Title: He Spent His Life Savings on AI API Calls
Source: https://youtu.be/Ukxt2GrpSwE
Generated: 2026-04-27T07:20:03.850Z
Note: Speaker labels are heuristic and may require edits.
[00:00:00] A: and i was like wow like it's it's better at python than me it's better at sql than me and i'm pretty proud of my analytical skills but then i was like i'm never gonna do that again i think the ceiling now is it's no longer like can it build this web app yeah it can build a web app like easy it's more like does it start operating more like an organization all i do is talk to ai but now i'm like i'm managing 45 different agents and then it's like i'm the bottleneck right so then you remove yourself and you allow the agents to manage sub agents and then all of a sudden the ceiling for orchestration goes up
[00:00:37] B: This is the Village Global Podcast,
[00:00:39] B: the biggest topics in tech explained by the best voices in our network.
[00:00:44] C: So Parth, welcome to the show.
[00:00:46] A: Thanks for having me, man.
[00:00:47] C: People are always surprised when I tell them that Reed has an AI guy.
[00:00:50] C: They're like, what the hell does that mean?
[00:00:52] C: And specifically,
[00:00:54] C: you've actually built out Reed AI,
[00:00:57] C: which is his digital twin.
[00:01:00] C: You can actually send out instead of him.
[00:01:02] C: It pretty much replaces him for interviews.
[00:01:03] A: Yeah, it's getting there.
[00:01:04] A: That's right.
[00:01:05] A: Yeah.
[00:01:05] C: The amazing thing about it, though, is that you don't have a software engineering background and there's not a software engineering team behind it.
[00:01:10] C: Like you've
[00:01:11] A: No, ReDial
[00:01:11] C: actually,
[00:01:11] A: is vibe coded.
[00:01:12] A: It's a vibe coded project.
[00:01:13] A: One of my most difficult vibe coded projects for sure.
[00:01:16] C: I mean, it's an ongoing living project. You've translated languages.
[00:01:19] C: Can you share a little bit of just the background of how Read AI came to be?
[00:01:23] A: I want to say November of 2023.
[00:01:26] A: maybe 24 maybe 24 basically when custom gpts came out from open ai um one of reed's teammates ben rellis he me and him were working out of a beach house in la and he comes over to me and he was like he knew i was the guy that could build chatbots um and these were no code chatbots just like a prompt and a set of documents and telling it to orchestrate those the like telling it to retrieve the right document and answer and so one of the one of the form factors of this was like can you create a personality that emulates
[00:01:55] A: relates a person and then retrieves from their body of work and reed has like five books you know a whole podcast and so we made the first version of it in chat gpt as a custom gpt
[00:02:06] A: But then people enjoyed the text-based interaction talking to this like this character that's pretending to be read and then it was like then we started getting we started working with these avatar companies It's like can we bring a visual layer to this an avatar and then a voice right like so then we started using 11 labs and in the beginning it was it was like pre-generated and then at like six months in some of these avatar companies were like Oh, we'd love for you to try our
[00:02:33] A: or unreleased real-time avatar technology basically us putting out every little experiment brought some attention to the project and then the different players in the space would reach out to us and be like oh what if we partnered on the next version of this real-time multi-lingual and so and i don't think i did alone like we've used a lot of ai to get here but then now it's like it kind of like manifests its own evolution as like we
[00:02:59] A: create more content
[00:03:00] B: yeah how do you end up meeting read like what's the actual story so you were a data scientist at clubhouse and then the next thing you do and you were working for reed just vibe coding tools can you tell us a little bit about that evolution
[00:03:11] A: so I was a data scientist at Clubhouse and ChatGPT came out in November of 2022
[00:03:17] A: and it was instantly the most popular topic in every like across this like audio app social network that we were working on and in and so then i was like seeing people using chat tpt in like all sorts of ways and then also some of my smartest friends were working at open ai on it and so i was like okay i wouldn't bet against these people like this is they've cracked something it already felt like when you know gpt3 chat tpt feels like it's 100 years early
[00:03:43] A: in my mind like you know all the science fiction that we've consumed you know c-3po things like that you're like wait it's already speaking every language and it's kind of getting like a b minus or maybe a c plus on like almost every topic which is kind of amazing and then um but then in march 14th march 14th of 2023 gpt 4 comes out
[00:04:06] A: And at the time it was 8,000 contacts token.
[00:04:10] A: Its context window was 8,000 tokens.
[00:04:13] A: And I was just like,
[00:04:15] A: okay,
[00:04:15] A: this is insane.
[00:04:18] A: It was basically the first time I'd experienced something that was way better than me at data science and analytics.
[00:04:24] A: It was better at writing SQL and Python than me.
[00:04:26] A: And I would ask it questions and then I would come up with the answer and I'm like, no way.
[00:04:30] A: Then I would be like, oh wait,
[00:04:31] A: that's just a more elegant solution than I could come up with when you look more closely.
[00:04:35] A: So at that moment I realized that analytics as a field was going to be done by these systems and that I would rather talk to this chatbot.
[00:04:44] A: to tell it to do analysis than for me to do it by hand ever again and then we had layoffs at Clubhouse but at that point in time I was like I was like three weeks into just like programming all the time with language models so I was kind of just like doing it after work so then when we when we got laid off
[00:05:02] A: then i was like okay cool now i'm just going to do this until i run out of severance so then i burned like four months of severance just talking to gpt4 every single day building things seeing what it's like the extent of its world knowledge at the end of those four months i was like okay i have way more questions than answers like this is this is insane and like i can't imagine going back
[00:05:24] A: to the real world right when i know that like this thing is like going to be powering the real world right so then i earlier through my 401k and i just like
[00:05:33] A: kept burning it on open ai api calls telling it to make stuff and it was funny this is the 8 000 token gpt4 model and and that november i was kind of just like oh that'd be cool if like i got access to the 32k context gpt4 model that'd be crazy then i could like you know we're going to mars like this is this is like i could i could run with this and by november they released 128 000 tokens gpt4 and then i was like okay that's
[00:05:59] A: So this is.
[00:05:59] A: The capacity to solve problems is growing faster than me every single day talking to it, trying to figure out how to use it.
[00:06:07] A: The ability for it to solve problems is moving faster than my ability to even process that.
[00:06:13] A: and that's all i do it's like all i do is talk to this model and it is getting fast it is getting better faster than i can like imagine what i would do with that new capability so it's a it's been and and uh yeah so that was the the gap between that was like after working at clubhouse i just went deep into the language models but then um
[00:06:31] A: I met Ben Rellis who works for Reid and he was like love for you to come up to the Bay Area and then I come up to the Bay Area I met Reid and some of his team members and we talked for five hours that day and then the next day he was like you should you should come work for me basically doing the same thing that you're doing but now you're not like it's not your life savings you can now like you have a team around you you have people you can bounce these ideas off of people you can build for
[00:06:57] A: before um so then i kind of like exited the cave re-entered like i guess the the real world in a sense
[00:07:04] B: You don't have to spend your life savings on tokens.
[00:07:06] A: yeah yeah although as i was running out of money was the very beginning of when i started getting contracting work to use language models so i felt pretty optimistic that okay there is like there is a new market for this new skill set which is the application of intelligence so i but i had to like
[00:07:23] A: kind of like leave like i had to put my old career aside stop thinking myself as a data analyst and start thinking myself as this like guy who knows how to use generative ai yeah
[00:07:33] B: There's a fascinating trend. It's really just end of two because it's you and Lindsay who's on our team here at Village of Global.
[00:07:38] B: You both were data scientists by training and both are very strong and vibe coding now.
[00:07:44] B: I almost wonder if there's something about being.
[00:07:47] B: technical enough to understand how to architect and speak to these models but still having enough humility that you can't actually software engineer right
[00:07:55] A: Yeah.
