How I'm Actually Using AI These Days
BDBen Davis@bmdavis419 · 45.9K subscribers
- Views
- 73.5K
- 73,524
- Outlier
- ×3.4
- Median 21.4K
- Length
- 25:08
- Published
- Sep 15
- Tuesday
- Comments
- 119
- 1.3K likes
Script
How it's written
- Words
- 5,838
- Words a minute
- 232
- Reading time
- 25 min
- Segments
- 17
The hook
0:00–0:19Since the models and tooling around them has gotten really good, the way I work and just use my computer in general has fundamentally changed. I've mentioned this before in the past, but I wanted to do a dedicated breakdown on the ways I actually get things done. Obviously for dev work, but also for day-to-day business operations, personal stuff.
Structure
Since the models and tooling around them has gotten really good, the way I work and just use my computer in general has fundamentally changed. I've mentioned this before in the past, but I wanted to do a dedicated breakdown on the ways I actually get things done. Obviously for dev work, but also for day-to-day business operations, personal stuff. I've gotten to the point where the Codex desktop app is effectively my interface for my computer, and that's email, that's Slack, that's Notion. Whenever I need to do a first pass on things, it generally speaking goes through Codex. It's kind of funny as a side note, Cursor, I love the stuff they do and full disclosure, I've run some ads with Cursor, I'm going to run other ones in the future, but they have nothing to do with this video. Cursor has excellent, like really, really good product services. Grok bot is excellent, Cursor Cloud is excellent, their desktop app has actually gotten good, but the models, especially now that they're part of SpaceX, they're really dependent on the Grok series of models, and until those get really, really good, it kind of sucks to have to use only those. I want to use Astra for my stuff, and as such, I've built most of my workflows entirely into the Codex desktop app. When you look at my setup here, there's a couple things that are a little off and strange. Like the first one is I'm not signed in with a profile, I'm signed in with this router down here. And the reason for that is I have multiple Codex subs. I do a lot of work with this thing, obviously coding, but also some of the workflows I'll show you later with, you know, like when this video gets edited and put up on YouTube, the processing that I do to prep the project
Watch from 0:00and get it ready to be edited is going to be done by Codex. A lot of automations that go off every day to handle things like triage and make sure that I'm semi on top of things happens with Codex, and even just basic searching and machine configuration stuff happens through Codex. So, I use this model very heavily. I have three accounts. I'm currently down to not very much of my limits. I I love Astra, but it really burns your limits badly. But since I have what is effectively an internal AI gateway that is running on one of my Linux boxes in my home server, that is then connected to all my different devices over Tailscale, I can then authenticate Codex, Claude Code, and Grok with that API endpoint instead of the normal OAuth sign-in. The OAuth sign-in happens directly in the gateway itself. The interface here is custom because I don't love the default interface from CLI proxy API, which is what this is a fork of. The core of this is really, really good. I love what they put together here. You authenticate your subs, and then you just serve requests to Codex or Claude Code directly using their normal API. I am not using this to stick my sub into things it's not supposed to go into. Uh that's mostly a note for the Claude sub. I am using Claude entirely through Claude Code cuz they don't like it when you use it through Pi or Open Code or whatever. They will ban you for that. I would not recommend doing it.
Watch from 1:31Both Codex and Grok are really chill about that. So actually some of my Hermes agents are running just through this normal gateway load balancing on my subs. Uh same thing with Grok. I have a bunch of random things offed on this. It's really convenient. You get absurd amounts of usage. It just kind of works. Just God, I need a better model. I don't want to spend time on like how to get this set up or whatever, and that's going to go for everything I talk about in this video. The way you should set these things up if they would be useful for your workflow, like if you're burning through all your subs constantly and you really need more usage, load balancing subs is a great way to do it. Having an API is great for that. If you need to set this up, send this link to Codex and tell it, "Hey, I want to set this up. What is the best way to do it?" Even better if you have like other little old devices on your network. And these Linux devices don't need to be super high-end. Like the box that is running my CLI proxy API fork is a pretty small Linux box that I think was like five or six hundred bucks when I got it a while ago. It's a old Ryzen CPU and like 32 gigs of RAM, which to be fair in today's prices is horrible. But if you have like an old laptop or something like that, this is a great use case to just hook that up, get it running 24/7 and use it as a background server and internal API thing. I love having these internal APIs, and the best way to set them up is to ask your agents and work with them, and they will do it very, very well. I also mentioned that I do a ton of work here that isn't normal dev stuff, and I want to go through a couple examples of that and how it works. Unfortunately, the reality of that is so many of the good examples require sensitive data that I can't post publicly. I spend most of my time building internal tools that are for the
Watch from 2:48agency 10XN that manage the ad spots and stuff for my channel and many others. I also use it for, you know, personal one-off tools, a lot of admin work, a lot of research, a lot of just like managing and making videos. It's not stuff that I can really show nearly as much, but I do have a few things that I want to show right after the sponsor break. Today's sponsor is Greptile, the most powerful AI code reviewer. And I call it the most powerful because, well, two reasons. One, their agent is second to none. The way this thing works is really impressive. It fully indexes your code base, then it sends off a swarm of agents in parallel to review all the changes and everything that needs to be seen to get you useful actionable code reviews. The confidence scores are a very high-signal way to just know how good or bad a PR is and how much work is left to do. You can chat directly with Greptile right in the PR, and also it will automatically generate unit tests for new code in every single PR. And this is amplified by their T-Rex system. It is a full sandbox environment that gets spun up to not just look at your changes or run some unit tests, but run the actual code changes to make sure they work. A great example of this is a PR I merged yesterday, which was fixing some download behavior on one of my projects, TX9. And you can see down here the confidence score, the important files changed. You can see it says it's safe to merge cuz it's already gone through the review loop of my agent makes a change, Greptile reviews it, agent fixes the reviews until it's up to a five out of five. Love that workflow.
