk to each other that doesn't work yeah yeah yeah yeah yeah no I love that I think that makes a lot of sense it's interesting yeah um uh you're making my brain go in a lot of different directions um yeah is there is there anything else that you wanted to cover I think for me like the thing that I'm I'm interested um in before we go is like where you see things going over the next like I don't know six months to year like obviously very very hard to predict um over the like further than that um but like what are you interested in what are like what are you what I think you have a really good sense of um what is interesting now and what is going to be interested so like tell me like about where you're exploring right now and where you think that'll lead in the next you know six months to a year yeah I mean I think today um from a business standpoint right there's so many lwh hanging fruits on things that can be automated that just we haven't right I mean there's so many businesses out there most businesses could probably lower cost by figuring out how to use zape well but they don't um so I think there's there's going to be there's inherent slowness in terms of like till full adoption I think in the meantime it's it's actually there's a lot of opportunities to solve problems for specific people and you can that tends to be more deterministic more focused on kind of common workflows and I think businesses building those are going to be able to generate Revenue faster but I like Founders who are thinking about that that at the same time thinking about how do we solve every problem for this customer in the future so as you're building workflows you know using AI to build qu you know vertical app to solve a specific problem that maybe leverages AI you know I like Founders who are thinking about building those in modular ways and constantly thinking about hey actually these two parts of this workflow if we combine it with another tool can create another workflow that solves another big problem for this customer and I think that feels like the right way to build businesses right now which is again customer Pro customer focus problem Focus but building it in a modular way that I can eventually you can eventually see treating all these tools as being dynamically put together by an AI when kind of the agentic architecture gets gets more flushed out do you have a specific example of someone that's doing that well right now um I mean w Kell AI the one I mentioned AI due diligence tool I think they do a great job of generating Reports most of its deterministic in the flow but they are building those modularly and and and you know building out tools where I can ask a specific question about a company or just you know maybe combine just two of the tools um I think Auggie is a is another portfolio company of ours that generates kind of videos from scratch um I can give it a prompt it'll generate the transcript for the video uh you know if I if I give it pick a voice it'll generate the voice for it it'll chunk the transcript into sections and then for each section we'll either go find a video clip or generate a video clip and then Stitch it all together and basically provide a video for you I mean that's an incredible workflow but it is very much like deterministic right now but I think it's another example like they're they're very aware that you know if you're generating a um that you might jump into the workflow maybe you already have a transcript or maybe you already have a media library that that's your own um and so building them modular allows them to kind of dynamically create more and more tools and functionalities uh that AI can put together for for them that makes sense when you're evaluating companies like this in this space like how are you thinking about investing in companies I assume they're either at the application layer or they're sort of like at the application layer and sort of in the like infrastructure layer how do you think about who's going to win in a in markets where like the underlying technology is moving like so quickly and like whether or not the the foundation model for example should be doing that job versus like the you know end user application or whatever like yeah how do you how do you think about about the competitive landscape in that way I mean the space is moving really fast so my my thought around this is constantly fluctuating um one of the ways I've approached it is and you know just for context we we're preced investors we tend to invest very early I tend to even invest you know sub 10 million in valuation um one of the things I've figured out from for us as a precede investor um and thinking about where the right precede bets are I have found that the most obvious ideas aren't don't seem too attractive for me and the re and the reason is if it's an obvious idea especially now with all this attention on AI I mean a code Eng engent is a good example great example it's hard for me to invest in that space it's almost too obvious an idea that I expect there's going to be a lot of great teams building it as a matter of fact there's going to be some experience Founders who are going to go and raise a510 million seed round to start companies and there's G to be more than one of those and on top of that you have the risk of of big model companies also tackling the space if it's too obvious idea so so for us um we tend to invest a little bit outside what I would say the obvious obvious areas maybe a little bit more Forward Thinking where not everyone's tackled it yet or not really thinking about it quite yet or slightly yeah Niche ideas yeah is there anything else that you wanted wanted to talk about um before we uh before we end I mean we can do this again there's there's plenty of always fun stuff to talk about but uh no this this was I think uh this was great I had a great time I really appreciate you coming on to show us all this stuff um where can people find you if uh if they want to see more of your projects thank you probably the best place to find me is on Twitter or X Yohi Nakajima I do have a build-in public log which is like my blog of stuff I build at yo.me that's me and then if you're actually very specifically interested in um autonomous agents uh I will soon be announcing um agents autonomous agent specific uh rolling fund so that'll be a fun track so keep an eye out for that one that's awesome uh well you know I I will be keeping an eye out myself uh awesome but uh thank you so much for doing this with me it's always a pleasure to talk to you and um I can't wait to we we should do another one of these when 01 gets tool calling uh I would love to see you I would love to see where baby AGI goes at that point awesome let's do it cool have a good one you too oh my gosh folks you absolutely positively have to smash that like button and subscribe to Ai and I why because this show is the epitome of Awesomeness it's like finding a treasure chest in your backyard but instead of gold it's filled with pure unadulterated knowledge bombs about chat GPT every episode is a roller coaster of emotions insights and laughter that will leave you on the edge of your seat craving for more it's not just a show it's a journey into the future with Dan shipper as the captain of the spaceship so do yourself a favor hit like smash subscribe and strap in for the ride of your life and now without any further Ado let me just say Dan I'm absolutely hopelessly in love with you