chapters transcript notes
click any line to jump to that moment in the video
0:13 All . Hello and welcome back to another Agentic Thinking podcast with Matias and Mike. Hello Matias. Welcome again. Yes. Hi Mike. How are you? I'm enjoying this twice a week conversation with you. If nothing else, 0:24 it's just helping me get out of my normal workflow. I'm pulling away. I'm I'm unpacking what I've been doing, what I've been learning. I have another really smart AI person 0:36 that I've been talking with. His name is Dan. A and he's started this habit with me on Fridays. He'll text me something. He'll say [clears throat] 0:46 this week I've learned and he tries to distill all of the things he's been work it's happening fast. He's trying to distill all the little nuggets and pieces and things that you've been 0:56 working on down to a single item. what is a two or three sentence? What did you learn this week? Could be could be remedial, but 1:06 reflect on your week and and and drop in what you're doing what did you learn? What are you working on? just something simple something simple around that around agents. And I'm finding it 1:17 extremely valuable because he's given me some really good nuggets that he's been learning and and pulling in. and he's really emphasized or re-emphasized my need to do better planning and more 1:28 plan mode with my agents to build out better requirements. You Matias as have also been really pushing me to do better at that additionally. anyways, all this to say just 1:38 these conversations are this to me. It's helping me reflect and go back over what we're doing. Fantastic. Yeah. You have any comments for any intros or 1:49 openers? I think you found some news here that we should unpack before our main topic today. not quite news I would say almost the opposite of news. , I've gone 1:59 into history a little bit recently, sort of figuring out, , how how did we how did we get to where we are today and I found a reference to a 2017 2:13 research paper, nine years old which is called attention is all you need. And we can we can share the link to it. and I'm in absolutely no 2:25 position to comment on this academically and not even going to try to. but it turns out if you look into things a little bit this one this 2:37 single paper apparently was really really pivotal in terms of enabling everything around AI that we rely on today. they introduced a concept 2:47 of transformers which effectively helps massive parallelization of GPUs and 3:01 ultimately leading to the the scale and performance that modern large language models rely on nowadays. you know everything up to that point was 3:13 quite sequential with respect to going through and understanding text and that didn't scale the hardware wouldn't have been able to keep up at all in 3:24 terms of the the the volumes we would need for a very very good large language model. it turns out that this paper specifically changed 3:34 the course of history you can almost say. attention is all you need. There you go. Almost sounds a cheesy song title but if if if anyone has some academic interest 3:47 definitely something you should check out. And I think this this was again very early days 3:57 the neural networks was this is when we're starting to really explore what neural networks are. Andre Karpathy I think is how you say his name. he was the gentleman who 4:08 was at I believe he was at Tesla or open AI in the early days with Elon building some AI things. moved over to Tesla, built a whole bunch of vision and things there. 4:19 Really deep diving on Transformers. , I think he was one of the there's a lot of people that regard him very highly on Twitter and on X I guess is X is appropriately the name 4:29 around what he's unpacking. And I every often someone shares a little video and says you should watch this instead of watching Netflix for two hours. You should watch this video from Andre and 4:40 see what he talks about for Transformers. And man, I was hanging with him for the first 45 minutes and then he went past my head. [clears throat] 4:50 what he was discussing and again this is this is the concept here is he was doing some very simple math and and what he was describing in this 5:00 video talk. It's a two-hour video talk that I and I got about halfway through it and I really got was really into it and then my ADHD kicked in. I did something else. I started going to something else. but he was 5:10 talking about when we do large language models, we send some text into it, it the weights of the model 5:21 just do things and we get the other side of the output. on the other side you get the answers that you're expecting. And what he was trying to do or he was 5:32 mathematically walking through the transformers and saying different transformers have different weights and when you adjust something on 5:42 the front side, how much of that upfront change in parameter or change in information in results in a 5:52 different waiting and how does that waiting push the output differently one one direction or another. And that's what he was unpacking was this very simple math he's doing 2 plus 2 equals 3 and and 6:02 doing a division symbol and he was saying look we know what the answer is but then we can then work our math back through. one of the things I think he was discussing the paper that was kind of interesting to me was we can watch 6:13 what's happening inside the neural network and start to understand how do the weights influence the outcomes and that's really what we're doing with these neural network things. 