[00:07:56] B: like do you think that there's actually something to that
[00:07:58] A: Yes,
[00:07:59] A: I think there is a pipeline here.
[00:08:01] A: The pipeline from data analyst to vibe coder is,
[00:08:04] A: well,
[00:08:05] A: at least I can speak to my experience probably to the other and the other data scientists I've worked with.
[00:08:09] A: Olivia.
[00:08:10] A: So Lindsay,
[00:08:11] A: Lindsay and I, we have a mutual friend, Olivia, and she works in Anthropic.
[00:08:14] A: topic before that she olivia worked at repl.it before that i worked with olivia at clubhouse and so we were sitting there watching the language models do data analytics and then she base and i was making agents and olivia started working at repl.it and i was like put an agent in repl.it they put an agent in repl.it that you know repl.it went vertical um and now we have anthropic's cloud code going vertical right so i think the relationship between um the the pipeline from data
[00:08:40] A: to vibe coding is is like we don't like three years ago data analysts like we just were not full stack engineers the only programming languages we knew were
[00:08:51] A: sql and python and so sql is great for communicating and extracting data from a database and then python is great for
[00:08:59] A: analyzing that data visualizing that data analyzing it and then i i'm not even an ml guy like you you studied ml like i have not studied ml but i consider my superpower to be like data visualization like how do you figure out what's actually going and then show it to people like how do you create an interface that allows us to understand the data is what my superpower was and then
[00:09:20] A: we when you ask the language model to do these things and then it starts doing those like sql and python better but then it starts using html css and javascript to do the visualization you realize like oh like code is a general purpose solution and you have the benefit of knowing how to design like like how to think about the organization of different data types ends up being very useful for thinking about like back end
[00:09:46] A: systems so there's like a you have enough enough advantage you have enough strengths that you can that you're not like useless in your interaction with AI and then you're also humble enough to know that like everything the AI brings to the table is going to be like now your superpowers
[00:10:03] A: Right. You benefit from everything that it knows that you don't. Right.
[00:10:06] A: So I wouldn't get even a C minus in any of those other programming languages,
[00:10:10] A: but it doesn't matter if the AI can get like an A.
[00:10:13] B: This is something I've noticed with vibe coders too.
[00:10:15] B: In my own experience,
[00:10:16] B: I was also a data scientist similar to you and Lindsay.
[00:10:20] B: And when I use coding agents,
[00:10:22] B: I have enough of a sense of how you architect and can use systems, but I actually don't deeply understand,
[00:10:27] B: especially on web,
[00:10:29] B: HTML,
[00:10:29] B: some of these things, how to actually.
[00:10:30] B: actually build stuff javascript so it's almost like you need to be a level of abstraction down to i think really be super impactful but it's almost if you're too deep on it you you're too picky does
[00:10:43] A: Yeah.
[00:10:44] B: that sound right
[00:10:45] A: Yes. I think also the other thing is scripting.
[00:10:49] A: like and this was because at the time gpd like 8 8 000 token gpd4 and then 128k gpd4 i was talking to reed when i the day i met him and i was like this thing is writing like pretty much perfect sequel like i mean perfect is way better sequel than i'll ever write in my life and pretty insane python and and i was like that's this for me at that moment i was like this feels like you know we're gonna have these things i don't know how many years it's gonna be before it can build
[00:11:15] A: facebook or like 15 million lines of code kind of like code-based complexity and reed was basically like at the time he's like yeah but that's scripting like that's just scripting it's not that's not creating an application that can connect like maybe this camera back to like your computer and i was like yeah but if it gets smarter and a context window increases the per token intelligence increases and the context window increases
[00:11:39] A: none of those things seem like intractable problems and and he was like oh we'll see and then i think if you look at what's happening especially in the last year the leap the and just in the last week what people are now doing with the models i think it's no longer just scripting there's just scripts were like within the models capabilities back then but now it's like full stack full stack applications but also
[00:11:59] A: oh the agents can you write scripts and then reuse those scripts to solve problems so the agents are very good at scripting and then the scripts are just like automated ways to like get information manipulate that information it ends up being so scripting is a
[00:12:15] A: Yeah, like you call like the new programmer like the script kitty,
[00:12:19] A: but I think of that as like well the script kitty is also like the LLMs are also like script kitties
[00:12:25] B: So you've been using a bunch of these tools and through read you have access to some of these tools before they're even necessarily launched to the public.
[00:12:32] B: And you've been kind of testing this stuff even since ChatGPT launched,
[00:12:36] B: right?
[00:12:36] B: And you've actually been spending every day inside of these models.
[00:12:40] B: How do you think about the evolution,
[00:12:42] B: especially of coding agents,
[00:12:43] B: since you first started using them up through where we're at today?
[00:12:47] B: Like, what is that evolution and how do you think like we are where where are we now on the curve?
[00:12:51] A: You know, GPT 3.5 16k context was when I this was like right before this is like February,
[00:13:00] A: March,
[00:13:00] A: April 2023.
[00:13:02] A: That was when I was like started programming with language models creating my first chatbot like the first thing I asked a language model to do chat GPT I was like teach me how to build like a GPT powered chatbot and then ends up being 20 lines of code and like wow this this is very meta you can just tell chat GPT to write another program that uses a GPT and then ends up being so like it's very small it's a small program and then at the same time two projects went viral.
[00:13:27] A: in the broader ecosystem it was auto gpt and then the second one was baby agi by yohei nakajima who
[00:13:34] B: Love
[00:13:34] A: you yeah
[00:13:34] B: your hair.
[00:13:35] A: yeah so baby agi was created by yohei and i'm looking at him saying wait he's a vc interesting and then i look at his i look at the code for he had open source baby agi and baby agi was this chatbot that you would give it a job and it would just like autonomously work in a loop towards that goal break it down into sub tasks and then try to crush those sub tasks in pursuit of that goal and of course it didn't like
[00:13:56] A: I mean it was it was more that the idea of baby AGI was like oh my god if you configure a chatbot as an agent then like how far will it go is the question like can you just wind it up and set it off in a direction
[00:14:08] A: and I looked at the code for Baby AGI and it was like a hundred lines of code and this is what's awesome I was like I understood that it was Python it was my programming language so I was like oh wait this is this is this I understand and also I was like wait Yohei I mean he's an investor so I was kind of just like okay interesting like this feels like a game that I can I can play and so then I basically followed down that path I forked Baby AGI myself gave it like a chroma DB knowledge base
[00:14:35] A: is and so the first agents i was building would use like 15 tools gpt 3.5 16k context and then um some vector store and then they would like operate in a loop and then i would give them voices and then i was like okay this is like every every like three weeks i would come back to a project and give it some new capabilities and like you know rebuild it from scratch and uh that was the that was like the
[00:14:59] A: this wave of like these chatbots as agents 2023 there were certain other projects that like clued me in there was okay one several years ago i think andre karpathy went on lex friedman and they were talking about agi what form factor would agi take and um he he proposed two possibilities one which was that um it would be some product evolution some iteration some product that's an iteration of github copilot
[00:15:28] A: And then that like this, this, this like basic autocomplete on code, if you extrapolate that out would eventually become some kind of AGI.
[00:15:35] A: And then his other theory was like, if that doesn't get you to AGI, it would be embodied robotics in the real world, giving them sensors and the ability to like reason in physical space looking at GitHub Copilot, which is the first.
[00:15:47] A: Which even pre-days chat GPT is the first commercial application of LLMs, like the first successful one.
[00:15:52] A: And I think a lot of people interacted with GitHub Copilot that ended up going to work at OpenAI that were like, oh.
