Ep 02 - “I’ve Never Seen a Model Say ‘This File Is Getting Too Big'

[00:00:00]Welcome to the next token. Today is Friday, July 17th. I'm Dylan Mroy and this is a podcast about three developers who don't have the time or energy to be hosting a podcast.

[00:00:11]>> Hey, I'm Ree Sullivan. I'm working on my startup Executor.

[00:00:15]>> Uh, hi, I'm Sunil Pi. I work at Cloudflare. Uh, so does Dylan and they sponsored the show for us.

[00:00:22]>> So, on today's show, uh, we need to start off by, uh, settling a bet. Then we're going to uh talk about the news and what is top of mind for uh the three of us over the past week or two since we last recorded. And then today is the first of probably many but at least three episodes focusing on each of our workflows with AI and aentic engineering and how we're doing this day-to-day at our companies and day jobs. Um and thinking about that. So to get started here, we need to uh settle a bet we made at AI engineer where we last recorded.

[00:00:58]Uh Sil, you want to explain to our audience?

[00:01:02]>> This is okay. So this is embarrassing because I consider myself a man of the world and I don't like making bets that uh that I lose. uh on while we were so we were recording right in the center of AIE Expo and Fable was still like the government was like it's still so uh dangerous and we're not going to let it out without uh uh controls and which is why I think even OpenAI were like yeah we can't really roll out 5.6 to anyone and Dylan you said that by the time the next episode rolls around we will have access to Fable as well as 5.6 six and I very confidently I was like well I'm going to brag >> uh I said well there's no chance because the US government can't find their hands uh uh out of their ars or whatever the phrase is uh and I bet $10 against that.

[00:01:57]Uh Ree, you decided to join me because I clearly seem like the most uh the smarter one of the lot and you said $10.

[00:02:05][laughter] >> I said it was going to be week or months.

[00:02:09]month.

[00:02:09]>> You said like months.

[00:02:10]>> That's a crazy I I either said weeks or months, but I was like, it's not it's going to be a while.

[00:02:15]>> And from the communications we've had from OpenAI on the 5.6 roll out, >> it's definitely >> It sounds like it's going to be like weeks. Uh and so >> sounds like I'm about to lose 10 bucks, 20 bucks.

[00:02:26]>> My timelines are bucks. July, late July coming back.

[00:02:30]>> Like I like it wasn't even about the money. I just enjoy being right. So we did we recorded the whole episode for like an hour and a half, two hours, we got off and within 10 minutes, Twitter was a place that Fable is back and 5.6 is rolling out.

[00:02:49]is >> so bad. Yeah, it was like literally we had like gotten up from the the podcast uh station which was like smack dab center of the expo hall like open AI's booth with like hundreds of people around it right beside us pull up our phones and sure enough uh Fable was released and then a few days later 5.6 was also released.

[00:03:10]>> You have to understand like my my my life is built around people like owing me favors. Okay. So, yeah.

[00:03:19]>> Well, uh I will uh anxiously be awaiting uh the Venmo transactions from both of you. So, I am up 20.

[00:03:26]>> I'm good for it. Just so we're clear.

[00:03:29]Like I'm in the UK, so this is like uh like uh paper. This money doesn't even make sense to me, but I'm good for it.

[00:03:35]Okay. Uh we're trying I we don't do Venmo here like uh but I'm just saying I'm good for it.

[00:03:41]>> Oh jeez.

[00:03:42]>> Yeah. Sorry.

[00:03:45]>> All right. in the mail. Checks in the mail.

[00:03:47]>> Perfect. Perfect. I'll I'll V them VM over request. Reese Rhys, you want to kick us off uh this week with what's top of mind for you over the since we last recorded? Yeah, I I wanted to talk [music] on OpenAI versus Anthropics comm's game because it's >> like they both had really bad communication uh 6 months agoish and OpenAI kind of they locked in a bit in terms of like I like the changes are so small in terms of they just kind of improved their communication game a little bit by being like hey this is what we're planning on doing. this is what we're doing actively and this is like what we're seeing. Um, and so I got some like uh tweets here where one was the first one of like how Anthropic handled the extension of rate limits on or the extension of fable access to on the max.

[00:04:42]>> Oh, right. Right. Right.

[00:04:44]>> So I like I don't know. I don't like I they can do whatever they want with a model. It more just like confuses me from a perspective of like if you're gonna have to make this decision anyways to extend it. Like this it seems like something that you could communicate upfront in terms of like like if I think about how OpenAI would have communicated this. It probably would have been something along the lines of like hey here's the new model. It's like limited capacity, but we want to see like how people are using it, what you can do with it, and depending on the capacity or usage we have, like we'll see see if we can extend it in the future. That kind of thing.

[00:05:21]>> Yeah. Just like >> um >> transparent almost like owning the reality of the situation they're in.

[00:05:28]>> Yeah. And the other thing is like I've I've seen rumors on Twitter of like them trying to do like a distilled Opus 5 or something like that. So like maybe that's another thing where like they have this other model that's coming out soon. Um >> do you think that OpenAI releasing 56 soul terra and Luna kind of forced Anthropic hand here to extend the limits? And in the same vein, I think a second question I have is like do you think this is coming from them having scarcity around compute and them genuinely struggling to be able to serve the inference for these models? Yeah, I guess that would be my my my two questions around around this topic because I I still to this day have not used Fable at all. uh I don't have an anthropic subscription and because uh Fable does not support zero data retention, we don't have access to it in Cloudflare and many enterprises don't because zero data retention is a a very important characteristic of really any uh third party service that enterprises use.

[00:06:40]>> Yeah, I mean I I assume that they're limited. I wish that I knew more about how the infant serving works because like I would have assumed that like the night before fear was taken away with a bunch of people hammering their max plans thinking that it like trying to get all of their usage used up cuz they're like well this is going away so I'm just going to like burn my tokens.

[00:07:00]>> Yeah, >> I like would have expected that. I would have expected to see more people reporting errors on Twitter of like oh it's at capacity or like like hitting those us like hitting uh capacity issues and there didn't seem to be any of those.

[00:07:14]>> I'm curious if like enterprises that are building products over API who are doing like way heavier load maybe have seen that in higher um in higher numbers.

[00:07:28]Actually that's a great question for everyone listening. If you work at a company that is building products around uh anthropic models uh especially Fable during this time period where they were kind of extending the limits for Fable post uh GPT56 roll out did you see availability issues error rates increase uh on your products and in your telemetry? Yeah, I'd be interested to know how people building AI products are also like like I I assume that you wouldn't just immediately swap to running everything through Fable in that like it's a good model, but the price jump the like price jump for output seems like if you if you didn't have it subsidized, it seemed weird to me. Let's see then the other ones. Oh, another one was like they did a research on July 15th. Uh Claude did um and that was also an interesting one to me where it was like like it it was just like a statement of fact. We've done a reset.