6:23 And it's really interesting. Again, I'm not a PhD person. I I can't go deep with these topics, but I find these immensely fascinating as to what people are 6:34 discovering these math that they're coming up with and it's too compute intensive for us to look at the whole system individually every single node 6:44 because there's just too many of them. But in general, we can generalize and look at what are the weights? How are weights being pushed on these models? really interesting stuff. These papers are phenomenal. There's people way smarter than I doing this 6:55 work and here I am just sending a text box to or or sending a text string to an agent and getting all this code out of it. It's it's phenomenal what this is doing this technology. what 7:06 really puzzles me, , totally from a position of ignorance obviously how when you when you chat with a large language model, , 7:16 when it doesn't produce code for you, but but text language, how it it only ever produces perfectly 7:26 correct, valid English grammar, assuming we're talking English, ? which is obviously not something a human would ever do, ? it as as far as I 7:37 can tell it never makes mistakes unless I just never spot them or . . it almost feels if you don't really know what's in the black box as if there is some 7:49 a grammarly a process that checks all outputs and and filters them before they get to you. 7:59 But apparently that's not the case. . I I I had some discussions with Chad TPT about that at some point because I was just really curious and apparently you always get perfect 8:12 grammar perfect outputs just because of the huge volume of training material those models have been built on and and because apparently 8:24 they've been selected that the models have only ever seen valid English sentences. It's absolutely mindboggling. Wow, that's interesting. Yeah, 8:35 definitely wasn't a study on what I've been talking about because my sentences are all messed up. [laughter] All don't make sense. , let's let's get into we talked a little bit about some some interesting things. 8:46 Let's let's transition the topic here. Let's go into move into a bit of our main topic here. And I think I want to unpack Matias. We were just talking about before this episode. 8:56 You're coming out of a deep thinking session, ? you you you and you do this quite often and we talked about this a little bit with the very episode one was the mind palace what does what does the mental model look for when I think things 9:07 agentically and I have been finding how I look at my work daytoday is greatly being reshaped 9:17 with the use of AI in what I do also what I think is possible the level of reach I'm looking to grasp is 9:29 much higher than I would have three months ago. I'm much more aggressive on taking on bigger projects, more specific code detailed things 9:41 knowing that I've got agentic support behind me. , let's unpack some of this. Let's rethink what work is looking . What are you describe for me before we got on this call, what were 9:52 you doing? What what was what was really boggling your mind? And what are you changing from how you used to do work to what you're doing with agents? Let's talk about that. Yeah, absolutely. I I think a lot is 10:03 going on here. and let let's let's talk about a myth first of all. you know, lots of people think or fear, you 10:13 know, that all our jobs are going to go away because AI is going to do everything, ? Yeah. My experience is quite the opposite to be fair. 10:24 same in I'm a hardcore coder I've spent most of my career writing code I have to say I have I have not 10:35 wrote a single line of code for many many months intentionally to some degree but also because I've just 10:48 I'm just using different tools and , I'm writing more pros, ? I'm I'm writing more prompts, inputs, 10:59 reviews, , requirements. , , that's what all my keyboard time is is spent on. and 11:14 it also turns out that we we are substantially more 11:24 capable in terms of what in conjunction with agents we can produce in terms of in terms of scale in terms of volume 11:36 in particular. in in terms of depth, , we we can we can write in in in in in 11:46 in quotes in languages that we've never known about, ? Yep. Agree. that's that. 11:57 but are we necessarily producing more let's say shippable things out out of that per se? Not necessarily. . we just 12:08 we've we've inherited new kinds of problems, ? , harvesting and and orchestrating and observing and 12:20 steering agents turns out to be a very non-trivial problem. And that turns out to be one where we're still very early in terms of getting the 12:31 processes and procedures and tools in place. And that's something you know we all have to do a lot of learning around 12:41 it it may seem you get you're getting a lot of stuff for free if you put a prompt into I don't know code or and 12:51 then it spits out some software that runs on your machine. , sure. Which is, , exciting, particularly when when you don't 13:01 necessarily have a have a coding background. , but, , if you want to do this seriously, , , in in in a in a 13:12 professional context, that's not really where things stop. If anything, that's where they start, ? And everything else that comes after that takes a lot of new thinking 13:23 around what engineering harness do you need to put around this agentic system. and again I would say we're very early on here. 13:33 I would agree with this statement and I think I'm I'm relating a lot to what you're describing. let me give you just let me give you a tangible example of something I'm physically working through . 