# How Codex Became a Directionless Mess

Channel: Ben Davis
Video: https://www.youtube.com/watch?v=uugH4I82PeI
Duration: 20 min
Language: English
Words: 4270
Transcript page: https://viewrankai.com/tools/youtube-transcript/uugH4I82PeI

---

[0:00] At the start of this year, Open AI released their Codex desktop app, and it was amazing at the time. Everyone really liked it. I moved over to it. Everyone I knew moved over to it. It was great, and it got better and better until things started to get weird. It started to have a lot of performance issues. There was some weirdness in the product. Then the Chat GPT and Codex merger happened, so now it's the Chat GPT desktop app, and it has unfortunately not improved a ton since. I actually think it's gotten a lot worse. Despite the fact that there are a lot of things that are getting

[0:31] good inside of it. Like there are cool primitives inside of the Chat GPT experience. Like the plugins are great, spaces are great, computer use is still the best in the industry. All of it is wrapped together in a product that feels like it is being created by five different companies who are being forced to ship it together in one app. The whole thing is

[0:49] deeply incoherent and doesn't feel good. And where we are in 2026, I wanted to talk about this from one, just complaining about Codex video because I'm annoyed, and also kind of talk about what product looks like at this point because we're seeing this weird split where the big two frontier labs, Open AI and Anthropic, have unlimited tokens on

[1:10] the best models in the industry. Anthropic models are definitely better right now, but you know, Open AI had Astra 3 months early, something like that. I I think they've said publicly it was somewhere around that range. I don't know the exact number, but they had it for a while. All the stuff that they shipped at Dev Day was mostly built by Astra, and yet still the products aren't great. Claude Code has gotten way better. The CLI is amazing, but it took a while to get there. And the only piece of the Claude stuff that really feels good is the CLI and Claude tag because those are the things that the Anthropic

[1:42] employees actually use. Outside of that, as far as products go, the companies with the best models who you would think would have the best products kind of just don't. They're really not winning on that front, and I want to go over why right after today's sponsor. Today's sponsor is basically the only way I let my agents search and grab data off the internet. It's Firecrawl. They have an incredible set of endpoints for grabbing all the data you need and also have some of the best surfaces for actually interacting with and using these endpoints. Be it MCP, CLI, API, SDK, whatever you need, they've got it. And

[2:16] they also have this new thing called Alexandria, which is a set of curated, very high-quality, powerful data sources that they have collected that you can now just actually query from your existing Firecrawl connections that makes your agents a lot more powerful and unlocks a lot of very weird things you probably wouldn't expect. Here's an example. I just asked my agent about what the recent SF startup activity has been and it gave me a breakdown based on YC, Apollo, people data to see what the different categories are, how it's all working. And that's just one of 177 different high-quality providers that they have with over 113 million indexed

[2:52] sources. This is such a useful thing to have access to. When you combine it with the fact that this is already the best way to interact with the web, it's a no-brainer. Go check them out at data7.link/firecrawl. I got inspired to make this video after crashing out when going back to use ChatGPT desktop for a bunch of tax stuff I had to do last week. It did the job, it works, the computer use is still great, but there were so many weird little things I ran into that it just it pissed me off. The first big, like, obvious issue is just general performance UX weirdness where, you

[3:21] know, you go and you load a thread, it's not nearly as snappy as it should be, especially an older profile. Like, that pop-in is not great. There's a lot of little things that just aren't quite right. They also, you know, you have the sidebar here which, look, I'm going to have to do this a lot in this video. It's going to hurt me more than you, but I'm going to have to be very nice to Theo here and say that, yes, T3 code is much better in most ways and I am using

[3:44] it for most of my stuff at this point. One of the things that Codex did, and they've said this publicly, they copied his new priority view sidebar and did not learn any of the right lessons from it and as such their priority view is a dumpster fire. It is not what the priority view is supposed to be. The priority view is supposed to be a to-do list of the threads that need attention so that you can go work on them and then when you're done with them, you dismiss them and they disappear down into the void. In Codex, that's not how they work. They This is basically just a

[4:12] timeline of what you have worked on in order. It just pops up your thread this way. It's not great, but it's still better than this. This view is so archaic. It was fine at the time when we weren't doing that many agent threads, but with the way parallelism works at this point, having to have like search around for the projects and move them up and down and it it just doesn't work anymore. This isn't how we do things unless you are using five gigantic pinned hell threads, which I think is how OpenAI does things. I think that's the reason why this hasn't improved is because internally when they use these