[00:08:28]It wasn't like a oh we like love you guys using the product like here's a like there was no fluff for it which is maybe appreciated but it seems like if they added fluff it would again be like a free kind of win just in terms of sentiment whereas like otherwise it just like it you have all these like little opportunities to like win like consumer favor I would say.

[00:08:50]>> Yeah. Yeah. And I feel like they don't take advantage of them. And but like the other thing is they've got like a pretty good product, you know, like and maybe that's they're just like in a position where they don't like have to care about that. Um >> I I feel like that is shrinking away and like I mean both of these companies are I think dealing with similar adoption stories that like both Cloudflare and Verscell for that matter dealt with. And I think both companies are very good at and is that uh and a lot of successful dev tool companies over the past 10-15 years is like target like the developers for their hobby projects outside of work consulting contracting and if you win the hearts and minds of them for that type of work.

[00:09:37]They're going to be your biggest advocates in the in you know the enterprise and companies where like the real money is. And right now, at least me personally, like I have like probably a borderline negative opinion of Anthropic despite them having been the company and their models that have, you know, sold me on Agentic Engineering being the future with Opus 45 and 46. um just the I don't like their stance or the way to your point Ree that they communicate that like it just feels like they know better than us and they don't care about all the feedback. It seems like that the developer community is blatantly like shouting about on Twitter and Twitter's a bubble. But like regardless right now like developers are and like coding use cases are the places where these companies are getting the majority of their re revenue from at least in my understanding and the the reporting that I've seen and like I am far more likely to go into Cloudflare and advocate for OpenAI models right now uh because they're just more transparent in the way they work. they seem like they're actually working towards better token economics and you know usage but I think both of these companies are in a really tough place um too which we'll talk about here in a little bit uh which is I want to get into talking about >> go ahead >> yeah so about like four years ago five maybe uh cloud the big problem that cloudflare was facing was people didn't know who they were uh like they were like oh you have some I thought you were a CDN company and the barely shows up in marketing.

[00:11:31]>> And the thing that Cloudflare did was well, we have all of these people, let them tweet about what they're working about online. Like we don't we don't even want to like curate the message.

[00:11:40]And what's happened now is Cloudflare doesn't shut the [ __ ] up on Twitter.

[00:11:44]Like they're just [laughter] non-stop all over it and every like not even like announcement weeks. Announcement weeks we drown out everyone for a week, right?

[00:11:52]>> So you can see the same thing happening in OpenAI. I don't know if it's connected to the TBPN uh acquisition, but there seems to be a direct line there. All the uh DX boys, the uh builders in residence and the Codex people are just talking about their stuff all I mean they just hired Pauline Narwas, one of the most uh beloved community leaders, right?

[00:12:16]>> Uh that doesn't seem like a mistake to me. I know for a fact Anthropic is very uh I don't want to say hush- hush but their employees aren't really encouraged to talk I mean like company secrets whatever it is they so there is a difference between Tibety saying hey I pushed the reset button versus the official claude uh uh uh Twitter account just saying oh like limits are reset and I think >> to be clear they're both trillion dollar companies and I don't want to s for any of them but obviously they're doing a a charm on the open side, right?

[00:12:50]>> Uh that maybe that's what maybe that's what we're seeing. Yeah.

[00:12:56]>> Yeah. It's just like it's it's weird.

[00:12:58]The the thing that's weird to me like you you like last point of like just like companies like there's not really a huge reason to be attached either. I don't think it would like impact them that much to just like slightly tweak their communication style. Uh >> I have a proposal. I have a proposal.

[00:13:14]So, by the way, like this is related to the last ad that Anthropic put down where they had houses burning and gravestones.

[00:13:20]>> It was [snorts] so weird.

[00:13:22]>> Such a bad idea. Okay, so here's my proposal. Okay, let them have the same messaging uh like Dario's words, even though they're kind of weird.

[00:13:31]>> Pay George Clooney and uh Sarah Ferguson $500 million a year to just say those words. Just look extremely attractive on screen and say that. I suspect uh it might help even a little bit and they could use all the help they could use have right now. Just don't make the houses burning videos. That's what I'm saying. You know, >> it is so crazy. And we should actually maybe at one point soon do an episode or have a segment about this, but like I recently was back home in I grew up in York, Pennsylvania, and I was around a bunch of people that are not in the tech community that are just, you know, typical. Well, I know Pennsylvania is not really the the rust belt or uh the Midwest, but it shares a lot of characterist characteristics like it.

[00:14:20]And like even my friends and family that I that I know are highly educated and stuff, man, they have such negative disposition about AI and they have very incorrect assumptions about how it works, the energy around it, all these things. And like these kind of ads, especially this one and the way Daario in particular talks about the future and these models is I think really really hurting the potential not only these companies but the the technology too and like honestly like improving our energy infrastructure.

[00:14:56]Do you ever notice how whenever like the AGI ASI people talk to you and they're like, "Here's all of the very detailed uh ways it can go wrong." And then you're like, "Well, what are the ways that what are >> you always leaves with ways to go wrong?"

[00:15:11]>> Always the ways it can go wrong. Then you're like, "Okay, what what am I going to do in the AGI world?" And you get kind of like ambiguous. [laughter] You have these like ambiguous answers that are like well you know it'll be a beautiful world.

[00:15:26]>> I actually before we move on I want I want to say >> got a hot take >> uh open invite to open AI anthropic space I guess I don't really use those models thinking machines uh to come on to the show and be specific about how the world gets better specifics. None of this woo woo like we will take over labor and you will uh uh like uh food and health and energy will be solved.

[00:15:52]No, like please come onto the show and give us 10 minutes of best case scenarios with as specific as you can get. I will >> I would love that.

[00:16:00]>> Uh I will pay $20 more for that. I don't mind losing money [laughter] for that.

[00:16:04]I'm just saying >> huge budget over here.

[00:16:06]>> Just saying. Okay. I would like just come on and like and uh like like like the three of us aren't well it doesn't really matter to us. Uh, and I assume there are way more high-profile podcasts or events they can go to. But I promise you I will be your biggest fan if you can come here and be specific about it.

[00:16:24]I'm like just just come on. It's fine.

[00:16:26]It'll be good.

[00:16:27]>> I would mention like in the short term I have like enjoy like I I'm very h excited these technologies exist. Yep.