13:43 often I get a lot of phone calls about, , wanting to do some support or some work. I have, thankfully, I made a decision when I started my business in 2019 to go 13:55 independent and do my own work. I I decided I can't stand working in Word anymore. I'm done. The formatting is just not helpful to me. I don't need it. In order for me to convey information 14:06 from my business to a customer, I need markdown or or that landing page of GitHub, ? If it has headers, it has bold, you can add images, most of the formatting pieces, 14:17 and I don't really care. The the formatting isn't Let me say it this way, Matis. If I gave you a Word document that was PDFed and I gave you a markdown file that was PDFed, 14:29 does the formatting sway you in any way to buy my services? Probably not. I probably I probably would argue if I put a picture in there and it has a 14:39 border around it or if it's centered or if it's less just you're probably less interested in the format of my document. You're probably more interested in the words on the page, the comments I'm 14:49 producing for you. thankfully I've been writing all of my re all of my statement of works in a single repo all in markdown. 14:59 Here come agents, ? over the last two weeks I have been rethinking my entire process around what do I need to do to distill 15:10 and get information and knowledge to my agent to help me rationally think through what is the statement of work? What do I want to emphasize? What kind of work do you need? What examples are 15:21 are you going to build? What are the primary outcomes for this this engagement? it has all of my old statement of works. The agent can see all of this stuff. I'm building 15:31 new skills. I'm enhancing this documentation. when you give me an email whatever the end of your domain name is, I'm giving the domain 15:41 name to my agent and say, "Hey, this is the new client. Go build me the basic information that I need. Go research them on the internet. Go find them. Go go find their website and 15:51 research everything about them. What do they do? How do they do it? And oh by the way, since agent I've described to you what my business does and you already have a summary, I've already had 16:01 to go through and summarize my other statement of works. what do we do based on what I've sold review against our services against what they think they might need? How could I find opportunities for them? 16:12 And I'm building a whole litany of skills around my very hyperspecific business and workflow that's not trivial. And I guess the 16:23 point here of this and the reason the example is I would just do the process because that would make that's what made sense to me. I have to step out of that process and think is that the way 16:34 to do the process how would I document it? What is the integration points between where I take my notes in loop where the teams meeting comes from? How 16:44 can I automate the transcription and the summary documents from teams into my agent? how do I get it there? there's all these other things that I'm trying to unpack and organize 16:56 and maybe that's an experience you're having. What are you finding? one thing which is quite related here which I thought about recently is 17:06 I'm finding it increasingly hard to select the the skills and tools when creating a new dedicated agent. , we've we've talked 17:18 about this presumably, , how the industry had a massive explosion of MCP servers last year and then end of last year, everyone switched to skills 17:29 as as the new kid on the block and skills were pushed as solving lots of problems related with MCP servers. 17:40 But if you look at plug-in marketplaces and skills repositories, it's insane. but particularly because they're easy to 17:51 make, ? It's just generally a natural language document. virtually anyone can can write a skill. what I find very very challenging 18:03 nowadays is how do you even how do you select given a particular problem you want to tackle how do you select high quality and useful skills 18:16 it's increasingly hard and hard it becomes harder and harder I don't really have a solution for it yet but it does not seem a great place to 18:28 be from from my point of view unless you restrict yourself really heavily and say I only use I don't know 18:39 three four or five particular sources slashmarketplaces and ignore everything else considered noise I think that's definitely where we need some 18:53 innovation and [snorts] let let me I this idea. May I propose a top an idea? Let me 19:03 propose a thought and maybe maybe you do this already. Maybe I maybe I don't maybe I'm just learning this as we go here as we