[4:41] things, they use it on the gigantic hell monorepo from OpenAI and they mostly do hell thread stuff. So, they don't notice a lot of this stuff. That's a very common theme that you'll notice in these products is even though yes, again, unlimited tokens, the models are not good at making open-ended product decisions that require, for lack of better word, human taste and discernment and decision-making. Anything that can't be expressed with a clean verifier function, the models still just aren't amazing at. Opus is the best at it of anything, but it's still not perfect or there. You know, they're solving these incredible novel math problems. They're grinding down full rewrites of gigantic

[5:17] projects in Rust, but all of these problems can be expressed with some sort of hill for them to climb. You have a test suite, you need to get them all green. That is a goal that the model can work towards. It can loop and run in circles until it optimizes itself to the point where it produces something that completes the goal. It's happy, it's done, cool. You can't really do that for product experience and UI. It's very hard to encode that into a verifier function. Maybe eventually someone will figure it out, but at least right now, they haven't figured it out and as such, this doesn't feel great. And then once

[5:46] you get outside of this, you run into a whole host of issues. There's one up here, the whole dots thing. I've already talked about this in other videos. I don't want to go down that rabbit hole, especially as I've gotten deeper and deeper into using Grok Bot. Like, I've pretty much centralized everything into it. It runs most of my stuff at this

[6:01] point. Dots feel like an archaic mess. Like, it is not a good experience to have one single thread for everything. It's not how it works. Having a bot for every individual little thing that I do, like for example, you can see over here, I have a ton of different bots for all these different things I work on. You know, you got projects up here, you got shopping, my company, agency, personal finance, personal stuff, YouTube, my main orchestrator bot, which is what the dot is, but the dot doesn't have all these. And by not having all of those, organizing and working with it long-term, it's just not a good

[6:32] experience. And also, just the polish and UX on Grok Bot is second to none. It is such a well-made, well-thought-out product with good primitives that I think is a great case study for if you burn unlimited tokens, you can still make great stuff. Being vibe-coded does not inherently make something bad. If you put the time and effort in, you

[6:53] can make something really, really good. You just have to engineer and architect these things in an intelligent way for that. You need to have easy ways to check and verify changes, to go try stuff, to have a coherent top-down product vision of what it should look like. And that is the thing that Codex and ChatGPT do not have right now. There is no cohesion across this product stack whatsoever. Uh there's the split up here

[7:16] of Codex to ChatGPT. Oh god, what even happens if I go into ChatGPT? I I don't even know. I I don't want to find out. We're just going to go back to Codex and pretend I didn't see that. So, you you got Codex and ChatGPT, that split, which already is very confusing to most people. I I get the vision here that they want to have the same agent be the thing that

[7:35] runs everything, which is Codex. If you want to do that, great. Just make it ChatGPT powered by Codex. Just power the number chat GPT with Codex. Don't do this work split. There There's like a chat and a work and a Codex. It's so weird. Just make all of them Codex under the hood. If I send a prompt that is a bunch of coding stuff, it'll do coding stuff. If my little sister sends a prompt that's asking about basic

[7:56] schoolwork, it'll just answer with that. It shouldn't have this weird split between the two. It just doesn't work or make much sense. And then, once you even get out of these and you start going down the depths of the sidebar and all the different random things they have, you end up with OpenAI DevDay 2026, which I haven't made a video on. This isn't a video on that, but I guess I'm going to have to talk about it because I think this is the perfect example of where things are kind of going wrong at a product level for OpenAI. Not going to go down the model rabbit hole in this

[8:23] video. Obviously, Opus is eating the world and Fable will soon eat the world again. I They're going to be in a rough spot for a while. I I wouldn't underestimate them or count them out. This has happened before. The model cycle is real. Codex and chat GPT will be back in a very big way by end of year. I mean sure. Again, I'm getting ahead of myself. I don't want to go down that rant. What I want to talk about is the insane number of products and things that they shipped and did in DevDay that I'm pretty sure 99% of people have no idea exist and don't even realize. I was

[8:54] at DevDay and I have been learning about new things as I was researching this video. There is so much stuff that they've done with no coherent through lines whatsoever. I I get that product discoverability is a hard problem these days, but you do a lot better than this. At the top of their feature list, they have, of course, Dots, the big headline announcement that, as I said, did not go all that well. Then they had 6.1 Soul, which, again, cool model, great price, not that interesting when Opus 5.5 exists. You have ultrafast, which is nice. It's cool, but it's A limited to the $500 a month plan, and B is way too