[00:16:33]>> Like I I very much see the upside in them. was it the ability to like your ability to jump into like a new codebase now and like start being immediately like productive in it is like pretty incredible. They're not quite there yet, but it's more like hooking them up to things, but like >> the ability for it to go synthesize a bunch of information out of all of your like past emails or like >> Yep. the >> in fact uh for everyone listening I may have to uh step away here briefly because uh I was using uh Codeex's computer use uh Skiller MCP which is very very good and if you haven't used it please try it out cuz it is like remarkably amazing but I have a package showing up likely during uh the recording of this that I did not plan for uh I had asked Codeex to go research what the cost or process to order a new Apple remote for Apple TV was. And uh it actually went through the whole flow and because I was logged into one password and have my credit card and one password uh it did the full checkout and I got a text this morning that I have a package arriving uh during the time frame of this recording. So if I have to go sign for that uh apologize but yeah check out uh check out computer use. It's very good. I'm surprised we haven't seen more um like accessibility call outs about computer use. My my assumption is that it would be a big benefit there, but maybe there is like already good systems in place >> uh for that. But yeah, >> if OpenAI gets their teams to like collaborate, I assume the plan is to get chat GPT codeex to like take over the computer like they finally like have the technology but they need to get all 300 teams to point in the same direction to do it. I think that's what they want to do. It is remarkably good. I use it now for like a lot of testing. Uh like letting the agent test stuff using computer use and it is uh it's very good. I'm excited for it to be faster and more reliable. I almost want like a code mode API round computer use.

[00:18:37]>> I'm surprised this is this is where I'm at with my testing stuff where I actually don't want to use computer use for testing because you can't like codify goodex computer use into like end to end tests in your repo. And so you can more you can have like a in you can instead of having it use computer use you can just have it drive like playright and then test against whatever you use which is pretty nice.

[00:19:01]>> Sweet. Well we'll uh we'll save that for uh maybe our next episode for going through your workflow. But moving on to the next topic. The thing I wanted to talk about before we dive into my workflow uh and we still have to get through uh Sunil's hot take on uh conferences uh is I am really really excited about thinking machines and [music] inkling and American opensource openweight models. So, for anyone that doesn't know, Mera Marotti, hopefully I'm getting her name right, I believe she was the former CTO of OpenAI. She spun off and created her own company called Thinking Machines. And in the past week, they released uh their first model called Inkling. And it is openweight, open- source. And the write up and release around this was phenomenal. Their benchmarks were incredibly transparent. Um, I think they got a lot of praise across Twitter uh for the way they demonstrated the capabilities of uh of this model. And one thing that is insanely cool and distinctly unique about Inkling is they shipped tools or a harness. I'm not sure which.

[00:20:14]Um, I haven't I only read over the the blog briefly, but they shipped tooling to essentially RL and fine-tune the model in your own hands. And the demon or the the demo they give is in an open code session, they they load up Inkling and they run the, you know, slashpost train or whatever the command is, I can't recall, and say basically stop outputting words with the letter E in it. and it runs this whole skill and updates the weights and biases are, you know, I don't fully understand the full, you know, post-training process, but it updates itself similar to like something like pi, and then it showed an example of uh them talking back and forth with the model, and sure enough, the letter E doesn't show up anywhere. And like I think that's fascinating and insanely cool. Uh like I like thinking of use cases like I I use my an agent to like budget and like if I could fine-tune it to like nail down and be really really good at using the tools and the knowledge around how to do my B my budget, excuse me. Um I think that would be insane. Um, so I'm curious on your guys' take on this and like my the other part of this is there was a good tweet from Shiaan Zoo saying, "My bet is thinking machines will soon make more money than anthropic AI, not by winning the race to build one standardized frontier model by becoming the palunteer for deployed engineer for enterprise custom models, the playbook. One, release the best American openweight model. Two, drive widespread enterprise adoption. Three, charge the largest companies seven to nine figures to post train and run custom models behind their own firewall.

[00:22:02]The model rests on three bets. Large enterprises will increasingly demand their own models with their own data.

[00:22:08]And this is how they differentiate and win. Enterprises won't need just one model. They'll continuously need new models for different workflows, departments, and proprietary data sets.

[00:22:16]That creates extremely sticky recurring revenue. Three, auto research will make custom model development increasingly scalable. Tinker can become the interface enterprises use to post-train their models with thinking machines providing the expertise and infrastructure behind it. Forward deployed engineers infra everything huge contracts for eventually maybe everyone wants their own model and auto research and training inside Tinker which Tinker must be the tooling that they have to postrain this and auto research and training inside Tinker on top of thinking machines base model will make it happen. [snorts] Meanwhile, Henry Forstyled standardized models will make no margins. Open AAI and Enthropic will have their API margins squeezed by DeepSeek, GLM, Grock, Meta, etc. And their consumer scriptions are lost centers already. The fat margin will move to customization, proprietary data, post- training, eval deployment, and infrastructure. If this thesis is right, Think Machines isn't building just another frontier lab. It's building the highest value layer between frontier research and enterprise model ownership.

[00:23:16]Turns out the best b the best business model for enterprise is not to sell commodity API access. Sell them their own model. I'm extremely bullish on this approach. Mirati may be the most commercially savvy frontier lab leader.

[00:23:29]I have to admit it. And I actually completely agree with this tweet. I think that this is partially aligns with like Cloudflare's bet on workers AI and hosting and running open weight models.

[00:23:43]Um, I think that both Anthropic and Open AI have got themselves into financial positions that are going to be increasingly hard to deliver returns to their investors on and a model like this could really really disrupt the two big labs. So curious on uh how you guys felt about this.

[00:24:05]>> Uh I I have two thoughts. First of all, I don't know if you know I'm a Miraadi stand like from the beginning. Holy [ __ ] extremely accomplished, knows how to get work done like big time Muradi fan. The second is a week ago from today. So on the 10th of July they published their internal manifesto thing called the future worth building is human where they actually go over these details and it's a remarkable document for two reasons which is manifestos are usually very surface level very very solar punk future looking but this actually like goes into specifics into some of like even the writings that have influenced them like they talk they have like a hayek uh document written in like 1945 about like the use of knowledge in society. I highly recommend anyone uh who's watching this to find the link in uh uh the notes for this episode and read it. And they go into the details about this because they say we are pursuing these directions. We train strong models. We build tools that enable people to make AI their own. We develop interfaces and we publish research. Uh which is a remarkable thing for also a company that's raised billions of dollars to say that this is the direction that they're going in.

[00:25:17]It's uh um I'm I'm a fan. Uh and the third thing, which is just a bonus thing, is thinking machines would have been too long for the Twitter handle, which is why their Twitter handle is Thinky Machines, which is the cutest thing I've ever seen in my life. I I I love it.

[00:25:34]>> Ree, do you think this puts Anthropic and OpenAI kind of in the hot seat?

[00:25:38]>> I don't know. I feel like it maybe gets I I think I'm maybe a little bearish on it in that I feel like it maybe gets a bit lessened away. What do you mean by that?