think through this one. to your point, I also see the same reflection. Lots of 19:14 skills every other day in my feeds. There's another whole look at this new skills repo for Claude. Look at this new thing I built. I built 70 custom agents all in a repo. pick what you 19:25 want. Awesome. Co-pilot has a whole library of a bunch of things that just continue to grow and the community is getting big. Awesome. Love it. Super cool. 19:35 What I'm finding I'm doing is I'm building skills and as I work with the skills, the skills usually do a lot of what I want, but as I refine my process and the skill, we're 19:45 evolving together a little bit. the skill does a lot of what I want , it doesn't always get it 100% all the time. Or my process is shifting a little bit. I need to slightly tweak my skill. 19:57 Are you finding or are you building a process where you're doing more retroactive skill review after you use it and having it refine itself or 20:07 enhance based on your most recent run? Do you do this at all? I I'm me I'm doing this a little bit and finding value from it. I don't know if you're if 20:17 you're experimenting with this. I'm tackling it a bit differently. I I see this as a problem of volume and scale, 20:27 ? And when when when I have seemingly a thousand different skills available that all came to do the same thing, I as a human possibly cannot select 20:40 which one is best. Sure. But my thinking would then kick in to say this is per this is the stuff I can hand to an agent, ? I'm I'm increasingly investing 20:53 in second level agent infrastructures where I use agents to run experiments because they can do things in 21:05 parallel. they can reason very and all I need to do is to explain to them what the scenario is and what 21:15 success means in terms of output and then an agent can assuming I've got AI credits 21:25 available or or run it on my own hardware they can they they can just test out a hundred different versions of it and ultimately report back to me And that's that's 21:37 where I'm at . I this too. my my my natural response whenever I see a a a problem of scale is 21:47 a human cannot tackle that or build testing and validation and you know problem statement is this success 21:58 looks that. Here's here's here's 60 differential equations. solve the solve the one that has the the the the maximize function, ? You're you're 22:08 just doing a maximize function on lots of skills with defining a clearly defined goalpost. And and I think this is maybe the point 22:18 around rethinking your work, ? I think what you're describing here is exactly what I'm finding myself doing more often, which is rethinking what are the inputs and 22:31 rethinking what does success look ? And I got to be honest, even in software development, people are not good at defining what does success look for your software. What does a win look 22:41 ? it's hard writing requirements, documents, and tasks and features because you really need to be specific at the end of the game. 22:52 What does that look ? as you hand it back to your development team you know when we're pre- aent world I had to be very specific around I want this this and this because they're going to put the nav menu in the wrong area and 23:03 they're going to have to use the wrong colors and we're going to deviate from xyz things we have to have standards in place to really be specific about this sounds we're able to then to your 23:13 point problem of scale but let the agents build verify and what we have to do is shift our work to think about what's the goal what's the output 23:23 I've had a few really good successes I would say in that space recently. exactly around the 23:33 definition of what's success and and what is 23:43 a positive outcome. ? what what I find very challenging and hard today in in an agentic world is creating good test harnesses. , it 23:57 turns out, if you are able to define to not just define a PR and and requirements, but to give 24:07 the the agent who's implementing your PR, a very specific ways for them to use and run the product they're 24:17 producing, , the feature they're producing, the the software they're producing. Yes. If if you enable them to use that 24:28 it's incredible how much how how the quality of of what's coming out of it is shooting up. wi without that and this is the default place that you 24:40 get for many coding agents today. you may have a reviewer coding agent and it may seem to be 24:50 very thorough and it may find lots of issues to address. But normally those are hygiene things. Those are cosmetic 25:01 things. Those are obvious typing or compiler things. it only gets you far and and a 25:11 reviewer that's not running your software may pass something that if you then try to run it as as a human user you will immediately 25:23 find this thing doesn't do the thing it's not working yeah button makes sense but that that is definitely something which I've 25:34 experienced recently where I've been able to produce very longunning agent sessions 25:45 because the reviewer as it was able to run the software themselves was then able to give much more specific feedback 25:57 and you had longer loops. and you had absolute certainty that at the end of the loop the thing that was produced produced the outcome. 26:08 this is not trivial to define up front. ? This is this goes substantially further than you writing a PR because you need to think about what 26:19 execution environment for instance do you need to provide you really need to think about do you need to provide login or or or give a way for your agent to to 26:31 authenticate maybe on your behalf or to bypass it or something that it it this is where the the real hard engineering 26:43 problems sit nowadays I would say compared to where we were a couple years ago where our engineering problems were about how do I write this function or how do I design 26:55 this API in the best possible way we no longer have that problem but I really see it as as a test harness problem 27:06 and that is harder or easier depending on, , what software you're dealing with, ? If you have a CLI app, for instance, or or 27:16 a REST API, it's obviously substantially easier to test that than, , a mobile or a UI or or even a a whole 27:27 fabric solution, ? Which is one of the hardest things to test in in my opinion. but anyone who is worried that they may no longer 27:40 have a reason to be employed as an engineer, there are probably more opportunities than they have ever been. they're just going to be 27:50 somehow different, ? And I see a lot of opportunity around testing as I just explained but also around experimentation to figure out 28:03 what's the agent harness for me to use? What's the model? What what are the model settings? what's the system 28:13 prompt? It an agentic system has many inputs in addition to your prompt. it's in insane the 28:24 different it's it's an infinite number of possibilities how you can configure a system. if you want to be particularly efficient, if you want to 28:34 have particularly good outcomes, you have no way other than running experiments to figure out which one of those performs best. Th this is the same concept. I'm I'm going 28:44 to bring this really back to PowerBI and Fabric. I think you made a note here on how hard it is to test these things. I really do want to push on this note. , let's go pre-agent world and how I think 28:55 about some of these things. If I wanted to know something was more efficient than a different system. Let's say for I'm going to use a very simple example. I use this all the time on other podcasts and I've talked about 29:06 this a lot. If I'm going to load data using data flows gen 2 and I'm going to load data using a notebook, ? Two different systems on how I build things. One's more business user focused and one 29:17 is more notebook and code ccentric. . And when I look at those two, in order for me to really evaluate the performance of both of those, I've got 29:27 to think of a a test to your point, a test, where does the data come from? How do I load it in? Let's run it through the CUS on the pipeline. Let's have an isolation environment. , , I'm 29:37 going to run this data flow three or four times. See how long it takes to do the the loading. , let's do it again with a fabric notebook. How long did that take? What did we get? Are the 29:47 outputs the same? Yes or no? . , this is something we would have to do one manually. And then two, I would wait a lot for not wait, but I would be very intuitive when 29:59 Microsoft would release new features. For example, the shoot, I forget the name of it. It's in the Spark Notebooks where you turn on the the new Veilocks engine. I know the names of 30:11 the terms, the gluten and vex engine on top of the spark engine. It's they're it's they're equivalent to the data bricks hypers scaling or whatever they're doing on the on the SQL 30:22 server. Anyways, when they come out with that feature, someone needs to go do, hey, I was loading this table and it took me two hours. I turned these things on and 30:32 here's what it did differently. I would wait for someone else in the community or observe. Mim was one of the ones I to follow from Australia. I think I think maybe he's from New 30:42 Zealand or Australia, that region of the world. , but he would write all these articles about it , "Hey, did you guys know about duck DB? You should check it out." it's super fast. That's all and he's doing all these 30:52 benchmark things. And I was eating that up because it was he was doing the research that the community needed to figure out what was efficient or not. And let me extrapolate a little bit 31:02 forward . . the new feature landed in fabric. We needed to have multiple decision paths on what we could build. This is the same 31:12 thing for what I see for data flows, copy job, mirroring. There's all these other features of how you can move data from the in around in fabric. I want to evaluate how long it takes and 31:22 how much to it cost me. with agents to your point here with fabric, it's hard. We can define here's the starting point, here's the ending point, here's 31:34 the different potential routes you can get to to get from A to B. You can define some of that. And you can just say agents, all , you go build seven different workspaces and go build 31:45 each architecture and run them three times, then come back automatically and tell me the results. That's , this is 31:56 that's what I want, ? I want to be able to do that and and have verification my process, my table, my SQL database, this is the 32:06 most efficient way to run this process in fabric, ? With agents, we can build at scale to do more of these testing type things. Yep. And ultimately, that's where you 32:18 got to be, ? Yeah. But this is your point. Your point is that to build systems that do that, that's where the challenge is . You can't you're not going to be able to 32:28 go in and step into an agent be , ", agent, test six things, go do it." It's not going to know all the nuances of deploying and fabric. And we we have to build tools. And I'm 32:40 going to point out today Alex Powers has agentic task flow development, ? he just built an agent that helps him use a large language model and helps you build the 32:51 architecture and has a whole bunch of he's got thousands of lines, dare I say, of input information to the agent to help it just build an architecture point 33:03 blank. That's not even doing the testing. That's not running data through it. That's not doing eval. That's just making the infrastructure. It takes thought. You're going to have to think through these things. this is I really love this point that 33:14 you're defining here is the testing is really important for us. and again I think our our work is shifting I'm not writing the Python scripts anymore. I'm going to agents and defining what I 33:24 want it to build and then say you figure out the best way to build this thing. and and let's let's define the result of what I expect. What do what's your thoughts? What are 33:34 your reactions? Yeah, absolutely. this is what people should be getting into. those are the hard but really 33:46 real and valid problems. non-trivial it takes a lot of effort and time and and persistence to to become 33:56 good at this. another example is last week I think we talked a lot about how the freelance is over and how 34:09 access to frontier coding models is getting more and more expensive very extensively if you've relied on on on GitHub copilot in particular 34:21 because they just changed their whole pricing model really radically. Yes. there's a real incentive and this is something I've invested into recently to do some 34:33 broad experimentation around how can you produce high quality outputs with cheaper models with with either 34:46 non-frontier versions of models or with alternative models that are substantially cheaper rather than you throwing every single even the most 34:56 trivial problem at the most expensive model. ? , , this is something which translates into really significant dollar savings if you 35:06 do it . And , it's a very worthwhile investment from my point of view. But again, it's also something you need to think about a lot in 35:16 terms of , how do you set up an experiment that? how do you make it valid? and and reliable. 35:27 this is ultimately building your own custom agents. there are lots and lots of Lego pieces and and available frameworks 35:40 and whatnot nowadays. every day new ones are being added. definitely the learning is not over. 35:50 Quite the opposite. Yeah. I'm I'm more than ever convinced that the aator the creator agent or agents to 36:00 create that what I don't know what that term is yet. ? Agents should be building things that are deterministic, ? You use the agents to build 36:11 systems and repeatable processes. . I I I get much if I think about the value I get of throwing an agent at a random data model, throwing an agent at a random 36:21 data table and saying go sift through this thing and figure it out sometimes it works, ? Sometimes I get and sometimes it it I have to steer it a little bit and and but man, when I get it to let 36:33 me give you some real examples. I have a lot of thumbnails that I use for YouTube videos. I'm I'm all in the video in the AI video space and I have three or four expressions of you 36:45 know the the home alone look or or angry and thinking or pensive or these different expressions. I had all these images with backgrounds on them 36:55 and you could always go randomly through the internet and find a website that has a background remover just I make it make this take this image take off the background make it a PNG save it to my 37:05 computer. That's what you want to do. I said,"I wonder if my agent could build me a tool that does this." I described what I want. I described the outcome. I need it to be this input. 37:15 I want to be that output. , I don't care what model you pick. You go figure it out. You do the research on what models need to be there. And then with a little bit of refining and tuning, it 37:25 wrote me a Python script, Python 3, that runs on my machine locally. And then it ran it all in CPU initially. And it would take, , I I did 80 images 37:35 in a probably about 5 minutes, 10 minutes. Wow. Impressive. Didn't pay a dime. Didn't have to go to Adobe. Didn't have to use Illustrator. zero effort. And then I said, " document 37:46 all the parameters." It was giving me alpha settings for front, foreground, background, all the things that these models could do. I was impressed. And then I said, ", wait a minute. Let's push oursel a bit here." 37:57 And , how about instead of me googling this, let's see if I can get it to run on my GPU. I have a GPU enabled background remover from five different 38:07 models that can all be run with a command line and any image I get backgrounds are gone that no problem. It's automated. It's it it's a 38:17 very simple process. And thi it's this stuff. It's I need to be able to materialize the the problem down to the solution. . 