[00:15:58] A: they could see this like connection but that was like it used to finish one block of code like it would write the rest of this function and you'd be like oh yeah cool i like it except but that was the beginning and then it was like there was another project that i saw that was um oh this was in september of 2023 which was called open interpreter and um it was basically you let gp you let these language models write code on the operating system level kind of like bash commands and
[00:16:26] A: and like arbitrary python code kind of radioactive because we weren't really sandboxing any of it but you were like wow this is so much cooler than like clicking around the screen because you could just it could just move through the computer more quickly than you could and
[00:16:42] A: An open interpreter,
[00:16:43] A: I think, is like a pre-evolved form of these CLI agents that we now have,
[00:16:48] A: Cloud Code and Codex.
[00:16:49] A: Cloud Code and Codex, kind of just like the evolved form of that.
[00:16:53] A: And Cloud Code came out in 2025 February,
[00:16:57] A: but it kind of went, for me, like I realized it was a general purpose technology probably in May of 2025.
[00:17:06] A: And by June, I was putting out, I was like burning 650 million tokens a month.
[00:17:11] A: in ClaudeCode because I was telling it to like do my expense reporting.
[00:17:14] A: I was telling it, I was realizing its applications were much broader than just traditional software.
[00:17:20] B: So we've pretty much gone from.
[00:17:22] B: copying and pasting code from ChatGPT
[00:17:24] A: yeah
[00:17:24] B: to having things like Cursor where it's actually a GitHub Copilot
[00:17:28] A: yeah
[00:17:28] B: where it's filling it in.
[00:17:29] A: I mean I use Cursor for like a year and a half like 25 which is my the usage only goes up right because you use it then you learn how you can use it and then you apply it to more things and then every day you spend more money and Cursor I was spending you know $15 $30 a day sometimes $50 if I just didn't sleep and it would just go up and I was like it's worth it it's like other people you know how much you spend on coffee every day
[00:17:51] A: Every day you should at least spend that amount on coding agents.
[00:17:56] B: So you were using Cursor for a long time and then now you switch pretty much entirely.
[00:17:59] A: Cloud Code and
[00:18:00] B: yeah yeah
[00:18:01] A: Codex.
[00:18:01] B: yeah when i when i use cloud code for the first time then i kind of stopped using cursor and it was surreal because um because i was like cursor is incredible i was the person that introduced all my friends to cursor and then when i use cloud code i was like oh my god what if we don't need the ide because all of a sudden when you're talking to cloud it's writing all this code you're not really looking at the code and then i think about the application of cursor is like uh it's like vs code right so it's a vs code fork with ai bullet
[00:18:28] B: I bolted on and I think about VS code, that's 12 years of an environment designed for human beings.
[00:18:35] B: to code but now the ai is coding so like what is so maybe that environment is like it definitely needs to evolve because i don't look at the code i look i like talk to the agent and then i you know i ask it you know i ask it for things that would prove that the code works versus me looking at because it's generating a lot of code and we're not most people are not looking at it and code review has ai for code review has to get better it will
[00:19:01] B: well but it's clear that people's preference is to make things and operate at a higher abstraction level you can always look at the code if you want
[00:19:11] B: But I think that that so that that that's when like the terminal wave like the people coding in the terminal because what kept people from going to the terminal was for at least for me, it was like, I don't know all the commands of the terminal. It's very arcane what these like different like using a terminal required knowing these commands.
[00:19:31] B: But what I learned from Open Interpreter was that you could just speak in English if an LLM turns those into terminal commands.
[00:19:37] B: So then you see the same thing cloud code.
[00:19:39] B: You're talking to Quad,
[00:19:40] B: but then it's translating your requests into the Bash commands and tool calls that would operate your computer faster than you can click around it.
[00:19:48] B: And so that was like a learning from, it was a learning from Open Interpreter that made its way back into Cloud Code.
[00:19:54] B: And then at the same time, codecs started getting better.
[00:19:57] B: um gave some feedback to my friends working on codex i was like i have my opinions from playing with these tools but then codex i think is actually smart i suspect codex is smarter than than claude it's just that maybe claude kind of created the category and uh so they have certain insights on like what people want but my experience with codex is i can tell it to go work on something for a whole day
[00:20:19] B: and then I don't really care if I'm not like having an interesting conversation with a chatbot because I'd rather treat it more like a contractor that just goes go work on this for two days and then come back with like a superhuman amount of work and it's a little different from Claude I feel like Claude is more like it's more like Jarvis where like I'm in the loop I'm like in dialogue I'm iterating with this this this co-pilot system
[00:20:43] B: it's not clear to me that like that the end goal is for us to be talking to these systems all the time like maybe the goal is for this thing to just work in the background while i live my life like that that so it's going to be a matter of preference but i play with all of them and i also like replit a lot so um
[00:21:00] A: It kind of just like your preferences change,
[00:21:02] A: but then a good sign is like which model is getting most of your hardest questions.
[00:21:07] B: So today,
[00:21:07] B: February 6th, 2026,
[00:21:10] B: because you're moving so fast that we actually have to date it,
[00:21:12] B: how would you stack rank both the applications? So we have cloud code,
[00:21:16] B: you have codecs, as well as the deeper models underneath them. Like, how do you think about who's winning the race? Is Gemini even in the race?
[00:21:24] B: Can you explain how you got your mental model of all of it?
[00:21:26] A: this is all my opinion um yeah
[00:21:28] B: For your friends who are at the AI labs so
[00:21:30] A: yeah
[00:21:30] B: they don't
[00:21:31] A: yeah
[00:21:31] B: get upset
[00:21:31] A: it's
[00:21:31] B: with you.
[00:21:31] A: yeah exactly especially now that they're all banning each other from using the models so like maybe my opinion matters but i think it's uh um so i think as a if we look at the harness the harness being like the wrapper um like how the model is organized and then the way that the person interacts with it i think claude code is probably the most like intuitive harness the human and loop experience
[00:21:56] A: I think Codex has a higher raw intelligence level.
[00:21:59] A: I mean, people complain about how much Codex thinks.
[00:22:01] A: I don't think that matters.
[00:22:02] A: I think like you go to your smartest friend with a hard question.
[00:22:06] A: You don't need the answer immediately. You need the highest quality answer to the unsolved problem.
[00:22:10] A: And I think that Codex is more often than not figures the arcane solution to like the hard problem.
[00:22:16] A: So like you hit a wall with cloud code and you give it to Codex. Sometimes Codex can crack that.
[00:22:20] A: So I feel that like the per token intelligence.
[00:22:22] A: versions of codecs might be better.
[00:22:24] A: The other thing I've noticed is that you can set up codecs to work by itself for extended periods of time.
[00:22:31] A: if you give it a proper planning framework so it's like here's how just if you want to do a complex refactor follow this kind of a meta approach to planning and you give it a little planning framework same thing with like here's how you should write architecture docs so then the model gets really good at working on larger context problems like more complex problems codex is really good at that i think cloud code is also better i mean yesterday opus 4.6 came out and codex
[00:22:58] A: 5.3 came out and just seeing the first examples of what people are doing with it
[00:23:04] A: there are more complex things than i've ever asked a model to do like apparently 4.6 opus 4.6 can build a c compiler largely by itself it takes like a week but it can and then um i mean even cursor using 5.2 got it to 5.2 5.2 codecs got it to build like a web browser
[00:23:23] A: um and that's like more multi-agent stuff running many of these concurrently so i think the ceiling now is it's no longer like can it build this web app yeah it can build a web app like easy um it's more like
[00:23:36] A: Does it start operating more like an organization?
[00:23:39] A: Because I think they are.
[00:23:40] A: And I think that, I mean, you should just put these things inside every single company.
[00:23:45] A: You should.
[00:23:46] A: You'll be surprised at how much infrastructure and how quickly they can build it and how quickly they can understand and attack more complex problems.
[00:23:52] A: Gemini, I mean, it has its strengths,
[00:23:55] A: large context,
[00:23:57] A: real-time multimodal. It's good for like voice.