[00:25:49]>> In that like there's you know how as the models get better you like can get rid of your like certain like guiding skills or whatnot like they stopped like doing certain antiatterns uh just as the nature of like putting more compute into them and uh scaling them up. Um, [clears throat] I feel like we've kind of seen like micro optimizations matter a bit less, but at the same time, I think there was some stuff from ramp that showed that they like were able to fine-tune fine-tune models and get uh like better cheaper results out of them. Um, so I think it's it's like cool to exist. I maybe just don't relate as much to like the um space that it exists in at the moment.

[00:26:37]Like for me, I feel like for the budgeting one, I would rather do like I would rather codify that as like a set of skills and instructions. Then I can like swap out the model as I need than like a fine-tuned model.

[00:26:49]>> Why not both if it's effectively pennies or cheap?

[00:26:54]>> Um, which I think is like the enterprise use case. Like so much of our workloads at Cloudflare that are not like individuals using coding agents like many of our automated business processes are running on openweight models.

[00:27:10]>> But do you need fine tuning for those?

[00:27:12]>> I think that we could achieve likely better results, increase efficiency, both better outcomes for our users of those products >> and better outcomes financially for us.

[00:27:25]um which I think matters as well.

[00:27:29]>> Yeah, it makes sense. I'm excited to see them.

[00:27:31]>> Do you think the best and this is just a completely random country. Okay, so the country of Kazakhstan, do you think a model will be able to from first principles derive the best way to run civic processes etc in Kazakhstan?

[00:27:46]because I don't I I suspect most of that knowledge is not openly available and I don't think they could be easily condensed into simple skills that you can fit in a context window. Do you see what like the whole fine-tuning thing is that you can chuck a whole bunch of data at it and get it tuned to your use case.

[00:28:05]So for example that I I I've been having a theory recently that you know like claw voice and openai voice you know not just m dashes but the x not y etc thing.

[00:28:15]>> Those are not problems that can be solved by just like skills right like and the other thing I was I don't know if we mentioned it in the last call which is I don't want to hire five different clauds for a team like uh Miami 50 was not made of five uh Mr. characters like they're like different.

[00:28:33]You want to like have like different models that can like interact with each others and they're not just the same model but with skills. So I there's something there and it's a shot worth taking. I guess that's what I'm saying.

[00:28:44]Like I would bet >> Oh yeah, I fully agree. It's like uh worth exploring. Um yeah, I'm interested to see how it works out.

[00:28:54]>> Sweet. Well, this is definitely something I want to keep kind of our finger on the pulse of, especially around Thinky machines and see where this goes because I I'm I'm personally very excited about it. Um, and I do think it does put Anthropic and Open NAI in uh a pretty in a much harder position given their financial positions and, you know, the returns they have to do, the revenue they have to make to satisfy, you know, data center buildout and uh obligations to companies like Oracle.

[00:29:24]Um, so we'll uh we'll keep a pulse on that over the the coming months before we the last topic to get into before we dive into my workflow and how I work with AI every day is uh Sunnil. You not only uh recently attended AI world fair with us, but you also uh attended a local first comp. So we were on the expo floor of AI engineer worlds swear 7,000 people 10 parallel tracks workshops uh taking over all of Muscone what have you and I did that I spent a week in New York where I went to the new CloudFare office. I went to uh Johns of Bleecker Street to get pizza. I love it. Uh I've heard it's a top five pizza of the world. And of course I bought the merch.

[00:30:13]I buy the merch everywhere. And uh then I went to Berlin. Beautiful city by the way. Especially if you want like it's made for cyclists. So you can have like two beers, get a bazone and just like cycle through the city. Amazing feeling.

[00:30:26]>> I don't recommend it. Oh, so sorry younger people. Oh, it's so bad. You should not get drunk and ride a cycle through a beautiful city. Don't do that.

[00:30:34]>> Um uh but local first conf is it was wonderful because I want to say it was like 300 maybe 350 people maybe less actually. I don't know.

[00:30:44]uh bunch of research nerds, people like from Inc and Switch, Jazz Tools, who are uh talking about data sovereignty, uh owning your stack and not having a big tech influence, which is funny because Cloudflare was sponsoring it and it did get a few people on Blue Sky mad about that. And >> I did I did catch that.

[00:31:04]>> I just like whatever you know I I I'll get mad about it on Twitter later. It's fine. Um but uh it did strike me about uh where change happens and how do I put this? I met Dan Ingol, one of the creators of object-oriented programming at local first conf like I don't think the organizers knew beforehand that he was coming and he was just like a dude in a t-shirt and short sitting around. like I saw his badge. I was like, "Wait, there can't be like two Dan Ingol in the world." Um, and it like a a conference like AI engineer World's Fair must exist. It's surely a sign that something is going right. And Sean Wang has built like this generational conference. Um, but I think if you want to be closer in the know of the people like doing the work, I've managed to interact with and I'm saying this is worth versus the 7,000 people. I did enjoy my time at local first conf and I met a lot more interesting people there non-stop over uh the 3 days of the conference. Uh, I don't really have a point here to make. I guess it was just it was just very strange to me that this happened. I guess uh it's kind of why our podcast exists in the first place.

[00:32:27]Like we are three practitioners and we are tired of hearing like hypemen and people selling things all the time.

[00:32:33]>> Uh which is why I want to see how Dylan like literally types on his keyboard, you know? Like that's that's the thing that like interests me. I'm I'm not interested in your AI agent framework. I want to see a dude like punch his keyboard and generate code.

[00:32:47]>> Well, I have bad news for you. I've been using a lot of dictation uh to drive agents. But this is actually um I think it's encouraging to me because I I think at least >> you and I talked a bit uh when we were sitting at the the pub after we recorded our last podcast, but I personally feel like I didn't get a ton out of AI World Fair. And I think that was a combination of one just like we've been going or I I feel pretty confident saying the three of us have put a lot of time and effort and care into figuring out how to use this tech.

[00:33:24]And I think maybe a a large portion of the talks at AI engineer were kind of maybe higher level kind of more hype driven or just stuff that like I haven't found to work thus far. So it's it it's encouraging to me to hear that like there is progress being pushed in the other niches of engineering like local first technology and that then you know the bar and the needle is still moving even if it's uh drowned out in my Twitter algorithm and I'm not currently seeing stuff like that. So I'm that makes me relieved to hear. Ree, any any thoughts before we uh look at how I produce slop on a daily basis?

[00:34:09]>> Uh now I'm Let's jump into the slop cannon.