38:27 And I'm building a lot of these micro tools that are starting to stitch together . I'm solving very small problems which are part of a bigger problem that 38:38 I can start stitching things together . And this is this is what really gets me up in the morning honestly. But can I can I ask you does you being able to build those tools ? Does it 38:49 mean you're less busy? No, it doesn't . . Coming back to the original point. Yes. Yes. , instead of spending it probably could have taken 15 minutes, 39:00 go to a website, click a couple buttons, and maybe 20, that would have been fine. But I will argue the next time I need to remove background images, it's get images down, throw them in a folder, 39:10 hit run, done. The next part of my to your point, ? Let's , I think you you you said this very early on in our conversation, 39:21 which was , is it production ready? . Is is it ready to go? Is it ready to land for real time? I'm not there yet. I'm not ready to land at real time. It's a lot of these I 39:32 I feel I've I've proliferated the amount of work I've been able to do. I can accomplish much more with these additional projects and and solutions 39:43 I'm solving. But is it removing time from my schedule? I'm not sitting here just talking to my agent and eating bon bons and fruit all day. this is not what 39:54 I'm doing. I'm stressing more. I'm working more. I'm I'm talking to my phone in the evenings. I'm chatting with it on Telegram. I'm the volume of things I'm doing is 40:04 increasing. I'm not getting less stuff. It's just more cap. I would never have contemplated building these other little solutions until . that I can do them. I have 40:15 instead of three or four unsolved projects, I've got 15 or 20 unsolved projects that are all really interesting in their own , solving very specific needs, but not necessarily 40:27 stitched together holistically yet. And that's maybe my next move is as I build these small, to your point, small little efforts of of solutioning, how do I start building 40:39 these little Lego bricks of problem solving? How do I start weaving them together into a bigger set that I can just tell the agent here's what I want 40:50 and it knows all the little micro tools that I've built to make things run for me. maybe that's where I'm starting to go . I don't know. Are you experiencing this too? Do you have more side projects than ever? 41:00 what's it look for you? Absolutely. In the past, ? pre- aents. the preent I think I think this is I think this is a shirt that we 41:10 need to have that's I was around pre- agents. Oh yeah, I certainly was. [laughter] [clears throat] we we all have finite 41:21 resources. We all have finite number of hours in a day, ? in the past coming re reacting to the point you just made how did we cope with 41:31 that? How did we preserve our time by saying no to many potential problems much more than we do . . this is one 41:41 thing that's changed. That's true. Lots lots more things are being started because it's certainly viable and possible, ? but the the thing 41:53 I'm really struggling with is the agentic solution is almost certainly never going to get you to 100%. And then 42:04 you end up with having a lot of parallel small projects that you started and you probably got to 80 or 90% of them. But the remaining 10 or 20% which 42:17 then require your input in one way or another either to sign it off or to I don't know provide some credentials or to unblock something. 42:27 Those compound quite significantly. And I've heard from many people that the the sum total of all those remaining 10% is very very 42:40 significant nowadays and ends up making people much more busy than they've ever been when it originally appeared that Asians would take 42:50 away all our workload. . it's conundrum which certainly I'm experiencing. I'm not sure about you, 43:00 but I've definitely heard similar stories from from other people I've talked to. Yeah, I'll I'll Yes, I've heard the same thing. And I would also argue 43:14 when software becomes a commodity and I just built a background remover. You can build a background remover. You can go talk to any agent and build a background everyone's got a 43:24 background remover. you don't need to go buy anything to go purchase that anymore. You don't tactically need to do it . However, to your point 43:35 earlier around skills, that I've got 50 background removers at my disposal, and this is why I feel GitHub has become my social media network at 43:45 this point. Also, last week we discussed GitHub pull requests and GitHub repos are through the roof. We're talking about 20 million new repos every month. 