[00:24:00] A: um my and this is maybe maybe not true anymore maybe i gotta play with the new gemini 3 pro it's pretty good at like design and ui stuff i just feel like that kind of stuff doesn't matter as much like i've it almost like it's like the interface is now voice the interface is now like language like does the
[00:24:18] A: like i don't want to click around a web app or even a desktop app anymore for for that matter i kind of just wanted to just like i want an agent to just do that behind the scenes i shouldn't have to remember where things and how where buttons are and stuff like that but i'm kind of like i've like taken that pill of like the ui is melting i prefer language as the ui gemini is really smart but i find that it doesn't have this like propensity to work
[00:24:45] A: um some people call it laziness where it's like you know you tell the problem you you give the task to the model and then it's like great here's what we should do and i'm like
[00:24:54] A: so go do it like the prompt ends i end up being like go do it go do it just start go and i hate that because i'm like the bias towards action just isn't there sometimes so this there's a it's like raw intelligence but then like how agentic is the model like how likely is it to just execute a bunch of tools and then do the thing versus having a good idea which is fine but like you kind of also want the model that just like does 15 things reflects does 15 more things reflects and like makes momentum
[00:25:21] A: and so it's and then there's the third thing which is like if it's a chatbot that you actually talk to then eq matters so i think that's where opus might have an edge like i think anthropic models tend to have higher eq in my experience it's like you can tell with like more people enjoy interacting with them i do wonder like are we just attached to us being in the loop a little bit too much should we take maybe we should be getting out of the weeds a little bit more out of the way of the model so that it can go fast and go
[00:25:48] A: Oh, at the scale of a machine?
[00:25:51] B: Cloud Opus 4.6 dropped yesterday,
[00:25:54] B: as well as GPT 5.3 codecs.
[00:25:59] A: 5.3 codecs, yeah.
[00:26:01] B: Can you just break down what your initial thoughts are on them and what you've been seeing in terms of what the state-of-the-art models are doing right now?
[00:26:09] A: Step one is I think a lot of people don't really have a setup where they can kind of run any of these, like many of these at the same time.
[00:26:16] A: So I spent most of yesterday just redesigning my desk,
[00:26:19] A: my workspace so that I can very quickly switch between any of these.
[00:26:23] A: And so I use terminal multiplexing.
[00:26:27] A: so tmux for like seven months but basically it's a way to use multiple terminals very and switch between them very quickly so you can very quickly spawn new agents and then run run them side by side the interesting thing is that clawed opus can now do like multi can basically do that under the hood so instead of me spinning up multiple agents clawed opus can also spin up multiple agents under the hood and it uses the same approach terminal multiplexing so it spawns external terminal planes and then delegates to sub agents that
[00:26:54] A: that do different parts of the job so i think that's where the the new game
[00:26:59] A: is like their ability to like fan out across a problem so like concurrency parallelization and to run many of these sessions in parallel on more complex problems and I think also like they're both good at using get work trees which basically like allows them to work on copies of the code base at the same time so now you can take you can you can say it's not just build what I want
[00:27:25] A: Build n versions of what I want.
[00:27:27] A: Let's compare all those solutions and pick the best one.
[00:27:30] A: And so now it's more like delegating to a system that can spawn a team to solve your problem.
[00:27:36] A: And that's a new mental model for work that it's no longer like my attention,
[00:27:41] A: which is single threaded.
[00:27:42] A: It's like me giving a problem to a system that is multi threaded that can attack the problem with parallelization.
[00:27:49] A: That's what's unlocking some of the newest.
[00:27:51] A: magical upside of the model like i think if cursor can use parallelized instances of codecs to build a web browser in a week then like it's proof to me that like we have basically swarms of like like one agent quickly spawns many attacks a more complex problem than one agent could same thing um opus spawns up i think they call them teams because swarm is probably a scary term but
[00:28:20] A: Spawns a team of agents that takes a complex problem and then like makes it more tractable for them.
[00:28:26] A: that's i think that's the new that's the new like that's that's the new game is orchestration of many of these at the same time and then i've already felt in the last six months that i'm at the limit of like my own ability to manage them i'm like okay this is i feel like a manager now i'm an ic all i do is talk to ai but now i'm like i'm managing 45 different agents and then it's like i'm the bottleneck right so then you remove yourself and you allow the agents to manage subagents and then all of a sudden
[00:28:55] A: and all of a sudden the the dunbar number or like the the ceiling for orchestration goes up and so that's that's the new game i think Can yeah
[00:29:03] B: you explain exactly how you have done the multiplexing and coordination like that still seems like it's a hard thing to do and it feels like this kind of level up for people who might want to go from, oh, I'm using flawed code to whoa, I'm managing these swarms.
[00:29:15] A: yeah basically i talked codex about the hardest problem i have all the time like every day you can go to the go to your smartest model and you ask it the hardest problem
[00:29:21] A: from and at the time I was like man I feel like I open up these terminal tabs and then I send these coding agents in different directions but then I'm limited to the number of tabs I can quickly access and so that means I'm limited by the number of fingers I have it's like command one command two command three I use a lot of hotkeys to switch between them then I was like but that's like a physical constraint and then there's the mental constraint of like how many of these can you actually remember that you have on what does that UI look like and
[00:29:48] A: And then also,
[00:29:49] A: like, if I close it, then I lose everything like, you know, you close a tab, you lose everything.
[00:29:53] A: And codecs was like, well, actually, what we want is persistent processes that run in the background where.
[00:29:59] A: you spawn it it exists it works especially when codex can work for like codex can now work for like over a day so what are you going to just like walk away from the computer like no you're going to spin up another one and attack another problem codex was like well what we could do is use is tmux terminal multiplexing tmux was invented in 2007 or 2008 so it's like an old school solution to to like creating persistent state running processes in the background on your computer
[00:30:29] A: and then being able to very quickly like pull it up when you need it but that was an old solution to this problem that computer engineers had already solved and codex was like oh we should maybe use t-max that might solve your problem and so then i i told it to configure my t-max i was like okay install t-max configure it for me i don't like the default hotkeys mod them so that it's more like starcraft like pick the hotkeys that are more ergonomic for me and i would basically like i when i do this i want to spawn an agent when i do this i wanted to split the pain
[00:30:58] A: and then give me a second like one on the side when i do this just give me a cloud it was very much like i want this go hack my computer to make that possible and then it would do that
[00:31:09] A: And then one day, three months ago, I was talking to Codex and I like asked for something and then Codex spun up another Codex agent.
[00:31:18] A: Like it spun up another session and attacked the problem in parallel.
[00:31:21] A: And I was like, wait a minute, because I was already using Tmux, I was like, I could see the process that Codex had spun up. I was like, oh, Codex is using Tmux now.
[00:31:29] A: So then I'm talking to Codex, I was like, wait, what's the likelihood that like Tmux is a long term, like me using Tmux is a long term.
[00:31:35] A: term thing and it's like well no actually it'll probably be a substrate that the agents work on you just like right now you're using it but the agents can also use it so then maybe like that's you were using it today tomorrow the agents will be using it and then i saw that happen so multiplexing now the agents can multiplex cloud code can multiplex it could basically spawn multiple copies of themselves to parallelize their solution to the problem so
[00:31:59] A: and i was talking to lindsey and i was showing her like i was like that kind of approach and she was like oh wow you're like the f1 driver of vibe coders and i'm kind of just like yeah but the agents will also be f1 drivers of vibe coders in fact they can manage way higher context than us i accept that what i do today might be done by the agents tomorrow and then i will move to maybe like maybe i'm like a ceo of this like
[00:32:24] A: agents that have their own swarms.
[00:32:27] B: So you spend about 14 hours a day talking to AI.
[00:32:31] B: Can you talk through a little bit more of your exact workflow or what a day looks like for you in terms of coordinating these agents? How are you using codecs alongside cloud code?
[00:32:40] B: Like give us a little bit more of the tactics of how you actually like open up your laptop and start attacking a problem.
[00:32:46] A: Basically I wake up, I order my coffee,
[00:32:48] A: I brush my teeth,
[00:32:48] A: you know, take a shower,
[00:32:49] A: I go down, pick the coffee up,
[00:32:50] A: and I sit down at my computer.
[00:32:52] A: And because I have Tmux, all of these sessions are just like on.
[00:32:56] A: Basically, the fleet of agents is just waiting for the next instruction.
[00:32:59] A: And when I think about how they're organized,
[00:33:01] A: every session,
[00:33:02] A: every project I have that I'm working on, like ReadAI,
[00:33:05] A: or it might be this like translation of our podcast,
[00:33:08] A: every single project has its own pod of agents assigned to it.
[00:33:12] A: And it starts off with one, but then I spin up using Tmux, I spin up more panes, and I assign other agents to different parts of the process.
[00:33:18] A: So it might be one's working on the app,
[00:33:20] A: one of them's working on like refactoring a copy of the codebase, another one's working on maybe analytics,
[00:33:26] A: and the third one might just be doing research on like other form factors of the solution.
[00:33:30] A: And so every single project has its own pod of agents.
[00:33:34] A: Because of multiplexing allows me to do that orchestration. Otherwise,
[00:33:38] A: I would be limited to three tabs.
[00:33:40] A: and three agents but i had to get beyond the three tab three agent limitation now it's like an arbitrary number of projects every idea that me and reed have just has a set of agents that are just like on that idea and that's their only job they exist in that folder on our computer
[00:33:57] A: working towards that project it might be an animation project it might be like just random creative a lot of it's like random creative stuff some of them are just working on like random game ideas that we have or like research stuff one one pod's working on my website all the time and so I'll have like a Claude a Gemini and a Codex working on the website but if you think about it where we're about to go is that that Claude will be able to summon other Clods the Codex will be able to summon other Codexes and I like my approach because I can be
[00:34:25] A: model agnostic like i can spin up cloud or codecs but you'll never have i mean it's unlikely that codecs will like opening i will ship a version of codecs that spins up cloud even though it can technically it can just spawn a cloud so you can tell gemini or codecs be like go send codecs at this problem it'll spin up another pain and then summon an agent um but i like i like being able to switch between all the model providers because randomly like i don't know
[00:34:51] A: I don't know.
[00:34:51] A: It might be that there's some interesting form factor.
[00:34:55] A: I think we're going to see the speciation of these systems, slightly different approaches to these problems,
[00:35:00] A: depending on the different model strengths and weaknesses.
[00:35:03] A: Some experiences might be more human in the loop than others.
[00:35:07] A: And if the coding agents experience this speciation,
[00:35:10] A: I think that's a good thing.
[00:35:11] A: Like, maybe there's a version of coding which is more compatible with like having a life.
[00:35:17] A: you know like not being at your computer every day i i hope that also pans out because otherwise it's like what are we doing here you're
[00:35:24] B: You're a huge fan of voice as kind of the next medium and you use whisper flow a ton 140 words per minute So
[00:35:31] A: 140 words per minute and my typing speed is dropping yeah
[00:35:35] B: you're slower at typing now and it's like 70 words per minute.
[00:35:38] B: I think you're telling me
[00:35:39] A: so uh yes last week i did a speed test on typing and i was at 70 a year ago i was at like 85 there's no way my typing speed
[00:35:45] A: speed goes up why would it go up there's no reason to be faster at typing
[00:35:49] B: Proof that AI is making us stupider.
[00:35:53] A: so it's about the quantity and quality of tokens and i think uh my person i like to talk i don't write as much maybe
[00:35:59] A: like i'm working on writing more i think writing is important i didn't understand that until more recently in life read would agree yeah yeah i mean it's like the compression of thought right it's just that when i'm interacting with ai it's more important that i get the problem out of my head the ai is smart enough to like clean it up and i noticed this when i do typing tests is that like i got so used to language models that like when i'm doing the typing test it's like i'll miss the punctuation and then i hit backspace and i'm like
[00:36:27] A: The problem with typing as a in my opinion is that we can't think faster than we type. You're not thinking about the fourth sentence after and so your thinking speed comes down to your typing speed and then every time you have a typo you go back and like you're like you hit the brakes on thinking to go fix the typo.
[00:36:48] A: Speech is closer to thinking speed than typing is.
[00:36:52] A: I think writing is important as a compressed form of thought,
[00:36:55] A: but it is not the right way for us to be working with AI.
[00:36:59] B: Speech is your main form of actually giving context to a lot of these LLMs and even the coding agents themselves.
[00:37:05] B: Can you explain why and like why is that an important thing you think for people that are using these tools?
[00:37:11] A: If I wanted to describe a problem,
[00:37:14] A: And if I can speak at 140 words per minute,
[00:37:17] A: then I can describe the problem very quickly.
[00:37:19] A: I can basically speak an essay in two minutes.
[00:37:23] A: And then I look at that transcript and it's like garbled. It might be like stream of consciousness,
[00:37:27] A: a little bit less structured.
[00:37:29] A: but then in llms on the other side of it it's like oh these are the main things you want me to focus on it like distills it down into the actual plan and then i can validate that plan so there's still a visual element like seeing the llm process your speech and then being like yeah this is this is where we're going versus typing three pages i'll never do that like i'd never like that's just that'll take like days and then i'll get caught up in like the structure and the neatness and instead of just
[00:37:54] A: What you want is that raw brainstorm capability and the speed and the depth at the speed that you can think.
[00:38:00] A: And so voice allows us to get this high bandwidth output from our mind into the model.
[00:38:07] A: and then llms are able to like make sense of that in a way that it's you're not writing a document you're like talking to something that like is capable of that level of depth and nuance this
[00:38:18] B: One of the things that you also do is you ask the coding agents to ask you questions before you go in. So you're like, prompt me,
[00:38:24] B: can you talk about how exactly you do that? And what are the types of inputs that you give it so that it can get the best inputs from you?
[00:38:32] A: is a common mistake i think
[00:38:33] A: I think most people make interacting with AI is that they go to the AI thinking they know exactly what the answer is.
[00:38:42] A: And I think it's better to have the opposite kind of like mindset is like, is like to go to you,
[00:38:49] A: like I go to the AI and instead of saying I want you to make this,
[00:38:53] A: I think this is the problem I want to solve.
[00:38:57] A: And I sit there and I talk for like a couple minutes.
[00:38:59] A: Like, here are all the ways I'm thinking about the problem space,
[00:39:03] A: but the shape of the problem might not even be how you.
[00:39:07] A: have already predefined it the shape of the solution might be different from how you you define it basically i assume that like the solution might require information and knowledge that i don't have and i go to the ai it's like tmux right i would never have pulled some arcane technology from 2008 to solve my orchestration problems but the llm was able to surface that as like a potential solution and i'm like okay let's try it and i was like open to the llm being
[00:39:34] A: smarter than me open to the llm having the answer and it's and it's in its processes and then for it to unpack the problem with me and then for us to like think about possible solutions it's like the system you're trying to design a system but you don't necessarily even know
[00:39:51] A: those parts and to get the right system as the on the other end you have to really just focus on describing the problem describe the problem then allow the intelligence to unpack possible solutions and then you kind of pick directions on which ways look promising and you're like follow-up question why would i pick t-mux over zealage and then it's like well zealage is actually more it's also it's a t-mux alternative
[00:40:14] A: for session management for the pros and cons i had a three-hour conversation yesterday with codex about tmux versus zelig and i was like well if i value the agent being able to to use the tool tmux might be the right way to go but if it's me and my fingers orchestrating the tool then zelig is the way to go because the hot keys are more human friendly um
[00:40:35] A: there's no such the agent can just use t-mox out of the box it's kind of more complicated for a human to use but it ends up being slightly better on one axis for for a coding agent but that's something i realized in conversation with ai about like look both technologies i'm not the master in but we're going to have to choose one of them and so i have a dialogue about the problem and then it's like okay let's what are the pros and cons of going in either direction one is more ai agent native and one is more human native
[00:41:03] A: innovative so i'm willing to to like spend three hours talking to ai about like the space of solutions with technologies that i've never used so that the next
[00:41:13] A: Four months of what we're using is like slightly better judgment.