[00:34:13][music] >> Okay, so transitioning to the segment that many of y'all have been waiting for and asking for. Uh this will be the first of at least three episodes or segments where uh Reese Sunil and I go over how we're working with AI on a day-to-day basis in real uh code bases with real, you know, users. Um, and I want to kind of start by talking about my tooling, kind of where I came from, where I'm at right now as far as tooling, and then I can walk through a example of a feature I shipped to um, Herder, which we'll talk about here in a second in the past week. And also coincidentally how my workflow uh since I recorded or not recorded but ran through implementing this the first time till today has already started to change a little bit. Um so I have been living in the terminal for basically the ba the past 14 years. I've been using uh T-Mux since 2014. I abandoned VS Code sometime around 2019 2020 and jumped into the Neovim world. Uh and that was mostly because of actually RSI pain. So uh my my workflow has been largely geared around the terminal and T-Max for the better part of my career. And up until Opus 45, I was not very AIDS. I was using cloud code a little bit. Then like many people I used 45 and realized very quickly that uh the technology changed and agentic engineering was going to uh be the future of where this was all going. And you know I had a existential dilemma over the holidays I think like many many people had had some like restless nights about figuring out like what what does my job mean? I've been programming since I was 11 years old. uh reading code articles daily um like what does it mean for me to be a software engineer when being a software engineer uh is so much of my identity for better or worse and um you know took a lot of time to get where I'm at today and I think that's probably one thing that's not talked about enough is that this is like a completely new skill set like I don't think you can go from the workflows we were doing in the past to agentic engineering and getting good results off the rip. I I don't think one the models are there. I don't think the toolings are there or the tooling is there and it's still incredibly hard.

[00:36:59]And the thing that you know I think we've talked about a little bit but like it is exhausting to work in this way with this technology despite it actually sometimes annoyingly uh making me genuinely more productive. So bearing in mind that I've been in the terminal forever, I am I've had, you know, a pull towards using the terminal based agent harnesses. So I started with cloud code in October. Uh and last year I was using claude, you know, on the side like I mentioned, but then really went full-time into cloud code through the holidays and into like, you know, January, February. And then I had switched to using Open Code for a month or two and then I found Pi. So that's probably the first important thing uh to in my current tooling is PI. That is the agent harness I use every day. And it is uh for anyone that doesn't know PI is made by our good friend uh Mario, friend of the pod uh at Arendelle. And it is advertised as a one shitty coding agent, but more importantly, uh it is advertised as a uh minimalistic coding agent. And so like what does that really mean in practice? Um yeah, here's the the the copy. Here's the website. We'll link it in the the show notes, but Pi, there are many agent harnesses, but this one is yours. Pi is a minimal agent harness. adapt PI to your workflows, not the other way around. So, PI has a incredibly minimal system prompt. I mean, it is like a paragraph or two, a couple outlines of the tools it has and that is it. Like, it is probably, you know, under certainly under 2K tokens, probably under 1K. Um, and when I first used Pi, this is all based on Vibes, but I think there are there are benchmarks out there uh that show uh similar results, but like I could feel immediately a difference between using PI and something like Open Code, which uh at the time, I don't know if this is still the case with open code, but I believe Open Code at the time was based on your model, it was trying to emulate the system prompt of like, you know, codeex or quad code when you switched models.

[00:39:25]This is just static. There's like four tools, you know, read, write, bash. It doesn't even have web search or web fetch tools. And you could feel off the rip that like to me it it felt more focused and more likely to do what I was asking with better results even like with models like Opus 45 and or like Kimmy K2. Um, and I I I attribute that to the minimal context that's going into it. And importantly, the team behind PI does not make updates to the harness very often. And I find that very important in a time where these models are very dynamic and stochastic and uh non-deterministic.

[00:40:13]I don't want the other parts of my tooling regularly changing behavior underneath of me because that makes it harder for me to find a stable workflow with consistent results. Uh, and PI really I think has helped me get to a place where I can get a stable workflow.

[00:40:31]The other remarkably cool thing about Pi which you've uh I believe that AMP has now cribed and Open Code is also working on I think for their 2.0 release is PI out of the box knows how to modify itself. So it has one of the most beautiful plug-in systems I've ever seen. So you can basically load up a PI session and say, "Hey, I I I want you to build a MCP tool and it will go read its docs. It has very clear instructions on its APIs, where its source code is, and how to basically extend itself." So, it's super easy to model PI around your workflow. Uh, one cool thing uh that I saw recently was from Ben Vinegar from Modem. He was talking about how he built he shaped a custom harness with Pi around his workflow for editing his podcast. Like he basically built extensions, tooling, you know, whatever it is with Pi. So like he could load up into it and it was specifically shaped for his workflow around doing that particular task. So that that's one reason why I love Pi. It's great. Um I have a handful of extensions. Um I don't use very many admittedly. there is like an extension market that you can go to and install stuff. But my would my advice to anyone trying Pi would be use it out of the box and don't add extensions until you start to feel pain. And then when you do feel pain, take a look at the extensions that are publicly available and decide if that fits you well. And if it does, install it. If it doesn't, use them as inspiration and then tell Pi to build something like it that is more equipped to you because it just works.

[00:42:20]So the extensions I use is I have a custom gateway that I use at work uh that we use through uh Cloudflare gateway. I have a PI cloak extension which basically just masks environment variables if uh they somehow get into the session. a pi skill toggle which basically turns skills on and off for uh letting the agent invoke skills or making me invoke skills. I have an extension for saving markdown. So I can do /md and that will save the last uh agent message to a markdown file. I have my web tools extension which is basically a port of open codes web search and web fetch tools. Um I have an answer tool which I don't ever use anymore. I can probably get rid of this.

[00:43:03]Uh if the agent compacts, which it really does, uh this just continues it.

[00:43:08]Uh a get git AI integration, a git interceptor, which I think is quite funny. Uh this basically will block the agent if it tries to use no verify because I use a lot of like linting uh rules with like oxlint or uh like as GP.

[00:43:25]Uh so if the agent tries to use no verify in a bash command, it uh spits out no verify is not allowed. Good get hooks exist for a reason. do not attempt to bypass them. Instead, fix the underlying issue that is causing the hook to fail or ask the user for help.

[00:43:39]Um, so just like very minimal like I don't have a lot of extra here. These are just some like kind of quality of life things that I added. Um, then I have an MCP tool I also built. So that is like the extent of my my PI agent and you can see here like this is all it is. U nothing super fancy. And then I also uh have been using uh Matt PCO's uh engineering skills quite a bit uh which we'll get into a bit.