43:55 agents are just making much stuff. How do we discern which background remover is the best background remover or the one that works 44:05 for my workflow? Can I just pick off one off the shelf and just go with it and be none the [snorts] wiser? Back to your point earlier, should I be building an agent to download these things and 44:15 test them out and what's the quality look ? What's what does a good output mean? , this is the the challenge I think we're going to also 44:25 start walking into as you build software, as you do things with agents is distribution. , how what are you doing to mature a distribution channel about what you've been creating? Who 44:36 follows you? that. That is going to be incredibly important as you think through the things that you're building in the future here because you could build stuff all day long and eat 44:47 everyone else's software all day long, but if you don't have a way of distributing that, you could have the best software in the world, but if no one knows about it, it's never going to sell. I don't know what this new world is 44:58 going to look . I think when I look at my business, I'm a small business. Mias, you're a small business as . I think we are with agents on our side, 45:09 we are extremely more advantaged than any large organization that has lots of employees that need to use agents. I'm looking at I'm looking at the the 45:19 story from Uber, ? In Uber, they've got thousands of employees trying to use AI agents and their bill 45:29 is in the billions of dollars around agent spend. You and I spend, let's just say we're spending 300 bucks a month . Double that cost. 45:39 We don't care. It's still adding much more value. Triple the cost. Make it $1,000 a month. I don't know. There's going to be a point where we say no, it's no longer worth our time for 45:50 the spend. But for our time is valuable as a small business and rolling out AI and agents into our teams. We get a huge amount of 46:00 advantage. We can play ball with big players because the agents are part of our suite and our software. I think this is we're we're stepping into another era of the 46:10 independent creator is got immense power with agents. maybe we should end on this topic because we've been we chatted probably way longer than we needed to. we were 46:21 trying to do a 30-minute conversation on this and this turned into much longer. Any any final reaction to that Matias? Maybe a final thought maybe as we wrap here. , on that one, I was going to say I think the the conversion of AI spend into value 46:34 created is probably substantially more favorable the the smaller you are the smaller the company is and it gets much more problematic in in in bigger 46:46 organizations. I don't have evidence to back that but neither do I. [laughter] Maybe and maybe there there is no data on that because how would there be 46:57 but it that's that's something which seems quite likely and which means we have a fantastic competitive advantage and smaller companies 47:08 small mediumsiz businesses in particular interesting new world indeed it is awesome thank you much for listening to Agentic thinking hopefully 47:18 this conversation was interesting to you Matias and I are real time unpacking how agents and agentic things are working in our workflows. We're not doing any less work, that's for sure. 47:29 It's not taking away our jobs from what we can tell. , my my recommendation for those of you who are listening, I mean, you already understand and you already know what's going on here. Push into the job, push into using agents, 47:41 figure it out because I think there's going to be a whole lot of net new work to understand and figure out in your organization. If I'm a business and I'm hiring people , I will hire the person who has more AI experience over 47:52 the one who has certifications and and legacy information or software and coding at this point because I want the new mind to think about agentic things. I think that's going to be a big 48:02 push and I think new software has to be really reimagined solely thinking about how do you to your point, how do you build software, how do you make things 48:12 happen and how do you have the agent observe and watch what's going on there. Last quick quick call out I'll call here. I found a really cool tool that people may be interested in. I do a lot of web app development. This is not 48:22 a paid sponsorship. Sounds a paid sponsorship. It's not. I found this really cool open source. It's called it's it's a play on the word annotation and they replace the front of it with 48:32 agent notation. it's it's agentation is the website agentation.com. I put it in the chat window. It is an ability for you to build a website, put a MCP 48:44 server element on your vibecoded website and you can select divs and pages on the report or whatever the website is built and you can make 48:54 comments and those comments can be directly fed back to your agent and you can say agent resolve the comments and it does a great job of you commenting on a UI and having your agent 49:04 handle those comments directly and then push them back into making new tasks or things as . , this is a really cool tool. I just discovered this. , I rediscovered it. I found it months ago 49:14 and I I lost the link and I'm , crap, where'd it go? I just red dug it out from the the library and I I find it fascinating what this is doing that I really want to recommend to the community. Anyways, agentation is out 49:27 there. Check it out. You might find some value from that if you build apps. Matias, as always, it's been a pleasure. Super fun. This is a blast. We'll do it again soon. See you on Friday.