[00:41:16] B: One of the places that people have tons of problems with these coding agents is that they just hit walls,
[00:41:20] B: right?
[00:41:21] B: Like there's bugs or something that pops up or they don't know how to fix a specific problem or maybe how to structure the problem right.
[00:41:26] B: Asking questions can help in making sure that you're getting that right input to start as well as, you know, using voice, I think, to include more context on those problems.
[00:41:35] B: How do you break through?
[00:41:38] B: problems that come up with a coding agent where maybe can't figure out a solution or the output is buggy and like the end product is actually not exactly what you want like what does QA process look like how do you actually break through some of these challenges
[00:41:50] A: The first kind of call that I make is usually it's like, okay, if I'm interacting with an agent,
[00:41:54] A: say it's like Repl.it or Claude or whatever,
[00:41:56] A: and we're kind of hitting a wall or like we haven't made any progress.
[00:41:59] A: in like two or three hours like i'm kind of like in a loop like a death loop usually i think everyone eventually experiences these like death loops usually that is a sign that the context window is now polluted with a bunch of irrelevant information so like you're you're kind of like after like you're like 25 interactions in on this problem and then the agent's kind of like trying to hold a bunch of irrelevant context in in like in his context window while it's talking to you so usually like
[00:42:29] A: first step is to spawn a new agent that has a fresh context and have it attack the problem with a clear mind um the second thing is and i try to do this before you spend three hours like you you might want to try some of these tactics like earlier rather than like accepting it's like you're at a wall the second thing is to give it to a smarter agent and usually it's like the smartest agent i mean i think you should just only be using the smartest models
[00:42:56] A: because they have a higher ceiling than everything else it's just better to use the smartest model maybe not serve the if you're building a wrapper and then you're going to serve that to 100,000 people maybe then you should be conscious about which model you use but for building the first version of anything use the smartest model available if that model hits the wall
[00:43:13] A: um like it might be this context rot issue and it might just be that like what you're what you're asking for is like like the way you frame the problem is a little bit like incorrect and so generally like taking a step back and then also you might benefit from waiting um for the next model improvement to solve that problem so it's like it's
[00:43:35] A: A lot of times, you know, you hit the end of part of it is like the model, like the problem's complexity might be at the limit of the current version of the model.
[00:43:44] A: But that capacity to solve problems keeps going up. Like, I don't know what the current limit of the coding agents are and like codecs and 4.6, I don't know what they can't do.
[00:43:55] A: And the only way to find that out is to keep giving them hard problems.
[00:43:59] B: Are we at the point yet where you could use...
[00:44:02] B: a coding agent to develop it and then you can use some sort of an agent on the UI end to then test it click different things and feedback to the coding agent like if you're thinking about actually how does it QA itself have you done that before and kind of what's the the best methods that you've seen
[00:44:18] A: My instinct usually is to build tools that the agents can use and then secondarily build a human layer on top of that.
[00:44:26] A: So first, it's like what tools can agents use,
[00:44:28] A: like APIs?
[00:44:29] A: um and cli tools so like you're talking to an agent that operates in the cli the agent can also use its own terminal and spin up more terminals so giving it a tool that it can use in the terminal um allows it to very quickly build essentially like the back end of the application or like the tool if the agent can use the tool and then it'll keep using the tool to and improve the tool until the agent is good at using the tool that's step one
[00:44:59] A: Step two is build.
[00:44:59] A: building an interface on top of it but the CLI tools are very testable like the agent can use it and then review the output it's a it's very native kind of experience for agents to just like they have the closed loop of like input output they can read the help docs they can update the CLI tool to be better and then the web UI stuff browser agents are better than they've ever been so like clicking around the screen
[00:45:24] A: to like test things so you can basically like you can take your website and you can say claude like you're working on your website with claude and you can tell it to
[00:45:34] A: show pretend you are a mobile user coming to the website for the first time and then i want you to like explore it and then give me an analysis on your usability and it'll basically resize the it'll render the the website it'll resize it so that it's simulating a mobile experience then it'll interact as if it's a user visiting your website so you get the role play of the model pretending to be someone that is coming to your website it's you get the like simulated interactions of like
[00:46:00] A: like scrolling and clicking on a mobile interface and then that feedback comes back into the coding agent and then it can make improvements so that's great because it's like a closed loop of like it can test the it can test the problem it can test the the current state problems that are not testable or where it's not clear what success and failure looks like that's where actually i think we come in
[00:46:20] A: uh where there's more subjective aspects to the problem that's where like we should be we should generally human beings we should just generally be in the space of more subjective stuff because anything that's like automated automatable
[00:46:32] A: we will not like bring as much value to the table
[00:46:34] B: One of the things that you're a huge proponent of is context engineering versus prompt engineering.
[00:46:39] B: You've hit on maybe a little bit of components of this,
[00:46:42] B: but can you explain how you think about that?
[00:46:43] A: so i didn't coin it it's coined by dex horthy who's like he's a one of the one of the fellow clawed addicts out there um but basically it's this i think we in the in the gpt3 era we the term came up a prompt engineering right designing prompts
[00:47:00] A: because you get the model to like output something but then i did like a taxonomy analysis this was like three years ago i did a taxonomy analysis on what are all of the different prompt engineering techniques and at the time there were like 180 at least everything from like show the agent a couple examples of how you want the task done then it's more likely to follow that pattern or like you know giving it and like allowing it to look at your screen allow it to look at the problem visually right and that's a technique that's interesting because
[00:47:26] A: When you provide an image alongside the instructions,
[00:47:31] A: actually I think an image is worth a thousand words.
[00:47:34] A: So you're actually communicating a lot that you wouldn't have otherwise typed.
[00:47:38] A: So like multimodal prompting ends up being very useful.
[00:47:41] A: um so those are all prompt engineering techniques like i think step by step was a prompt engineering technique that now is the cornerstone of like reasoning right thinks the model thinks step by step and then it's like use tools while you're while you're thinking to like solve the problem so tool use they all became like prompting techniques
[00:47:59] A: But now we have so context engineering is a little different. It's more like I would say it's the way I see it is like thinking about the thinking about the like AI's bandwidth like for solving a problem as like a canvas that you can fill with context.
[00:48:15] A: And so like maybe part of that is like these are the tools you're going to need.
[00:48:20] A: to test the application you're working on these are the documentations of how we currently do these do this process like what's the relevant context that you need to bring to the ai such that it has a more likely chance of solving the problem and being very intentional about making sure that the irrelevant context is not there because actually like models do better when
[00:48:41] A: They don't have irrelevant context.
[00:48:45] A: They're singularly focused and everything that they need to solve the problem is like right there.
[00:48:50] A: or easily retrievable and then they have like clear access to that so the context window is kind of like a canvas and you should kind of you should be very pretty intentional about like not introducing garbage into the context or like rethinking like is the coding agent trying to is it being distracted by a lot of things that didn't work that are not relevant to our problem and then also this is goes back to why I like why we like CLI tools over say MCP so like creating tools for agents there's a whole wave of
[00:49:17] A: wave of mcp and the side effect of using mcps is that every time when you build an agent and i i have like a one of my agents on my mac mini has like 25 mcp servers and that was just like oh we're just i'm just going to give it all these tools but then the way it works is is that like
[00:49:36] A: every single tool has a description on how to use the tool so now the agent is trying to memorize
[00:49:40] A: how to do use 25 different tools but most of its interactions are just conversational so like it's holding in context a bunch of tools that are irrelevant to the main the task at hand right you may not need web analytics for every single normal interaction you might not need like slack integration for every normal interaction with the agent and so what happened with mcp i think is that like
[00:50:02] A: Like,
[00:50:02] A: we built,
[00:50:03] A: we basically allocated an increasing percentage of the model's bandwidth to tool, guides on how to use tools that are irrelevant to every interaction.
[00:50:13] A: So the reason why CLI tools are becoming more popular.
[00:50:17] A: is because part of a CLI tool is like you can if the model just says a dash dash help when it uses that CLI it gets a guide on how to use a tool so on demand it's able to see how the tool should be used but it's not trying to pre-memorize everything because you only have a certain amount of context window to work with so when you and this is a property progressive disclosure read the doc when you need the tool then you can use the tool it ends up being better than like memorize all of these tools
[00:50:46] A: Up front.
[00:50:47] A: So the model that like can dynamically retrieve the relevant context when it needs it ends up performing better because more of its mind is free to think about the problem.
[00:50:55] B: When you're talking to founders or people who are trying.
[00:50:59] A: convert their teams into being AI native or you're talking to teams that are already AI native that are working at kind of 11 on the spectrum.
[00:51:08] A: What are they doing differently?
[00:51:10] A: What are you coaching people on in terms of how they can actually build their businesses to be ahead of the pack?
[00:51:16] B: So usually it's like take the most technical person that you know and give them the best the best technical model there is like codex or cloud code and then
[00:51:26] B: actually i think you kind of have to play defense for them you need to take everything like you got to let them spend go deep and i think you need that person to like spend so much time with ai that they reimagine like their own capabilities like they like learn how to use these superpowers
[00:51:41] B: and then they can start reimagining the business and part of it is that coding agents move through cyberspace at speed and quality at a speed of knowledge work that was never possible when humans were when compared to humans and that person basically has to do a mini version of what i did which is like exit their own work and then like learn this new way of doing work and then apply it to the business
[00:52:10] B: and part of it is like coding agents are really fast rapid prototyping so you shouldn't you shouldn't have like a three-week timeline of like making the first version like why can't we make the first version of something today that's something i've always you know me and reed always talk about it's like what if we just like ask codex to do it see how far it goes today and then at least we have some of the hypotheses are unpacked and like that breaking that cycle of the speed and like breaking the the like molasses of the existing org
[00:52:36] B: work that's why you kind of have to take this person and like isolate them and allow them to like rip the smartest model and you got to just let them you got to invest in them like you got to like let them burn a bunch of tokens you're not going to learn how to use these tools if you're being conservative about how many tokens you're burning actually that number needs to go up
[00:52:53] B: like every day like um and i don't think that people have like realized this yet and i think that it starts with the you should start with the most technical like that person like the very amplified single high agency person that can now and then they should be allowed to attack any problem because they'll start viewing the entire business as a context window that they could be possibly plugging into and then the other thing i've noticed is you
[00:53:18] B: like because you can make things very quickly you can actually like not every team can do that so then you're if your competitors can't really think this quickly cannot build this quickly that's actually your opportunity and maybe like you can personalize your solution towards them um you can you can like
[00:53:38] B: you can build things just in time like very quickly to solve their unique problem and that can be a huge advantage and then now your sales reps can kind of like move in a different way knowing that the solution it's not like sales you know the classic dynamic of sales reps and engineers sales reps over promising something and then there's some timeline that's going to be like the engineer is like well we're not going to make this well actually if the engineer is like
[00:53:59] A: massively amplified with coding agents that's more tractable you can actually do something create something custom very quickly and i don't think most organizations think of themselves as like that efficient so they kind of like they if you don't i think the thing is like if especially if you work in pure software
[00:54:17] A: You're screwed if you don't think like this.
[00:54:19] B: What do you think are going to be the biggest jumps in the next year for coding agents?
[00:54:23] B: What do you expect to be coming out and what do you think is coming down the pipe?
[00:54:27] A: I think one of the big themes is orchestration.
[00:54:30] A: One,
[00:54:30] A: our human ability to orchestrate many of these systems,
[00:54:33] A: but also the agent's ability to orchestrate subagents on complex problems.
[00:54:38] A: Because then...
[00:54:40] A: I think then the larger context problems become more tractable.
[00:54:45] A: It's not just like,
[00:54:46] A: oh,
[00:54:46] A: it's good at writing a script.
[00:54:47] A: It's good at writing like building my website.
[00:54:50] A: That's actually really easy.
[00:54:51] A: But like asking it to build something as complex as, say,
[00:54:56] A: a social network or, you know, just things,
[00:55:00] A: problems that are much bigger than a single context window.
[00:55:03] A: Like.
[00:55:04] A: problems where the the problem itself and like the solution is much larger than a single agent single context window that stuff is now within scope of like surface area that we're going to be attacking so people so i mean small teams will be able to very quickly build much more complex uh systems very quickly and and the speed is another piece so one the model is getting smarter and the context window increasing the other piece is
[00:55:31] A: the tokens per minute like the models the inference speed is going to go up like i think okay people complain about how codex thinks a lot i'm like if it thought at very high speed then you wouldn't feel this like oh it thinks for 20 minutes then it starts working but if it could do that 20 minutes like if the inference speed goes up by an order of magnitude then that thinking time goes down but then it'll be able to think even more and do a more complex thing in the same amount of time if we start getting these
[00:55:57] A: Getting these coding agents that can just rip at like.
[00:56:00] A: thousands of tokens per second and they're very smart then what you ask for the turnaround time closes and you get you can ask for a more complex thing and expect it back more quickly mostly this all of these like speed advantages and like the parallelization all that is most apparent in pure software problem spaces so anything that's a pure software problem or anything where the entire solution can be done behind a computer that's where you experience like this you
[00:56:27] A: hyper speed hyper competitiveness across like there's just going to be people with agents competing with people with agents and anyone that doesn't have agents is basically not even competing like that's the like you can basically eat those guys the the people who use agents will eat the people that don't use agents in the problem space that's like hyper speed
[00:56:45] A: hyper like cyberspace but then then it creates another opportunity which is the real world where everything is slower where everything requires people and moving of things in the physical world that's where the I think the the slower adoption comes like where people are
[00:56:59] A: process or like it's a problem that really requires moving something in the real world.
[00:57:06] A: And I think that people who don't want to compete with the speed will be able to bring some of those advantages,
[00:57:11] A: but then like focus more on the human relationships and the human problems,
[00:57:15] A: the physical world problems end up being a more manageable speed for people who don't want to, because this is going to get crazy and it's going to be highly automated and not everyone is like an F1 driver.
[00:57:26] B: So tell me more about that. Where do you expect coding engines to be in this, say, five years?
[00:57:30] B: Like if we're thinking like really far out.
[00:57:33] B: Is it just going to one-shot totally insane full production systems?
[00:57:38] B: Should these big public companies like Workday or Salesforce be as afraid as people have kind of thought this week in terms of the stock market?
[00:57:48] A: they should i mean that's exactly why i mean i think that that's why the stock market is experiencing this but i've been thinking for a long time that like i mean even in my one-on-one interaction with these systems i'm like man i'm the bottleneck like i it's like me not knowing about
[00:58:02] A: t-max is why i discovered it only six months ago like me not knowing and like the agent had this in his pre-training right so i'm like wait so i'm being overly specific about what i think the solution is i'm saying no no you need to take a step back and like recognize the solution might require many things that you don't know like the last 10 years of software was like interface lock-in
[00:58:21] A: and then you talk to an agent that can that hits the API the interface doesn't even matter I'm right like I can skip the interface and then the AI can the AI can basically like render an interface that's uniquely relevant to me I notice this with like I prefer the CLI agents because they can just hack my own computer to work for me instead of me tweeting at someone at OpenAI to add a feature and I'm like whoa I could just tell the AI to add the feature and then we'll close the loop right there right
[00:58:47] A: and so that's the other thing is that buy versus build you'll be able to build things more quickly than you can go buy them the problem is that like we're in the way of like the system that's actually now much more sophisticated without us like guiding and designing the next system is more important right now than like for us to be in the weeds of the process ourselves unless it's like a personal like something that uniquely requires like a person right like sales i think does need people i think
[00:59:13] A: if you play this like high-speed deploy agents at scale to solve the software problems then you benefit from actually hiring a bunch of sales reps to like take the market right so there might be a new play style a new way a place of allocating the human uh resources this sounds so dystopian by the way but i don't know i see it like my my mom um like her whole team has basically been laid off for the last couple years and she's been working at this company as a software engineer for my entire life and she's like
[00:59:42] A: well maybe then i just like i probably i gotta learn ai like what what books do you tell me should i read i was like there's no book you literally just ask me and that's our best bet is for us to just start hacking stuff and like there's no there's no course there's no nothing right like and it's if we waited for someone to write the book it would already be outdated by the time the book's published
[00:59:59] A: So how do you recommend people like your mom or folks get kind of red-pilled?
[01:00:05] B: Use this for something that you're passionate about.
[01:00:08] B: something that you wish you had more time for because then um you know no one needs to tell you the default is use it for work and that's like a weird thing because like your boss tells you oh you got to use ai now like gotta use ai the board's like oh you gotta use ai and i'm kind of just like the problem there is that like your intrinsic motivation might not be work and also like your company might not make it let's be honest don't tie your entire identity up with like your job we are multifaceted like
[01:00:35] B: like maybe use it to like for my mom it's like she's using it now she's using the multimodal models to help visualize her art projects and like learn how to like learn different styles of like sketching and watercolor and stuff like that but the multimodal models are basically like i mean the image models are so useful for things not just generating cool stuff but like generating
[01:00:57] B: interstitial artifacts that allow you to become even more creative like visual pre-visualization workflows animation stuff so there's a whole new thing that you can do but you would miss if you only used it for your job and also like you're not most people don't love their job anyways so if you don't love it you're not gonna get good at it so if you tie it to something you love you're gonna go faster you're gonna enjoy it no one needs to tell you to do it so you're intrinsically motivated
[01:01:24] B: that's where like you figure it out because you it's for you it's better to start with things that are like or like use it for like learning a topic or like planning a trip for your family there's so many use cases of AI that are not work related that you can experience the upside and then get up to speed and if you really want to like go deep on this like hyper engineering this like different thing do it because like but you have to be ready to roll up your sleeves and like
[01:01:52] B: um you have to deprogram yourself because it's the ai native people of the future are going to be like they they won't have this like baggage that you have and also like your company is just like filled with people that don't even believe or like they're banning the tech or like that's what's holding people back is like other people and you kind of have to just aim it at your person where it's like you're you're not even approved to use cloud code or codecs imagine it takes four months to approve codecs at a company at a massive company
[01:02:18] B: And like the person who spent four months in codecs is like, they could never imagine working for that company.
[01:02:24] B: So you got to use it in your personal life.
[01:02:26] B: Otherwise, like you're letting the corporation cut your learning rate.
[01:02:30] B: And the most important thing is your learning rate.
[01:02:32] A: What's the most non-consensus view you have on the future of AI and coding agents?
[01:02:36] B: AI is a power user technology.
[01:02:40] B: Yes, it's generally a superpower and everyone benefits from it.
[01:02:44] B: And it lowers the barrier to do almost anything.
[01:02:48] B: But the people who think really deeply about how to use it, the metacognition, like thinking about how to think, thinking about how to apply intelligence,
[01:02:57] B: that is a compounding advantage.
[01:02:59] A: Notice that only goes to the people that are power users.
[01:03:03] A: Yes,
[01:03:04] A: everyone will benefit on an everyday level.
[01:03:07] A: AI will start helping them. But like thinking about how to use AI is on a meta level,
[01:03:12] A: a very rare skill.
[01:03:14] A: And to hone that skill is to apply intelligence. And the application of intelligence is a
[01:03:21] A: I mean, it's where I spend all my time and I'm still only at 1% of like seeing the iceberg.
[01:03:26] A: Most of it's actually underneath.
[01:03:28] A: And for me, the thing that's holding me back is the need to sleep.
[01:03:31] A: That's why I have like so much caffeine,
[01:03:33] A: but that's also not sustainable.
[01:03:35] A: So it's a very,
[01:03:36] A: I think it's a power user technology.
[01:03:39] A: Like you have to be a little obsessed to get the most out of it.
[01:03:43] A: it's kind of like i think of it as like video games the gap between the top one percent player in starcraft or league of legends the way they think about the game the speed at which they move through the computer is like you can't you like they'll never drop some drop a game to someone that's in the 98th percentile like the top 200 in the discipline are on a different they're in a different game than everyone else and while ai is going to be designed for everyone else
[01:04:08] A: the top 1% experience of wielding it is an emerge it's a it's still unfold it's all just unfolding and there's no there's no you don't wait for the textbook they're just kind of like people talking to each other using the tools to see what they can do and then pushing the the frontier and I think that's just going to be a small group
[01:04:28] B: Absolutely.
[01:04:30] C: If you want to put a bat signal out to people,
[01:04:32] C: anything that you're thinking about in terms of projects that you're working on or just experts that you would want to meet, like I guess in summary,
[01:04:40] C: as we finish out the podcast,
[01:04:43] C: who would that be?
[01:04:45] C: What are the things that you want to learn more about?
[01:04:47] A: I want to meet more people that are experiencing like the frontier capability,
[01:04:52] A: just playing with the models,
[01:04:53] A: not just like in a work sense,
[01:04:55] A: but like pushing them creatively,
[01:04:56] A: pushing them like and like the meta game around like orchestration, thinking about how many how many agents are you able to wield across how many different types of contacts?
[01:05:06] A: Think about the multi agent coordination problem,
[01:05:08] A: getting them to work meaningfully on complex stuff.
[01:05:11] A: stuff.
[01:05:12] A: And if you find something out, you know, you can find me at my website,
[01:05:15] A: parth.club.
[01:05:17] A: and all my socials are there and so like and i engage i try to engage on socials as well so i'm just trying to meet people that are also just like oh yeah i guess like if you're familiar with cyberpunk like i'm trying to meet netrunners um these this this new i think that's what this is it's like i wouldn't call it ai engineering i'd call this netrunning which is like this like what does the f1 driver of vibe coding look like that kind of person i'm very interested in or if you you're aspiring to be someone that's closer to that
[01:05:43] A: I kind of want to meet those people.
[01:05:44] C: Awesome. Thanks so much for joining us today.
[01:05:46] A: Thanks, man.
[01:05:47] A: It's been great.
[01:05:50] C: Hey,
[01:05:51] C: this has been Kez Noka, co-founder of Village Global.
[01:05:53] C: Thanks so much for tuning into the Village Global podcast where we go deep on all of the biggest topics in tech.
[01:05:58] C: If you enjoyed this conversation,
[01:05:59] A: Subscribe to our YouTube channel. You can check us out on Spotify,
[01:06:02] A: Apple,
[01:06:03] A: wherever you get your podcasts.
[01:06:04] A: We'd love to see you for the next one.