[00:44:10]The other second important tool in my uh tool chain is Herder uh which this is what is kind of encapsulating or wrapping around my PI session here. That is what the sidebar on the side is. Um these different workspaces, these different agents running. So, Herder is a lot like uh T-Mux, but for the age of agents. And as I mentioned at the top of the show, I've been using T-Mo since 2014. I've been dragging my T-Max config along with me for a long time. And I was really skeptical of Herder when I first used it. I admittedly have been waiting for Mitchell Hashimoto to finish his conquest of getting rid of terminal multiplexers through uh Ghosty. Uh, but that's not here yet. But the next best thing is that herder is built on lib ghosty. So, this actually ended up solving a bunch of my pain points with T-Mux, which is things like uh an escape sequences, not supporting kitty image protocol. Like, there's just a lot of weird things that make T-Mux not optimal when working with like a modern uh terminal like Ghosty or Kitty or, you know, whatever else. I was able to basically convert all of my bindings from T-Mux to Herder in one shot. And now I have basically this working product. And some cool things about Herder is that it has agent tracking and like their support. So I could like come in here and be like review the code, right? And you'll see when I run that it kicks off. So I can see a status of like if it's blocked, if it's running, what's going on? um which I love. Uh which I didn't have with T-Mux. When I was running T-Max, I was like using sound notifications to like grab my attention back when agents finished and that was, you know, kind of chaotic.

[00:46:04]>> Um >> did did you convert the bindings or did uh >> Oh, yeah. I I literally just told GPT to uh set up Herder and copy all of my key bindings and settings over to Herder and it basically just like one-shot it. And I >> I think that that's a fun quick call out of like the cost of adopting new tools has never been >> Yeah.

[00:46:25]>> lower. And so like you really get to experiment with what makes the best workflow for you.

[00:46:29]>> Yeah. I I love it a ton. Um, and then other than that, uh, another important tool is, uh, Planotator, which is a tool that launches a local web app that lets you basically review files. Uh, it lets you do local code review, and it lets you kind of like annotate files, code review, or even like the previous agent message and send those results right back into the coding agent. We'll dive into that here in a couple minutes. I'll give like a real example of how that works. But that's been a really important tool. And then I find myself very rarely diving into my code editor, which is Neovven, uh, anymore, which is quite sad. But, uh, from time to time when I do need to do manual editing, I'm using Neoven. So, that's basically like my full tool stack at the moment.

[00:47:25]So to give an example of something I built in my actual workflow here, I started working on a feature for Herder.

[00:47:35]So when you create a new tab in her, I can come in here, I can create a new tab. And when I create a new tab, it you know, it will ask me to name the tab. So if I do like fooar, it's going to name it fooar up there, right? So I can also come in here and create a new workspace.

[00:47:52]However, when I create a new workspace, it does not let me give me the option to rename it. It just kind of adopts the current directory and then I have to do this second step where I have to, you know, rename it over here. Now, fubarbass, right? So, I wanted to basically add the feature where when I create a new workspace, I can get that same modal to name it before it opens.

[00:48:14]Let me uh grab the session here.

[00:48:20]If we do pi dash resume right here. Oh, that's still not the right one. pi dash resume.

[00:48:33]So, a, uh, one thing you'll see in here is pi has this feature called tree, which is probably like the most sticky feature of pi for me. Uh, beyond the things I've already talked about, tree lets you jump around any point of the conversation. Uh, so like you can go down some path. Let's say I want to write unit tests for a feature. I can like go build that out and then I can jump back to the root of the conversation and like you can choose when you jump to another point in the conversation, you can choose to summarize it. Summarize it with a custom prompt, no summary or just jump back and forth. This is almost like doing manual sub aents uh which I'll demonstrate here in a second, but like I literally cannot live without this feature in a uh coding agent uh anymore. It's probably probably the most important feature feature of Pi to me.

[00:49:22]>> Does slash tree come with PI or did you add it with an extension?

[00:49:26]>> Slashtree is built natively into PI.

[00:49:32]All right. So where I started here was I started by and I forget the the exact name of this strategy but I will often start work with an agent by asking it questions that I either know the answer to or I myself need the answer to. But I I'll start with asking questions, right? So I asked how does the feature work for herder where when creating a new tab in a space a user can be prompted to name it first. What is the code path? How's the abstraction designed? How's the config value to enable/d disable that feature read and parsed and passed to the create tabflow/feature.

[00:50:13]So if we jump through the conversation here we can see that uh it ended up coming out here with this basically overview of how this feature works, right? And in fact, this is a big blob of text that is really hard to just like read in the terminal. So, this is a great uh example of like a place where I will use planitator. So, I'll do planetator last. And what this will do is it's going to launch a web app where it will load the last agents message into a web app right here. And I can go through and read it. And I can, you know, it just makes it easier to digest.

[00:50:49]And on top of that, I can come in here and I can be like, you know, uh, are you sure about this? Uh, can you be more clear?

[00:50:58]Right? Something like this. Right? This is just a throwaway example. And I can save that. And it basically aggregates these comments as I go through. Um, and then I can, if I were to click send feedback, which I'll show in a second, this would send all of this back into my session and let the answer or let the agent respond. So here my intent was to basically have it go off and do research almost like a sub agent would, right?

[00:51:24]Like I'm being my own sub agent. My I have a hot take that I think uh sub aents autonomous sub aents in coding harnesses have a tendency to take thinking away from the developer uh too easily. So I I like kind of doing this this thing on my own and I I don't want to defer the prompts and the context gathering to another agent orchestrating a sub agent like I want to be in control of that. So you know we can click send feedback it will send it and you can see here uh it sent the message annotation message feedback and it would go address that. So I will oftent times like go back and forth with agents planning work this way, asking questions, gathering context. Um but in this case I was pretty happy with uh the outcome of this. I felt you know this feels like enough information about how this feature works. So then I ended up doing tree and I jumped back to zero context and I uh started a new thread in the same session. Right? I'm mal back to 0% context and then I was like how does the create new workspace feature abstractions code path configuration types and data paths work for herder and I let that run off and eventually it resolved um to let me find my labels here create workspace research. You can also label anywhere in the conversation and jump around them. So here it came out with the same thing and we can go look at this implantator again and it outlines how the current create workspace feature is which is the thing I want to change right so here's this nice very succinct uh compact way that this works so again I was happy with that outcome and now I jumped back to uh I jumped back to 0% contacts and Actually here uh I used the save markdown skill and I wrote this last message to a markdown file which you can see is right here. This is exactly what we just looked like. Saved it to a markdown file.

[00:53:34]So I jump back and then we jumped to here and I say uh let's open this up in newim.

[00:53:44]You can see I rarely use my editor because I have a bunch of updates. Um, I ask read at create workspace MD, which is the file I just made from that last branch to learn how workspace creation works in her. Here is how tab creation plus renaming works. And I just copy and pasted that first branch where we got the summary of how the tab rename works.