Watch from 4:20but you can really see the magic here on one of the reviews it left where it found this issue with bounding the asset downloads and then it left these three artifacts which were actual reproductions it made within their sandboxing system. When I click that link it takes me directly to the artifact to see what all happened in here, the code it wrote, the steps it took. Companies like WorkOS, Nvidia, and Brex all use Reptile and having had it on my repos for quite a while, I can tell you I wouldn't want to live without it. And you should really add it to your projects at david 7.link/reptile. This first one is weird, but it uses a lot of the pieces that I think illustrates the point. I had this weird issue that I ran into where Axiom, one of the things that I use for logging on all the services I work with, I have a pro account on one of my accounts, but something weird somewhere happened and I forget which login I used. I was like, okay, I don't know what to do. I sent Codex off to figure it out, tried a bunch of different things, it didn't work, and ultimately what ended up happening is I had it send off an email to support like, "Hey, what is going on here? How do we fix this?" and then got a reply from support and then drafted the reply back to support with this. Was really cool and it uses a lot of the different pieces that are required for day-to-day work. The first of which is the in-app browser within Codex. I found myself using this more and more as a normal browser. At first it was just the localhost preview for things, but now I've got Google signed in here, GitHub signed in here, and I'll often just open something like, you know, in physical I'll open up Axiom and then if I need to make some change that requires these within a project, all of that is being handled through another like central
Watch from 5:48gateway that I have hosted on my Linux machines. It is Executor, which is basically MCP gateway that you can self-host or use their cloud version, which is pretty good. Full disclosure, I am invested in Reese's company, he's a good friend, but I invested cuz I really like it and I think it's dope and I use it for everything. Since it's set up and running on one of the Linux machines, you can see with this like connect to an agent command, this is a Tailscale URL and I am using that Tailscale URL to add that MCP into my Codex instance. It's also in some of my Hermes agents and actually even in my Grok Bot instance. So, this has my Firecrawl, my Google Calendar, my Google Drive, my Gmail, my Notion, my Slack, my Todoist, and some read-only YouTube Analytics. Most of these have multiple accounts on them cuz like I have five Gmail accounts that I have to keep track of. All of those go through here. Multiple Notions, multiple Slacks, you get the idea. Since all of these are in one place and the agent can interact with them via code mode type stuff, I'm able to do things like, "Hey, can you just send an email using my personal email to support asking what's going on?" It'll give me a draft. I'll review it. I'll be like, "Yep, that looks good." Tell it to send. It'll do the send for me. And things like this where I needed to troubleshoot this Axiom issue, now it just kind of happens automatically through Codex. It's a really powerful workflow. This side of it, the browser side, can be done in something like Helium. Like my main browser that I use day-to-day is Helium.
Watch from 7:18But, the issue of just letting the agent go on computer using the Helium is mostly just that it will start clogging up my computer and using up not even my resources, but just my screen. If it's doing something in Helium, I only have one Helium, so I can't do something as well. I have to go to a different browser. If I have it in the in-app browser, I can have as many threads going as I want and each of them can have their own dedicated browser to do whatever miscellaneous thing they want. I think the best way I can explain this side of the workflow and the whole like using Codex as an interface for my computer thing is that the apps I use day-to-day, like Notion for example, I still go into Notion a lot, but I'll often start within Codex. So, for example, the video that I'm recording right now, I have a Notion entry for it. So, uh use executor, go into my Notion, uh or sorry, it's the 10xN Notion, find the video page for my current setup and the way I'm working for things. Uh it should be under currently filming or to film, whatever. I don't remember how it's was But, go find that one and get me a list of the points that I wanted to make in that video. So, it's going to go off, it's going to call the Notion MCP, get all of that data for me, present it to me in chat, and then I can start taking actions from there. If I want to directly edit something, I'll just have it send me a link that'll deep link in, and I can go and do that thing. I just start here. Same thing with like an email type thing. If I'm like, "Oh, I need to go deal with this specific company." I will just have it pull up all of the context and info that I need for that one. And then, I can work off of it from there. I need back and forth, pull in more resources, do more search.
Watch from 8:39Just having a very powerful, capable agent that can go off and both pull the things for me, take actions for me, and do little random things on the side. I set up earlier, I was considering doing a video on Fable 5.1. I'm not going to do a video on Fable 5.1. This is actually what that video was. It'll make sense later, I promise. But, you can see in here, I started with a voice-to-text vomit prompt that included a link to the video template that I put together. That's the way I like structuring videos. I wanted to start that for the Fable 5.1 video. Then, I did some more voice-to-text vomit to have it go pull some sources that I knew of off the top of my head, but I didn't want to go find the specific links for. Then, after a couple minutes, I was like, "Oh, wait, there's another thing. I want to make a skill so that in the future, when I want to start ideating on a new video, it will always pull directly from that template." So, I just put in that same thread, "Go send off a sub-agent to make this new skill." Put out that all together. It ended up doing all that, as well as putting in the links that I was asking for the anti-distillation stuff. Then, I had another video that I wanted to reorganize, the stuff I'm working on with local models. I was like, "Okay, put it in the new format that I have for these videos." It did that for me. And then, after that, I did a completely different thing, and I was like, "Okay, go find the Claude code sessions." Oh, wait, voice-to-text said Claude code. That's fun. But, the machine's smart enough, it knows what that means. It doesn't really matter. You can just kind of say things to it, it'll redirect automatically. It's just The models are so smart and capable that as long as they have the tools, you can kind of just keep throwing random stuff at them and just slowly orchestrate all these pieces together in one place in a way that just feels really, really good.