[9:29] hungry on your usage limits. Like even when you're using 6.1 Soul ultrafast, which came out, it's great, it's fast, it still incinerates the usage on your $500 a month plan. It doesn't feel all that viable to use. Private intelligence, great enterprise feature, but not what we're talking about here. Then you go down to Codex in the cloud. I don't know anyone who has really done much of anything with their cloud stuff because frankly they're pretty behind on it. If you want a hosted cloud solution, Cursor is currently the best option. If you want an on-prem custom self-hosted cloud solution, T3 Code is the best

[10:01] solution. This just doesn't quite get there. They released the CLI, which is actually awesome. This was a great release which will flew super under the radar. I found out about this actually cuz I was talking to some of my friends at Open AI and they're like, "Oh yeah, you should go try the new CLI." And it's awesome. I've used it a ton. They did really good with this, but it got buried under I think there's like 30 to 40 announcements. I'm not going to go through all of them cuz again, there's just a lot of stuff. Like there's the code review, uh security, decisions API, their Jeff killer. Again, great ship,

[10:29] but nothing groundbreaking. Uh API computer use also really good. Bedrock. But then you get down into like a lot of the product stuff here. And Spaces was one of the big new headline features that they shipped. And it is quite good. I actually am using it a lot. I'm actually I planned this video in Spaces. I've been using this for managing and working with my channel. I have very repeatable processes for things. And one of the things you can do with Spaces is if we just go into this video, you can see I have all this stuff here which is copied out of a template that includes a

[11:00] bunch of steps written out for how to prep a video in Final Cut. And the reason I have that is because you can save prompts inside of it to fire off into a Codex thread and make those happen. So, my editor can go into our shared Spaces here, go into this video. He can hit the prompt here, prep the video in Final Cut just with this pre-rolled prompt. And then I have similar things for thumbnails. I'm using this style lately. I have this codified to like pull out a bunch of screenshots and faces that I could use. Checklists for uploading the videos, description templates, all that stuff. All set up

[11:31] and working in here. It's quite nice. It's basically notion with a bunch of AI stuff built into it with a lot of rough edges around it. The composer for working with these things is a normal codex thread, but it's also not because it's super closely tied into one of the space pages or one of the pages or I don't know what the terms here are honestly. I use this and I don't know

[11:50] what it is. Uh I think it's tied to this page. So, if you send a prompt, it'll make some changes, but when it makes changes, it will force your view, your entire computer back onto that page to focus it so that you can watch the model make the change for some reason. I don't know what reason. Really, really annoying there. And then when you go like actually try to view one in the sidebar here. So, I'll go back to priority view. I go to prep a video on final cut. It'll open. So, this is the thread. It's here and I'm on this page, but then if I hit this, I can open it in

[12:19] full view and it's a normal codex thread, but also it's here and it's just it's weird. It it feels very very much like you have the threads and you have the spaces and they are trying to stick them together and they're just not doing it very well, unfortunately. And that's the theme with all of this. The cohesion just isn't there at the product level, especially across such a big and ambitious thing like codex because they are trying to build 30 plus products to work for everybody. I got a really funny text from my little sister that partly made me realize how much of a bubble we're in, but also just how

[12:56] weird the UX of this is for someone who doesn't live in the stuff 24/7 the way we do where she was asking me like, "Hey, how do I use a different model or like not run out of usage?" And I told her, "Okay, you hit GPT-6 Astra to change the model, then drag the slider to the second position. WTF is the

[13:14] difference? I need my homework done. This stupid bot won't do it." And I just had to explain, "No, it's a smaller, cheaper model." Like they don't know what the difference between a 6.1 soul and a 6.0 Astra is. No normal person knows the difference between these two things. We do and we care and we need to have these things available in such a way that we can toggle whatever model we want and make it work. But normal people don't, which is again, why I think they tried to make

[13:39] this split between Codex and ChatGPT. But ChatGPT isn't powerful enough to do a lot of the things that Codex does. I It showed her Codex because she had to do some stuff with spreadsheets and I'm like, this is like the best way to do spreadsheets, you should use this. Blew her away, uses it a ton and now the whole thing just doesn't make a huge amount of sense. And all these weird confusions and edges and just cohesion problems are very symptomatic, I think, of what happens to a lot of companies and products in 2026. There's been a lot of, you can call it enshittification, where things just keep getting more and

[14:12] more features and random nonsense. It's never been easier to ship a feature, to make something new and put it in production. That is very, very easy to do. The models can do it. You give a model a task, it will perform that task. The problem is, they're not going to do it in the best possible way and a lot of weird edges are going to naturally start to form. If you don't have a very strong vision and especially way to say no to things. Codex and ChatGPT are not the only example of this. They're just the most glaringly obvious one, which is why