[00:54:06]I copy and pasted that into just like a details tab. And then down at the bottom, I said uh I closed the details, right? I I don't know. This is all vibes putting it inside those like tags but just trying to like distinguish the different signal there. I said how would you add design and implement the feature capability to herder for introducing a similar feature to tabs with a new configuration option so that when you create a new workspace you're given the option to name the workspace at creation time. Walk me through the code paths data structures you would create update or change the types you would add change delete anything relevant like that. I really want you to focus on code paths, types, abstractions, and data flow.

[00:54:49]I do this thing when I'm planning with agents where rather than having these kind of highlevel uh almost like PRD product docs, I like my specs looking more like pseudo code and I focus on having the agent outline types. I basically do type- driven development uh with the agent. So my my specs end up looking like an outline of code and I also do this thing that's gotten some attention on Twitter where I have it like outline the call stack uh of like what the agent plans what the call stack is today uh and how it wants to change it and at each level of the call stack I ask it to say what is the input type what is the output type what are any side effects that can occur and what are any errors that can occur and I make it just outline that and that I I find at least on a vibe basis that that really helps keep the agents on track and stay aligned with like my expectation of uh like abstractions.

[00:55:50]So I ran that off and it ended up coming out with uh the initial spec and you can see here we have this document and I have that loaded into neoim which in fact we should just look at in planitator right so let's say planitator and we can use annotate and here I can mark any file uh and I think this is yeah create a name space spec.

[00:56:17]So if we look at the spec in here, me close this. We have, you know, this outline of code paths focusing on types already. Um, you know, I'm not going to walk through this all line by line here, but you can see that there's like a real emphasis on code and types here rather than just like, you know, as a user, I want to do XYZ. And it also outlines uh I have it outlined like red green refactor test and doing like much inspired from Matt PCO skills like vertical slices like what is the narrowest part of this feature that you could implement that proves it end to end and then let's add to that over time.

[00:57:00]>> Where' that note on terminal control in the planning doc come from?

[00:57:06]Uh where was it?

[00:57:09]>> It mentions uh testing and terminal control. It's uh up a little bit. I'm curious if that's coming from agent MD or from you.

[00:57:18]>> I after it produced that spec and I read through it. I uh I have a skill for terminal control and I said add verification and proof with terminal control to our spec so that a human can verify the correctness of your implementation. So that was me after the first spec came out. Uh, Sunil, you you had a question.

[00:57:38]>> So, I the the idea of having it almost like generate types and then having it like fill in the blanks is very appealing.

[00:57:47]But the question is like why have you not is it because you feel strongly that the function boundary should be defined by you or are you saying that models are still not good enough at generating those boundaries or have you noticed just better like I I want to understand like why you do it this way it is very appealing by the way like I and the reason I'm asking is like does it scale later?

[00:58:10]>> I do it because one that's almost how I worked previous to AI. I always started at the type level and thought about types and interfaces first. And you know when I used to create a new JavaScript module or file like I tried to keep a single module centered around a single core type and then build functions and combinators around that type. And if I feel myself starting to operate over different types, that's probably a good sign that I need to break out to a new module. And I kind of compose those together. And I definitely do not think that agents are good enough or the models today are good enough at doing uh like abstraction and system design. uh they even with guidance and skills and everything they still quite frankly produce really bad slop. So this gives me the ability upfront to align on how the system's going to fit together and how data is going to flow through it. And like looking at a call stack I can like very clearly see how data will flow through it. And it because you know it's so much code is coming out and I mean I still read all the code in review but like this also helps cement my understanding of the abstractions and how data and code flows through through my services that I'm working on.

[00:59:43]>> I've never seen a model go oh this file is getting too big I need to refactor it. [laughter] >> Yeah exactly >> like they don't they don't feel the friction in the way that I really like this flow.

[00:59:55]This is great. They also do this thing where like they make tons of tiny abstractions like I can give them uh Matt PCO's skill and like deep modules for example which I think is very useful and I still use and we'll inject into the context for some of this but it will like make these very thin interfaces and create just like one-off adapters and it's just like I want you to look at the abstractions we currently have. I want you to look at the interfaces we already have. Does it fit anywhere? Can we adapt any of them? Is something wrong? Uh rather than just being so trigger-happy to make a new abstraction. So, I've been trying to like put resistance between introducing new abstractions with the agent, if that makes sense.

[01:00:43]The other thing I really love is this planotator is incredible and I regret not trying it even though you've told me like 10 times to do it because the UX model of just being able to like paint a like just brush strokes over areas and like comment on it and then pass it on is very different from this idea of the entire source code behind a chat window and the chat being the only thing that you interact with. How has every other agentic ADI ID not stolen this year?

[01:01:14]This is just absolutely remarkable UX. I love it.

[01:01:17]>> Codeex just in the past couple days introduced the ability to like annotate the previous agent message and send feedback in. So like [gasps] I actually like a lot of times I will not settle on the first output of a plan, right? I'll come in here and be like I don't like the name of this type. be like uh this should be this should be workspace uh name action instead and then I'll come down here and be like um you know this is wrong you know back and forth and then I'll have it update the spec and then I'll launch back in and just go back and forth using planetator like this because I can make these larger brush strokes to the full plan rather than being like having to scroll up and down in my terminal and like copy and paste pieces is and like aggregate that together in a tiny text box. It's it's so useful. Planetator has been without a doubt the most important tool I think to like getting good outcomes. I feel like it would also call out that this is probably not how the labs the like labs want you to be working with the models like uh like that's why like that's why their like interfaces aren't designed like this in that like they don't they I think that they probably want you to be deferring more to the models and like trying to solve these problems through like loops rather than >> you square that with codeex adopting the ability to anodc or annotate the previous agent message now >> that it's like nowhere near the same I think it's like a parallel feature. It's not like trying to solve s you know what I mean.

[01:02:55]>> Interesting.

[01:02:55]>> Like it's it's more for like highlighting a bit of text than it is um >> like planning and doing work.

[01:03:02]>> Yeah. I I will say this is maybe more for like cloth than it is for open a eye. I would say on like claude desktop.

[01:03:11]>> I haven't used claude or any of that stuff for so long that I I I can't uh comment on I don't I just don't know that what the state of the art is currently beyond pushing to loops and dynamic workflows and stuff.

[01:03:24]>> But also I I think that this is a better flow than like the loops and dynamic workflows. And it's it's so interesting how much are how how much the way how many different ways of working have evolved over the past three years.

[01:03:38]>> Yeah. Uh, like you have this, you have whatever the hell the bun rewrite was, which is like incredible.

[01:03:46]>> Oh, we've got dogs barking.

[01:03:49][laughter] That might be the delivery of uh >> of your Codex Apple TV remote.

[01:03:58]>> Yeah, I might have to step away for Yep.

[01:04:00]I'll be right back.

[01:04:03]>> New Apple remote courtesy of >> courtesy of Codex. Is this now an unboxing stream? Is that what we're doing?

[01:04:11]>> Um, >> see it. Oh, dude. This is chaotic.

[01:04:16]>> I need to set I want to set up codeex computer use to just order groceries for me automatically.

[01:04:20]>> Dude, I'm telling you, it autoered an Apple remote for me and I didn't even know it.

[01:04:25]>> Yeah, >> thank you. Codeex, >> I guess I could just give it its own credit card. I'm still stunned by the planotator UX because now I'm thinking about how it would generalize for like like regular life use cases. Like it doesn't have to be just for coding agents, right? I use it when I'm budgeting like it'll put out like I'll be like how can I rec reconcile like my MX and I'll like it outline like what it thinks how to square you know transactions to different budget categories and I'll just launch that into planetator and I'll correct it where it needs to and then it'll go do go through executor with the wab mcp generated uh by the open API spec and just do it. Yeah, it's fantastic. So I want to show you another feature of planetator which is its local code review capabilities.

[01:05:10]And this is something that I was like really hyperfocused on early in the year. Like I was tweeting about it a lot. I was generating prototypes. Um I I felt very early on that local code review was an issue. Like I did not want to push code to GitLab or GitHub and have it public yet. Like I didn't feel like that flow was good for reviewing code. I didn't want like co-workers to have to get notifications for pushes that weren't ready yet. So I was really like trying to find a solution and planetator was that. So once I have the spec finished, right, we have our spec file here. It has all the stuff outlined. It has the types, the code, blah blah blah. Um, validation, verification, tests, everything. So once I kick this off and it goes through implementing it, I will let it run, let it finish, let it do its first pass over to anything. I currently have been experimenting with a skill uh my own skill for doing code review that um I have a skill for my own coding standards. Basically, it's like a two-track. It actually uses Matt PCO's skill for kind of two track comparisons.

[01:06:19]One, it checks one sub agent will check against the spec. One sub agent will check against my coding standards and then do an initial pass. I've found that like I've spending more time than I want doing code review and constantly have to like nitpick the same things. So, I'm just trying I've been experimenting with how to get a good first pass over that.

[01:06:40]But once the code's done, you we can come in here and do something like planetator review. And this will launch up and if you diff it against origin main, you can come in here and kind of see basically all the changes. And then I'll come through here and this is where I this is how I still read all of the code, right? Like I'll come in here and I will read the diffs and let's say I don't like this like I do it you know X Y and Z add comment you know just step through file through file and this is all built on PR diffs so it's lovely diffing and you can basically just step through everything see it and then you can send feedback and that will feedback into the agent code review feedback and then it will go uh address it and fix it and then I basically just repeat this process process until I'm happy with it.

[01:07:31]So that that's basically like where my workflow has been at for the past 2 months. It gets pretty close to producing code that I would have written myself before AI and agentic engineering was here and I'm able to work on two or three features whether that's in work trees or several projects kind of at once and jump back and forth them while the agents are reviewing or writing code. It is way more exhausting I think than before, but I am more productive and that's how I'm getting good outcomes. Um, one thing I do want to talk about quick is I mentioned at the start of uh me talking about my workflow that I've started to experiment with Matt PCO skills a little bit more and in particular leaning into his kind of wayfinder workflow which if anyone's familiar with uh his grill me and grill me with docs which have been quite popular um those two skills basically help you align with the agent on like what you're actually trying to build or solve. it will like kind of relentlessly interview you and go down every path.

[01:08:35]And he built on top of that to build this thing called Wayfinder, which is for bigger work that might take multiple planning sessions. Um, and this will outline like a map of like either grilling sessions that need done to get the feature completed, prototypes, uh, or other tasks. And you can see an example of this. I started trying to build, uh, a new app here. And this is the map that was produced. Uh, and you can see all these different subt if you were just doing like grill me with docs or I was just trying to build my own text back up. Like this would have been a ton of work. Like this is like way too much for one tech spec or one passover things. Uh, some of these you'll see are labeled as research. Uh, some other ones are going to be grilling, which this was like an interview session with me and the agent.

[01:09:27]Some of these are prototyping.

[01:09:29]um or tasks. So it'll go and create throwaway implementations.

[01:09:34]Um you can see here it has like the proof of like I went and like implemented a potential like interface for this project. Like what do you think about it? And it'll eventually build up this text spec in the way that I showed before, but it goes through all of these places that there was ambiguity in my initial plan and it resolves all the ambiguity over these different sessions.

[01:09:57]and then I'll eventually spit out kind of issues for the agent to go implement.

[01:10:01]So, I'm playing with that. Uh I'll keep everyone updated on how I feel about that in the next coming weeks. But, uh that's kind of my my workflow right now.

[01:10:10]>> Cool. Thanks for sharing.

[01:10:12]>> Uh this is amazing because uh yeah, I keep thinking that you and I have like fairly similar output in terms of standards, but it is so completely different from how I do this. Thank you very much. I'm like quite grateful. I'm uh I'm excited to see how you guys work work as well. Well, thank you everyone for tuning in. That's it for the next token episode 2, aka the third episode.

[01:10:38]At some point, we're going to have to figure out uh this zero indexing problem we've got ourselves into. Uh we'll be back in 2 weeks with more opinions, predictions, workflows, and uh whatever Sunnil happens to be tilted at at the time. Follow the show wherever you're watching or listening and tell us what you think in the comments.

[01:10:55]Uh, one of the things that's been on my mind is I also want to talk about this industry-wide phenomenon of burnout that's happening. And if I may be vulnerable for a second, I'm actually deep in the throws of burnout. Like, I've actually gone for the last 2 weeks with approximately zero lines of code, which is weird for me because I was averaging about 50 to [music] 60,000 lines of code a week before that. Uh, and I'm actually so which is why I started 2 weeks of leave. you know, you see the ship burning and you immediately take some leave. So, uh I saw the I saw the Odyssey today. I did this podcast with my friends and I'm going to set up the Mac Mini that I bought for OpenClaw 6 months ago. Uh but I would love for us to like talk about that. I'm and I'm looking forward to seeing Reese's workflow. I'm looking forward to sharing my workflow, but I do want to talk about how we survive as an industry through whatever the [ __ ] is going on with everyone at the moment. But still so grateful, Dylan, for you showing us uh how you [music] do things.

[01:11:56]>> Well, go uh go get rested up and hopefully recover here. Uh I'm not far behind you on that front. I am I am straddling burnout myself and uh I think once we get artifacts to open beta here in the next couple weeks, I'm probably also going to take uh two weeks off uh just to kind of step away and let my brain get away from the AI mush. Um well, thank you everyone. We will catch you next time.