Watch from 10:09What I love about all of this is it unlocks so many weird new novel workflows that at this point you're kind of only bounded by how creative you can and want to be with these things. If you have a good foundational understanding of how all this stuff works, you know, how the agent works, where it lives, what it can connect to, what capabilities it does and doesn't have, you can start chaining these together in very insane ways. I had this one earlier today where this is a giant hell thread that E- Ethan Eiser, another good friend of mine, was over earlier today and he was recording his video on effects just like why it's really good and why it's taking over. And for fun, on that even for fun, this is how I'm going to do this video, too. I used the skill I put together for prepping a Final Cut Pro project for editing. It doesn't do the full edits cuz it's not quite there yet. I don't trust it for the cutting. That's still actually done by an editor, but a lot of the miscellaneous tedium of go set up this clip, go make the multicam, do the syncing, add in the color grading, do some corrections, get it all prepped and ready to go, make sure the audio is good, then give me the project that's ready to edit instead of just a pile of raw assets. Now, when my editor gets these videos, he gets the actual Final Cut project that's ready to go to edit and it's much quicker and nicer to deal with. That I did that same thing for this video and as it was going, it sent off this stuff, it did all this work with computer use cuz computer use on Astra is just unbelievable. What's really cool is it didn't actually do this work on this laptop. It did all of it on my It's hard to see here, but it's like down back behind me. I have one of my old MacBooks plugged in set to never sleep. It has an Ethernet cable on it
Watch from 11:40and that guy is effectively now a background agent that's any of my normal Codex instances can call as a sub agent. So, it sent off what is effectively a sub agent prompt from here to grab these assets, prep the thing, follow the workflow, and it did a really good job. There were some kinks. It did not go perfectly, and it also took way longer than it should have. So, you can even see like down here like Ethan was just balding at it cuz it was being annoying, understandably so. But after all of that happened and it was staged and set up, I just had to go through this thread and be like, all right, go through everything, see how well it went, how can we improve this, what should be changed? You can see it worked for a little while, gave me some info on this. Then afterwards, I responded to each of its points, told it to make the changes to the scale, make the little tool for helping send stuff up and down, work for 20 minutes, gave me this really nice visual breakdown within the actual editor. This is one of the reasons why I still lean on Codex heavily, despite the fact that I do have a lot of problems with this desktop app. Their priority view sucks. Like it's very similar to the T3 code sidebar. Well, it's kind of inspired by it, and we'll talk about T3 code later, but like it sucks. It is weird. It's not what I really want, but it's good enough to where I can live with it. The place where Codex is good is right here in the main chat view for stuff like this, and also the in-app browser is really good. I like their artifacts view where, you know, if I open up an agent's MD, this is a nice real editor where I can make changes and easily read what it did. All these little things work super well, computer uses first class, all that stuff. Having this visualization, which is just some
Watch from 13:12custom HTML that it made and injected into the page, looks really good, gave me a good idea of, okay, here are all the steps, the preparation, delivery, and receiving. Gave it another giant Voice to Text vomit prompt telling it to like, all right, cool, this works. Let's do a V2 of this and test out how the workflow would work with the changes we made. But obviously I didn't have a video at that point, so I was like, okay, let's just use the same one and test it there. But this thread is poison. This thread already knows too much. What I had to do is since Astor is incredible at swarming, I've talked about this in my dedicated Astor video, but like this thing is really good at working with itself. You can have 20 Astors communicating via message board or direct message or whatever, they will get good work done. So, this main Astra instance spawned up another Astra instance, which was just a new Codex thread, and it was basically pretending to be me. How I would interact with it, the prompts I would send it, it sent the same prompts that I did at the very start of this thread, and watched as the workflow went. It baby-sat it, it saw what it was doing on that machine, it was just keeping a tabs on all these agents. It actually even sent some steering messages to it to be like, "Hey, don't do it that way. Change this, change that." Worked for about an hour, gave me V2. I looked through it, was much better. And then at the very end, I made sure everything was persisted, and when I'm done recording this one, I'm going to fire off this workflow again, and just have it go prep and chop the assets on that other video, get it on the NAS, and it's all set and ready to go. These workflows are insane. There's nothing like it. Be creative, find weird ways to hook these things together. I don't know if this exact shape is the future, but this kind of thing where
Watch from 14:43you're working back and forth with an agent almost at a co-worker level, or really an assistant level, it hasn't been there before. And I think the biggest problem we have right now is that there are not enough models that are really capable of doing this well. There are other ones that can handle single-threaded stuff. There are other ones that can handle sub-agents. That was very much what the last generation was good at. The open models caught up, the American labs that aren't OpenAI and Anthropic mostly caught up. It's just Now OpenAI and Anthropic have delivered their new next generation in Astra and Fable 5.1, and we now have this new fundamental capability unlocked, which is that swarming behavior. And also with Astra, the computer use stuff, nothing else quite has it. Also, the writing of Astra is so much better than any other model I've ever interacted with. For knowledge work stuff, it makes it so much better. So, for now, it's just Astra, but give it 6 months, and this will be everywhere, and the ways we interact with this will also probably be pretty different. But if you are using this stuff right now, I do recommend setting something like this up. Which does then kind of lead me into, okay, cool, this is all neat, but how does the dev stuff work now? I do a lot of my dev work still just in Codex on this machine. A good example of a thing that I built recently is this little WhisperFlow alternative. It's just a local voice to text thing. I believe it's open source. I'll go double-check.
Watch from 16:13I honestly don't even know cuz I just had a bunch of Astra instances go off, build the thing, test it a bunch. And what was really neat about that is it would use computer use to make sure that the thing actually worked. And I did a ton of iterations on it. And now at this point it's using a local model on my computer. I believe it's the Whisper Turbo model. Fully open source, nothing leaves my device. I have all my voice transcripts. That was really the motivation behind this is I wanted to keep the transcripts just for myself as well as the raw audio recordings for myself so I can use those in the future if I feel like it. I like hoarding data at this point now that models have gotten so good. Part of the reason why I have my router gateway thing here is that I am using this to hoard all of the logs and tool calls and traces that all my agents do. So that in the future if I ever want to do anything with this, I can do it. Hoard data. It's very good to hoard data right now. But again, as I was saying, all of this work was done in Codex. A lot of computer use on this machine. And it works really, really well. But there are certain projects that either one require much longer running background stuff or also I just don't want to have my computer on all the time. This was very much an in-the-loop project where the core of it isn't all that complicated. Most of the time was just spent on UX stuff, improving how it looks and feels and behaviors. This menu I spent a lot of time smacking it to stop putting stuff in there. When it first made this, this was about twice as long and had five x as much text. Most of my work with UI on models these days is telling them to do less, make less. And then within the management UI, again, a lot of polishing on how I want this to work. Spent some
Watch from 17:30time, like I really wanted this input priority list. So like if I have this mic, I want to make sure that that's the default. Then if I don't have that, use my Rodecaster. If I don't have that, use my Studio Display. If I don't have that, use the built-in mic. That priority list works really well. I have a dictionary, which I've actually haven't been using. Uh you can test dictation, see the history of all the different things I've been voice-to-texting, see what models are being used. Again, I had to do a lot of polishing to get this where I wanted it. But, at this point, I'm like really happy with where this app is. It works great for me personally, and you know, it was a side project I did with Astra over the course of probably 5 to 10 hours. It was good. It works really well. Then, for a lot of my other work, there is T3 Code. This is where I do all of my Claude Code stuff. I don't really interact with Claude Code through the CLI, even though its CLI has gotten quite good. I've been very impressed with how much that has improved over the last 6 to 12 months. Actually, not 12 months, like 6 months. It was really rough in January of this year, but now every time I open it, I'm very happy with it. It feels good. It works well. I just don't like working in CLIs all that much anymore. I find myself losing work more, and it's really just not where I want to be. It's much easier to have a desktop app for this stuff, especially since I parallelize work a lot. And for more normal projects, which would be like a web app type thing or a custom server or whatever random thing I need to do, T3 Code works quite well for that. You probably heard about it if you're watching this video. I'm not going to spend too much time here. It's great. The remote stuff's great. I really like their sidebar. I I don't know. I honestly don't have too much to say other than it's good. If you're
Watch from 19:02doing coding work, offloading it from your laptop is a very good thing to do. If you want more details on that, go watch my Linux video that I just put out. Uh yeah, I'm not going to spend more time on that. So, last thing I just want to quickly cover before I wrap this up is a quick comparison between the models. I've touched on this briefly earlier, but as I said, the main models I'm using these days are Astra, Fable 5.1, and sometimes Grok-4-6. I actually do use Grok-4-6 more than I'd expect through the Grok build CLI, but one of the biggest reasons for that is that it has direct access to the XAPI without having to pay their API prices, which are obscene. So, it's very useful for when I need to go grab some post or I need to go fill in some data set on this is the post for this video or whatever. It was great. I really like it. I'm very hopeful that they get better models in the future, but at least for right now, the big two models are obviously Astra and Fable. I found myself using Fable a lot more than I expected to cuz whenever I'm doing, you know, normal dev work type stuff, most of my main interaction is with Astra. I still adore the model. I know it has its quirks, but the quirks work for me and it's great. Great. Fable is very good though for cleaning things up. I'll often have Astra go make some change or add some feature or whatever. I'll look at the PR, I'll see plus 5,000 minus a hundred for what should be a very simple thing. I'll just be I'll throw it over to Fable and be like, all right, what are we actually doing here?
Watch from 20:33How do we simplify the hell out of this? And it can generally speaking get that way trimmed down. I kind of just use it as cleanup crew for the open AI stuff. The code Fable 5.1 writes is admittedly very beautiful. I It's just hard to push it super hard because the limits are so brutal and bad. Astras are not great either, but they're still better than Fables, and the tooling in places where I can use it is just so much more limited. I don't want to use the Claude desktop app. They don't have the same computer use level stuff, so it's just much harder to push the model, and the really crazy top-of-the-line stuff you can get from both these models, like the swarming stuff, requires you to effectively incinerate an entire week of usage just to do it. On Claude, you borderline can't. Like I I had a thing that I did a little while ago where I was looking at taking my A drive, which is just my self-hosted custom Agents Google Drive that I use for everything so that I can be like, all right, this is a good plan or this is a good breakdown of something. Can you just host this quickly, give me a URL for it, and then I can send that to people? It was incredibly useful for day-to-day work with the team, and also when we were doing like Def Con and puzzle solving. It was a great way to just send artifacts over to people. I wanted to see, okay, what would it look like to host this and make it a thing that other people can use. I went to bed, but before I went to bed I was like, "Okay, you can work all night, have fun, go try and implement this plan we've been working on." It ran for almost 8 hours, and when I woke up, my usage had gone from 80% down to 20%. It's just not viable. Unless you are burning five to 10 of these subs, you just can't do it, unfortunately. It really sucks, but until the tokens get cheaper and we get
Watch from 21:52these same capabilities in other models, it your workflows are inherently limited. I do think that sucks. It's a provisioning issue. Hopefully, if the curve continues, we'll get 10x better models at 10x cheaper prices in the future. We'll see. But for now, Phind 5.1 is a good model. I like interacting with it. For the first time, I generally speaking hate Claude models, but I do like this one, and it does the job. I just can't use enough of it. Astra, I have already talked about Astra a ton. You know how much I love this model. So, yeah, that's uh that's kind of how I work. I I know this is kind of chaotic and a mess, but that's really what it is. I'm figuring out these workflows as I go, and I'm sure within a month or two this will be completely the future is definitely trending in the direction where you spend more and more time interacting directly with the agent as a proxy to other things, and then once it's done some work, you go into that other app and finish the work. The analogy I've been thinking about a lot is, you know, how we were using Cursor a year ago, where we were mostly just using the sidebar and then had the code open, so it was in-editor type thing, very much a Copilot type experience. And now we're at the point where I'm mostly using something like the Codex desktop app, which is completely abstracted away from the code. I can still read it when I need to, but it's not right in my face. This isn't an editor. But they've gotten so good that that's fine. It's fine for me to be abstracted away more and just more interact with the agent than the code. I don't trust that at the level for anything like video editing or thumbnails or even just a lot of like general admin management work. Like I still want to be able to go in Notion and edit the Notion document. But it's still quite nice to have it do some passes, do some research, and dump artifacts in there. You're now kind of just pulling back and forth in a co-piloty type experience. Go experiment with this stuff, see what you think. Let me know how you're using this stuff down below. If you enjoyed, make sure you like and subscribe, and I'll talk to you again soon.
Watch from 23:23Full transcript
- 0:00
Since the models and tooling around them has gotten really good, the way I work and just use my computer in general has fundamentally changed. I've mentioned this before in the past, but I wanted to do a dedicated breakdown on the ways I actually get things done. Obviously for dev work, but also for day-to-day business operations, personal stuff. I've gotten to the point where the Codex
- 0:20
desktop app is effectively my interface for my computer, and that's email, that's Slack, that's Notion. Whenever I need to do a first pass on things, it generally speaking goes through Codex. It's kind of funny as a side note, Cursor, I love the stuff they do and full disclosure, I've run some ads with Cursor, I'm going to run other ones in the future, but they have nothing to do with this video. Cursor has excellent,
- 0:42
like really, really good product services. Grok bot is excellent, Cursor Cloud is excellent, their desktop app has actually gotten good, but the models, especially now that they're part of SpaceX, they're really dependent on the Grok series of models, and until those get really, really good, it kind of sucks to have to use only those. I want to use Astra for my stuff, and as
- 1:04
such, I've built most of my workflows entirely into the Codex desktop app. When you look at my setup here, there's a couple things that are a little off and strange. Like the first one is I'm not signed in with a profile, I'm signed in with this router down here. And the reason for that is I have multiple Codex subs. I do a lot of work with this thing, obviously coding, but also some
- 1:24
of the workflows I'll show you later with, you know, like when this video gets edited and put up on YouTube, the processing that I do to prep the project and get it ready to be edited is going to be done by Codex. A lot of automations that go off every day to handle things like triage and make sure that I'm semi on top of things happens with Codex, and even just basic searching and machine configuration
- 1:44
stuff happens through Codex. So, I use this model very heavily. I have three accounts. I'm currently down to not very much of my limits. I I love Astra, but it really burns your limits badly. But since I have what is effectively an internal AI gateway that is running on one of my Linux boxes in my home server, that is then connected to all my different devices over
- 2:05
Tailscale, I can then authenticate Codex, Claude Code, and Grok with that API endpoint instead of the normal OAuth sign-in. The OAuth sign-in happens directly in the gateway itself. The interface here is custom because I don't love the default interface from CLI proxy API, which is what this is a fork of. The core of this is really, really
- 2:25
good. I love what they put together here. You authenticate your subs, and then you just serve requests to Codex or Claude Code directly using their normal API. I am not using this to stick my sub into things it's not supposed to go into. Uh that's mostly a note for the Claude sub. I am using Claude entirely through Claude Code cuz they don't like it when you use it through Pi or Open Code or whatever. They will ban you for
- 2:46
that. I would not recommend doing it. Both Codex and Grok are really chill about that. So actually some of my Hermes agents are running just through this normal gateway load balancing on my subs. Uh same thing with Grok. I have a bunch of random things offed on this. It's really convenient. You get absurd amounts of usage. It just kind of works. Just God, I need a better model. I don't want to spend time on like how to get this set up or whatever, and that's
- 3:07
going to go for everything I talk about in this video. The way you should set these things up if they would be useful for your workflow, like if you're burning through all your subs constantly and you really need more usage, load balancing subs is a great way to do it. Having an API is great for that. If you need to set this up, send this link to Codex and tell it, "Hey, I want to set this up. What is the best way to do it?" Even better if you have like other little old devices on your network. And
- 3:28
these Linux devices don't need to be super high-end. Like the box that is running my CLI proxy API fork is a pretty small Linux box that I think was like five or six hundred bucks when I got it a while ago. It's a old Ryzen CPU and like 32 gigs of RAM, which to be fair in today's prices is horrible. But if you have like an old laptop or
- 3:49
something like that, this is a great use case to just hook that up, get it running 24/7 and use it as a background server and internal API thing. I love having these internal APIs, and the best way to set them up is to ask your agents and work with them, and they will do it very, very well. I also mentioned that I do a ton of work here that isn't normal dev stuff, and I want to go through a couple examples of that and how it
- 4:09
works. Unfortunately, the reality of that is so many of the good examples require sensitive data that I can't post publicly. I spend most of my time building internal tools that are for the agency 10XN that manage the ad spots and stuff for my channel and many others. I also use it for, you know, personal one-off tools, a lot of admin work, a lot of research, a lot of just like
- 4:30
managing and making videos. It's not stuff that I can really show nearly as much, but I do have a few things that I want to show right after the sponsor break. Today's sponsor is Greptile, the most powerful AI code reviewer. And I call it the most powerful because, well, two reasons. One, their agent is second to none. The way this thing works is really impressive. It fully indexes your
- 4:51
code base, then it sends off a swarm of agents in parallel to review all the changes and everything that needs to be seen to get you useful actionable code reviews. The confidence scores are a very high-signal way to just know how good or bad a PR is and how much work is left to do. You can chat directly with Greptile right in the PR, and also it
- 5:12
will automatically generate unit tests for new code in every single PR. And this is amplified by their T-Rex system. It is a full sandbox environment that gets spun up to not just look at your changes or run some unit tests, but run the actual code changes to make sure they work. A great example of this is a PR I merged yesterday, which was fixing some download behavior on one of my
- 5:32
projects, TX9. And you can see down here the confidence score, the important files changed. You can see it says it's safe to merge cuz it's already gone through the review loop of my agent makes a change, Greptile reviews it, agent fixes the reviews until it's up to a five out of five. Love that workflow. but you can really see the magic here on one of the reviews it left where it found this issue with bounding the asset
- 5:53
downloads and then it left these three artifacts which were actual reproductions it made within their sandboxing system. When I click that link it takes me directly to the artifact to see what all happened in here, the code it wrote, the steps it took. Companies like WorkOS, Nvidia, and Brex all use Reptile and having had it on my repos for quite a while, I can tell you I wouldn't want to live without
- 6:14
it. And you should really add it to your projects at david 7.link/reptile. This first one is weird, but it uses a lot of the pieces that I think illustrates the point. I had this weird issue that I ran into where Axiom, one of the things that I use for logging on all the services I work with, I have a pro account on one of my accounts, but something weird somewhere happened and I forget which login I used. I was like,
- 6:36
okay, I don't know what to do. I sent Codex off to figure it out, tried a bunch of different things, it didn't work, and ultimately what ended up happening is I had it send off an email to support like, "Hey, what is going on here? How do we fix this?" and then got a reply from support and then drafted the reply back to support with this. Was really cool and it uses a lot of the different pieces that are required for day-to-day work. The first of which is
- 6:57
the in-app browser within Codex. I found myself using this more and more as a normal browser. At first it was just the localhost preview for things, but now I've got Google signed in here, GitHub signed in here, and I'll often just open something like, you know, in physical I'll open up Axiom and then if I need to make some change that requires these within a project, all of that is being handled through another like central
- 7:18
gateway that I have hosted on my Linux machines. It is Executor, which is basically MCP gateway that you can self-host or use their cloud version, which is pretty good. Full disclosure, I am invested in Reese's company, he's a good friend, but I invested cuz I really like it and I think it's dope and I use it for everything. Since it's set up and running on one of the Linux machines, you can see with this like connect to an agent command, this is a Tailscale URL
- 7:41
and I am using that Tailscale URL to add that MCP into my Codex instance. It's also in some of my Hermes agents and actually even in my Grok Bot instance. So, this has my Firecrawl, my Google Calendar, my Google Drive, my Gmail, my Notion, my Slack, my Todoist, and some read-only YouTube Analytics. Most of these have multiple accounts on them cuz like I have five Gmail accounts that I
- 8:02
have to keep track of. All of those go through here. Multiple Notions, multiple Slacks, you get the idea. Since all of these are in one place and the agent can interact with them via code mode type stuff, I'm able to do things like, "Hey, can you just send an email using my personal email to support asking what's going on?" It'll give me a draft. I'll review it. I'll be like, "Yep, that looks good." Tell it to send. It'll do
- 8:23
the send for me. And things like this where I needed to troubleshoot this Axiom issue, now it just kind of happens automatically through Codex. It's a really powerful workflow. This side of it, the browser side, can be done in something like Helium. Like my main browser that I use day-to-day is Helium. But, the issue of just letting the agent go on computer using the Helium is mostly just that it will start clogging
- 8:44
up my computer and using up not even my resources, but just my screen. If it's doing something in Helium, I only have one Helium, so I can't do something as well. I have to go to a different browser. If I have it in the in-app browser, I can have as many threads going as I want and each of them can have their own dedicated browser to do whatever miscellaneous thing they want. I think the best way I can explain this side of the workflow and the whole like
- 9:06
using Codex as an interface for my computer thing is that the apps I use day-to-day, like Notion for example, I still go into Notion a lot, but I'll often start within Codex. So, for example, the video that I'm recording right now, I have a Notion entry for it. So, uh use executor, go into my Notion, uh or sorry, it's the 10xN Notion, find
- 9:27
the video page for my current setup and the way I'm working for things. Uh it should be under currently filming or to film, whatever. I don't remember how it's was But, go find that one and get me a list of the points that I wanted to make in that video. So, it's going to go off, it's going to call the Notion MCP, get all of that data for me, present it to me in chat, and then I can start taking
- 9:47
actions from there. If I want to directly edit something, I'll just have it send me a link that'll deep link in, and I can go and do that thing. I just start here. Same thing with like an email type thing. If I'm like, "Oh, I need to go deal with this specific company." I will just have it pull up all of the context and info that I need for that one. And then, I can work off of it from there. I need back and forth, pull in more resources, do more search.
- 10:09
Just having a very powerful, capable agent that can go off and both pull the things for me, take actions for me, and do little random things on the side. I set up earlier, I was considering doing a video on Fable 5.1. I'm not going to do a video on Fable 5.1. This is actually what that video was. It'll make sense later, I promise. But, you can see
- 10:29
in here, I started with a voice-to-text vomit prompt that included a link to the video template that I put together. That's the way I like structuring videos. I wanted to start that for the Fable 5.1 video. Then, I did some more voice-to-text vomit to have it go pull some sources that I knew of off the top of my head, but I didn't want to go find the specific links for. Then, after a couple minutes, I was like, "Oh, wait, there's another thing. I want to make a skill so that in the future, when I want
- 10:50
to start ideating on a new video, it will always pull directly from that template." So, I just put in that same thread, "Go send off a sub-agent to make this new skill." Put out that all together. It ended up doing all that, as well as putting in the links that I was asking for the anti-distillation stuff. Then, I had another video that I wanted to reorganize, the stuff I'm working on with local models. I was like, "Okay,
- 11:11
put it in the new format that I have for these videos." It did that for me. And then, after that, I did a completely different thing, and I was like, "Okay, go find the Claude code sessions." Oh, wait, voice-to-text said Claude code. That's fun. But, the machine's smart enough, it knows what that means. It doesn't really matter. You can just kind of say things to it, it'll redirect automatically. It's just The models are so smart and capable that as long as
- 11:31
they have the tools, you can kind of just keep throwing random stuff at them and just slowly orchestrate all these pieces together in one place in a way that just feels really, really good. What I love about all of this is it unlocks so many weird new novel workflows that at this point you're kind of only bounded by how creative you can and want to be with these things. If you have a good foundational understanding of how all this stuff works, you know,
- 11:52
how the agent works, where it lives, what it can connect to, what capabilities it does and doesn't have, you can start chaining these together in very insane ways. I had this one earlier today where this is a giant hell thread that E- Ethan Eiser, another good friend of mine, was over earlier today and he was recording his video on effects just like why it's really good and why it's taking over. And for fun, on that even
- 12:14
for fun, this is how I'm going to do this video, too. I used the skill I put together for prepping a Final Cut Pro project for editing. It doesn't do the full edits cuz it's not quite there yet. I don't trust it for the cutting. That's still actually done by an editor, but a lot of the miscellaneous tedium of go set up this clip, go make the multicam, do the syncing, add in the color grading, do some corrections, get it all
- 12:35
prepped and ready to go, make sure the audio is good, then give me the project that's ready to edit instead of just a pile of raw assets. Now, when my editor gets these videos, he gets the actual Final Cut project that's ready to go to edit and it's much quicker and nicer to deal with. That I did that same thing for this video and as it was going, it sent off this stuff, it did all this work with computer use cuz computer use
- 12:56
on Astra is just unbelievable. What's really cool is it didn't actually do this work on this laptop. It did all of it on my It's hard to see here, but it's like down back behind me. I have one of my old MacBooks plugged in set to never sleep. It has an Ethernet cable on it and that guy is effectively now a background agent that's any of my normal
- 13:17
Codex instances can call as a sub agent. So, it sent off what is effectively a sub agent prompt from here to grab these assets, prep the thing, follow the workflow, and it did a really good job. There were some kinks. It did not go perfectly, and it also took way longer than it should have. So, you can even see like down here like Ethan was just balding at it cuz it was being annoying,
- 13:37
understandably so. But after all of that happened and it was staged and set up, I just had to go through this thread and be like, all right, go through everything, see how well it went, how can we improve this, what should be changed? You can see it worked for a little while, gave me some info on this. Then afterwards, I responded to each of its points, told it to make the changes to the scale, make the little tool for helping send stuff up and down, work for
- 13:58
20 minutes, gave me this really nice visual breakdown within the actual editor. This is one of the reasons why I still lean on Codex heavily, despite the fact that I do have a lot of problems with this desktop app. Their priority view sucks. Like it's very similar to the T3 code sidebar. Well, it's kind of inspired by it, and we'll talk about T3 code later, but like it sucks. It is
- 14:18
weird. It's not what I really want, but it's good enough to where I can live with it. The place where Codex is good is right here in the main chat view for stuff like this, and also the in-app browser is really good. I like their artifacts view where, you know, if I open up an agent's MD, this is a nice real editor where I can make changes and easily read what it did. All these little things work super well, computer
- 14:39
uses first class, all that stuff. Having this visualization, which is just some custom HTML that it made and injected into the page, looks really good, gave me a good idea of, okay, here are all the steps, the preparation, delivery, and receiving. Gave it another giant Voice to Text vomit prompt telling it to like, all right, cool, this works. Let's do a V2 of this and test out how the workflow would work with the changes we
- 15:00
made. But obviously I didn't have a video at that point, so I was like, okay, let's just use the same one and test it there. But this thread is poison. This thread already knows too much. What I had to do is since Astor is incredible at swarming, I've talked about this in my dedicated Astor video, but like this thing is really good at working with itself. You can have 20 Astors communicating via message board or direct message or whatever, they will
- 15:21
get good work done. So, this main Astra instance spawned up another Astra instance, which was just a new Codex thread, and it was basically pretending to be me. How I would interact with it, the prompts I would send it, it sent the same prompts that I did at the very start of this thread, and watched as the workflow went. It baby-sat it, it saw what it was doing on that machine, it was just keeping a tabs on all these
- 15:42
agents. It actually even sent some steering messages to it to be like, "Hey, don't do it that way. Change this, change that." Worked for about an hour, gave me V2. I looked through it, was much better. And then at the very end, I made sure everything was persisted, and when I'm done recording this one, I'm going to fire off this workflow again, and just have it go prep and chop the assets on that other video, get it on the NAS, and it's all set and ready to
- 16:02
go. These workflows are insane. There's nothing like it. Be creative, find weird ways to hook these things together. I don't know if this exact shape is the future, but this kind of thing where you're working back and forth with an agent almost at a co-worker level, or really an assistant level, it hasn't been there before. And I think the biggest problem we have right now is
- 16:23
that there are not enough models that are really capable of doing this well. There are other ones that can handle single-threaded stuff. There are other ones that can handle sub-agents. That was very much what the last generation was good at. The open models caught up, the American labs that aren't OpenAI and Anthropic mostly caught up. It's just Now OpenAI and Anthropic have delivered their new next generation in Astra and
- 16:43
Fable 5.1, and we now have this new fundamental capability unlocked, which is that swarming behavior. And also with Astra, the computer use stuff, nothing else quite has it. Also, the writing of Astra is so much better than any other model I've ever interacted with. For knowledge work stuff, it makes it so much better. So, for now, it's just Astra, but give it 6 months, and this
- 17:03
will be everywhere, and the ways we interact with this will also probably be pretty different. But if you are using this stuff right now, I do recommend setting something like this up. Which does then kind of lead me into, okay, cool, this is all neat, but how does the dev stuff work now? I do a lot of my dev work still just in Codex on this machine. A good example of a thing that I built recently is this little
- 17:25
WhisperFlow alternative. It's just a local voice to text thing. I believe it's open source. I'll go double-check. I honestly don't even know cuz I just had a bunch of Astra instances go off, build the thing, test it a bunch. And what was really neat about that is it would use computer use to make sure that the thing actually worked. And I did a ton of iterations on it. And now at this point it's using a local model on my
- 17:47
computer. I believe it's the Whisper Turbo model. Fully open source, nothing leaves my device. I have all my voice transcripts. That was really the motivation behind this is I wanted to keep the transcripts just for myself as well as the raw audio recordings for myself so I can use those in the future if I feel like it. I like hoarding data at this point now that models have gotten so good. Part of the reason why I
- 18:07
have my router gateway thing here is that I am using this to hoard all of the logs and tool calls and traces that all my agents do. So that in the future if I ever want to do anything with this, I can do it. Hoard data. It's very good to hoard data right now. But again, as I was saying, all of this work was done in Codex. A lot of computer use on this machine. And it works really, really
- 18:27
well. But there are certain projects that either one require much longer running background stuff or also I just don't want to have my computer on all the time. This was very much an in-the-loop project where the core of it isn't all that complicated. Most of the time was just spent on UX stuff, improving how it looks and feels and behaviors. This menu I spent a lot of time smacking it to stop putting stuff
- 18:47
in there. When it first made this, this was about twice as long and had five x as much text. Most of my work with UI on models these days is telling them to do less, make less. And then within the management UI, again, a lot of polishing on how I want this to work. Spent some time, like I really wanted this input priority list. So like if I have this mic, I want to make sure that that's the
- 19:07
default. Then if I don't have that, use my Rodecaster. If I don't have that, use my Studio Display. If I don't have that, use the built-in mic. That priority list works really well. I have a dictionary, which I've actually haven't been using. Uh you can test dictation, see the history of all the different things I've been voice-to-texting, see what models are being used. Again, I had to do a lot of polishing to get this where I wanted it. But, at this point, I'm like really
- 19:29
happy with where this app is. It works great for me personally, and you know, it was a side project I did with Astra over the course of probably 5 to 10 hours. It was good. It works really well. Then, for a lot of my other work, there is T3 Code. This is where I do all of my Claude Code stuff. I don't really interact with Claude Code through the CLI, even though its CLI has gotten quite good. I've been very impressed
- 19:49
with how much that has improved over the last 6 to 12 months. Actually, not 12 months, like 6 months. It was really rough in January of this year, but now every time I open it, I'm very happy with it. It feels good. It works well. I just don't like working in CLIs all that much anymore. I find myself losing work more, and it's really just not where I want to be. It's much easier to have a
- 20:10
desktop app for this stuff, especially since I parallelize work a lot. And for more normal projects, which would be like a web app type thing or a custom server or whatever random thing I need to do, T3 Code works quite well for that. You probably heard about it if you're watching this video. I'm not going to spend too much time here. It's great. The remote stuff's great. I really like their sidebar. I I don't
- 20:30
know. I honestly don't have too much to say other than it's good. If you're doing coding work, offloading it from your laptop is a very good thing to do. If you want more details on that, go watch my Linux video that I just put out. Uh yeah, I'm not going to spend more time on that. So, last thing I just want to quickly cover before I wrap this up is a quick comparison between the models. I've touched on this briefly earlier, but as I said, the main models
- 20:50
I'm using these days are Astra, Fable 5.1, and sometimes Grok-4-6. I actually do use Grok-4-6 more than I'd expect through the Grok build CLI, but one of the biggest reasons for that is that it has direct access to the XAPI without having to pay their API prices, which are obscene. So, it's very useful for when I need to go grab some post or I need to go fill in some data set on this
- 21:13
is the post for this video or whatever. It was great. I really like it. I'm very hopeful that they get better models in the future, but at least for right now, the big two models are obviously Astra and Fable. I found myself using Fable a lot more than I expected to cuz whenever I'm doing, you know, normal dev work type stuff, most of my main interaction is with Astra. I still adore the model. I know it has its quirks, but the quirks
- 21:34
work for me and it's great. Great. Fable is very good though for cleaning things up. I'll often have Astra go make some change or add some feature or whatever. I'll look at the PR, I'll see plus 5,000 minus a hundred for what should be a very simple thing. I'll just be I'll throw it over to Fable and be like, all right, what are we actually doing here? How do we simplify the hell out of this? And it can generally speaking get that
- 21:55
way trimmed down. I kind of just use it as cleanup crew for the open AI stuff. The code Fable 5.1 writes is admittedly very beautiful. I It's just hard to push it super hard because the limits are so brutal and bad. Astras are not great either, but they're still better than Fables, and the tooling in places where I can use it is just so much more limited. I don't want to use the Claude
- 22:16
desktop app. They don't have the same computer use level stuff, so it's just much harder to push the model, and the really crazy top-of-the-line stuff you can get from both these models, like the swarming stuff, requires you to effectively incinerate an entire week of usage just to do it. On Claude, you borderline can't. Like I I had a thing that I did a little while ago where I was looking at taking my A drive, which
- 22:36
is just my self-hosted custom Agents Google Drive that I use for everything so that I can be like, all right, this is a good plan or this is a good breakdown of something. Can you just host this quickly, give me a URL for it, and then I can send that to people? It was incredibly useful for day-to-day work with the team, and also when we were doing like Def Con and puzzle solving. It was a great way to just send artifacts over to people. I wanted to
- 22:58
see, okay, what would it look like to host this and make it a thing that other people can use. I went to bed, but before I went to bed I was like, "Okay, you can work all night, have fun, go try and implement this plan we've been working on." It ran for almost 8 hours, and when I woke up, my usage had gone from 80% down to 20%. It's just not viable. Unless you are burning five to 10 of these subs, you just can't do it,
- 23:19
unfortunately. It really sucks, but until the tokens get cheaper and we get these same capabilities in other models, it your workflows are inherently limited. I do think that sucks. It's a provisioning issue. Hopefully, if the curve continues, we'll get 10x better models at 10x cheaper prices in the future. We'll see. But for now, Phind 5.1 is a good model. I like interacting with it. For the first time, I generally
- 23:41
speaking hate Claude models, but I do like this one, and it does the job. I just can't use enough of it. Astra, I have already talked about Astra a ton. You know how much I love this model. So, yeah, that's uh that's kind of how I work. I I know this is kind of chaotic and a mess, but that's really what it is. I'm figuring out these workflows as I go, and I'm sure within a month or two this will be completely the future is definitely trending in the
- 24:02
direction where you spend more and more time interacting directly with the agent as a proxy to other things, and then once it's done some work, you go into that other app and finish the work. The analogy I've been thinking about a lot is, you know, how we were using Cursor a year ago, where we were mostly just using the sidebar and then had the code open, so it was in-editor type thing, very much a Copilot type experience. And
- 24:24
now we're at the point where I'm mostly using something like the Codex desktop app, which is completely abstracted away from the code. I can still read it when I need to, but it's not right in my face. This isn't an editor. But they've gotten so good that that's fine. It's fine for me to be abstracted away more and just more interact with the agent than the code. I don't trust that at the level for anything like video editing or thumbnails or even just a lot of like
- 24:46
general admin management work. Like I still want to be able to go in Notion and edit the Notion document. But it's still quite nice to have it do some passes, do some research, and dump artifacts in there. You're now kind of just pulling back and forth in a co-piloty type experience. Go experiment with this stuff, see what you think. Let me know how you're using this stuff down below. If you enjoyed, make sure you like and subscribe, and I'll talk to you again soon.
Description
Wanted to talk about the ways I'm using all the modern AI stuff for my day to day work (both dev and non-dev stuff)
Thank you Greptile for sponsoring! Check them out at https://davis7.link/greptile
cli proxy - https://github.com/router-for-me/CLIP...
executor (connecting my tools) - https://executor.sh
hardware setup - https://www.davis7.sh/en/home-server
My X, Discord, and more: https://davis7.sh/