[14:39] I'm using them and dunking on them here. But this is happening everywhere. Very few companies are doing the whole like using the agents to build really fast, make something much higher quality than they could before while not falling into this trap. I I like Grokbot for an example here a lot because Grokbot is in a ton of different places. It is in my Tesla, it is in my desktop app, it's in a mobile app, it's in XDM's now for some reason. It is in

[15:07] the grok.com website. It is everywhere. And this primitive of a bot that they created has been reused and works super, super well across all these different services to where it always feels polished and clean and first class in a way that I haven't really seen for much of anything else. I I've said this multiple times, but SpaceX has the best engineering of any of the labs right now. They are really, really good at product and making things. They just kind of suck at research and models, which is, you know, it will be an interesting thing long-term to see whether or not product is enough to make

[15:40] up that difference. Maybe Grok 5 can get there. I'm very skeptical. I'm not really sure if they're going to get there, but right now we're getting to see Grokbot with Opus 5.5 running as the model powering it is insanely powerful. I am using this, tons of people I know are using this for very heavy, useful work, which is obviously what they wanted dot to be, and dot is, well, was

[16:02] powered by a better model for a while. This had Astra when Grokbot just had Grok 4.7, and I still went to Grokbot instead because the product was so much better. I don't know who's going to end up winning in this split here. If the models ever get to the point where they can make product decisions on their own, like OpenAI can just tell GPT-7, "Go make product cohesive and good." Then that's just AGI and it can RSI itself and it's over. Like, at that point if the models are capable of that level of non-verifiable reasoning and decision-making, then they're intelligent and they have actual Sentience is maybe the wrong word, but

[16:35] they're they're good enough to improve themselves at that point because that's kind of the thing that's missing. Anthropic has posted about this a lot with their risk reports on RSI, where the big thing that's missing now is this like mystical researcher taste making these very difficult calls, these fork in the roads that aren't super verifiable things. That's a whole 'nother rant for another day. Just for today, my point is the products don't have to be like this. You can both use the agents and get a very clean, well-made, polished products. T3 Code,

[17:06] again, is another great example of this. Julius and Theo and all the contributors are burning obscene amounts of tokens on T3 Code, but it is still improving over and over and over again because there are clear dictatorial product decisions being made repeatedly that are moving it in one direction that actually makes sense. Codex isn't. So, will OpenAI be able to fix this eventually? Probably. I'd be surprised if they didn't. Again, underestimating OpenAI is almost always dumb. People like doing it and then

[17:36] Astra happens. Like they're very good. They will figure this out eventually. If I were in their shoes and I was tasked with, "Okay, go fix Codex and ChatGPT." What would I do? What would be the first changes I would make to fix this? I think the first thing I would do would be gutting most of it. Like I I would start cutting features before I made new ones. There are a lot of things in here that should be much cleaner, better primitives that just aren't. A Codex thread should become the base building block for everything. That should be the building block for the CLI, for the

[18:08] desktop app, for the web app, everything. Everything should be a Codex thread. Dots should be turned into effectively named Codex threads which have attached to them routines, memories, connections, whatever you want to have in there. Very similar to what Grokbot does. Grokbot, for those who don't know, is Cursor Cloud agents. These are Cursor Cloud agents. It was developed by Cursor quite a while ago in parallel with Cursor Cloud and that's the reason why this works. These are Cursor threads. You get like one Cursor Cloud VM for your Grokbots. And through whatever magic they're doing under the hood, they're able to make this core building block primitive that works really, really

[18:44] well. And that's something that Codex does not have right now. So, I'd fix that primitive. Then spaces, I would just do an insane amount of polishing. Once you have that primitive really locked in, then I think it becomes a little bit easier to then make spaces a separate primitive sitting over here. You have your threads, then you have your spaces, and then a space has like all these different pages. That's like the building block in there. That is like an entity that can then be manipulated by a thread directly whenever you want to. Should also be very easy to take a thread and throw it in the cloud or have it local. All of

[19:15] that should just work nicely. Spaces can easily be controlled headlessly. Right now they can be, but not really. I wouldn't have threads tied directly to the spaces. Like keep this composer view, but when you do this, it just opens a normal thread. And then that thread you can just type, do whatever. It just has this in context as part of the prompt because you fired it from here. And then honestly just kind of

[19:36] keep iterating and building from there. But they they need to fix the foundation first, and the foundation is fundamentally pretty broken right now. I trust they'll do it. There's tons of talented people there. They'll figure something out here. But for now, yeah. Enjoyed this video. Make sure you like and subscribe, and I'll talk to you

[19:52] again soon.

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt