# EP #40 The Two-Billion-Token Weekend That Cost $125 (For Now)

“EP #40 The Two-Billion-Token Weekend That Cost $125 (For Now).” AI at Work by AscendAI, 2026-09-04. https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/

Credit the named speaker when quoting a passage, and link to the episode or the passage.

Transcript licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

### [Kevin Williams · Passage 1](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-1)

Hey Matt, is it wrong that I'm sort of obsessed with my own avatar in the intro? I mean, those of you who are listening don't actually see it, but but he's a silver fox. And you know, he's gonna he's a good looking

### [Matt Graham · Passage 2](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-2)

Yeah, he's a good looking guy, not gonna lie.

### [Kevin Williams · Passage 3](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-3)

guy. Like, I don't know. This isn't good. Yes, yes,

### [Matt Graham · Passage 4](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-4)

I'm gonna be really disappointed when I see you in October in person.

### [Kevin Williams · Passage 5](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-5)

exactly. I need to have like like silver tipped ends and spiky hair and

### [Matt Graham · Passage 6](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-6)

Yeah.

### [Kevin Williams · Passage 7](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-7)

probably a tan, and yeah.

### [Matt Graham · Passage 8](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-8)

Well, on on that topic, there are so many, I think it's a really fascinating space, these like influencers that are fake AI avatars. you know, and I think I I think there's like two groups. The ones that are very clearly transparently obviously AI, and then there's like a subgroup where they're trying to, you know, kind of come off as human, but they're actually not. I think most most of the influencers, AI influencers out there are obvious and transparent about it. But I mean, I I keep a tra I keep a eye on them and they are getting really good, Kevin. Like really good. It's scary good.

### [Kevin Williams · Passage 9](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-9)

They are getting really good, particularly in consumer, where it's like skincare or it's health care and or or you know, whatever it is. And they these things can really adapt to signals as far as what the marketing environment is yielding. And for you or I to create, you know, 10 different marketing spots, even using the tools that we have at our disposal is a decent lift, and then to respond, to look at the data, to tweak it in the other direction. But if you can put out a hundred avatars and get five positive signals, and then you put out a hundred more avatars on those five positive signals, you know, eventually you're gonna find some green fields for that stuff.

### [Matt Graham · Passage 10](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-10)

Yeah. And it so I think there's a scalability. Like we know this shooting content. Like, you know, myself personally, I have to shoot maybe on like Saturday. It's like a second part-time job for me to like shoot content with my team here in Boston. and yeah, you're just limited. So like having the ability to like instantly spin up, instantly test things. I guess what makes me nervous though is I think there's a certain segment of the population that doesn't know the stuff is not real. you know, like older crowds or You know, people that just are not in tech at all and they just don't know that much about AI in general. Like I think these things can really fool a lot of people. It'll be interesting, like election cycle, like if any of this stuff gets used. Yeah, yeah, true. True.

### [Kevin Williams · Passage 11](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-11)

Everything will get used. I'm sure like all bets are off. I mean, on the one side, one reaction is, does it matter? Does it matter if it if it is creating the education, the the entertainment, whatever the value is that's being put out there? You know, it's not like scripted television is real. Like there's a lot of not real things that are out there. the problem becomes sort of that uncanny valley of trust, I th I think. I'm actually I'm curious. so on this show in the past, we've talked about some of my experiments with avataring myself, in a genuine way, like not to like

### [Matt Graham · Passage 12](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-12)

Mm-hmm.

### [Kevin Williams · Passage 13](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-13)

create media, but but say for courses. So we do a lot of courses, and ironically, I think I'm wearing my course filming black polo sh shirt today, but Like if I need to correct a two minute, three minute piece in one of those cont pieces of content, it would be awfully nice to be able to use an avatar. And I've I have yet to find a platform that really does a great job of self avataring. Like they have their own avatars that's that are like

### [Matt Graham · Passage 14](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-14)

Mm-hmm.

### [Kevin Williams · Passage 15](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-15)

trained and they're pretty decent. I don't know. I'm curious. Have you tried that? Have you tried to build your own avatar?

### [Matt Graham · Passage 16](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-16)

Yeah, so I've been trying this for like three years. because also like when you do cold outbound or you do any type of like solicitation, it's helpful sometimes to have a video, have a voice, and it's hard to do that at scale. And so when these tools first came out, we were very big on like eleven labs, hey gen. and like you said, in those days, it was much easier just to have a fake avatar than to try to make an avatar of yourself. And so those two things I think most of the influencers out there, we see them just inventing a new avatar. And that's a s way higher production value, way like looks way more real. And as soon as you try to make one of yourself, the quality always goes down. So I've been checking like every eight months to see, you know, has it improved? Has it improved? So far, the only way we've actually used this stuff is let's say we have a short of me talking, but there's like a particular clip that's missing, like and like a filler. And it's just a s you know, maybe a microsecond, you know, when it's jumping from screen to screen. And so we've had some success with that. But if you pause it right at that microsecond and you look, you can see some details in my face that aren't real. So it's not to the point where it's good enough for me to talk straight to the camera, not even close, but it's it's been some filler gaps on some of our content to just make the storytelling a little bit better.

### [Kevin Williams · Passage 17](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-17)

So what are you using? So I you you mentioned Hajen, we use we've used Synestia in the past, we use Descript. what's sort of your go-to tool there right now?

### [Matt Graham · Passage 18](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-18)

Yeah, so Hey Gen, like we you know, Hey Gen has the ability to do like a really professional, like, you know, set up your best camera you got and like record everything. Same thing with eleven labs, like get a professional mic and record something for several hours and it'll try to clone your voice. What the problem is, so it's mostly like Hey Gen and 11 Labs that we've stuck with. The problem is is just the variability is very wide. You can get like a few clips that are pretty good, but if you need to make a thousand clips, You start to see a wide range of different outputs. And you're like, who the heck is that? Or the an eleven labs with the voice, you're like, that is I mean, the first couple were good, but then you know, f number five hundred and sixty-six sounded like a totally different person. so it's interesting because you want to add some variability to your settings to be able to get get it to like pick you up well. But then when you do that with your settings, it actually creates like a ton of different outputs that are really off the charts wrong.

### [Kevin Williams · Passage 19](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-19)

It feels like a category that's just going to keep getting better, but it felt that way in two thousand twenty three too. So

### [Matt Graham · Passage 20](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-20)

Yeah, exactly. That's the same thing. Like I've we've been trying for three years to get this like production ready and it's still like only good for little tiny clips and that's it. On the video side. I think the the audio mm audio's still like only so so as well, I think, when it's trying to replicate my actual voice.

### [Kevin Williams · Passage 21](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-21)

Yeah, and another another place is video editing. So you know, we still use a human editor of as far as and and no offense to the human editor, Gabriel, you're good, I promise. That Gabriel,

### [Matt Graham · Passage 22](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-22)

Love you, Gabriel.

### [Kevin Williams · Passage 23](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-23)

yeah, yeah. Heart side to Gabriel that but still I'd like I'd like to make him more efficient such that some of the slicing and dicing of these videos and other videos w can go faster. And we've tried

### [Matt Graham · Passage 24](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-24)

Mm-hmm.

### [Kevin Williams · Passage 25](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-25)

Opus Clip in that respect. you know, some other tools. And it's just, you know, a lot of it just isn't quite ready for prime time. That said, on like major media side, there's so much money that's being invested in these technologies that I think That even though we've got a bunch of like AI market leaders that popped up that sort of service this sort of need. I mean, Hey, Jen, very well funded, very successful. Eleven Labs as well. you know, Descript and Riverside and all of these, it's probably gonna be something that creeps out of Hollywood that's like a dedicated, very well funded thing that just appears. whether we mere mortals can afford it, I think, would be the question there.

### [Matt Graham · Passage 26](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-26)

Well, I I think it's interesting. It goes back to what we talked about last week, how we were talking about real real quick recap. These models solved a bunch bunch of mathematical problems that humans haven't been able to solve for a hundred years. And your point, Kevin, was like it's very interesting that these models have been trained on a wide range of different types of skills. And even though it's maybe not the, you know, wasn't trained to be a lawyer, trained to be a doctor, because it was trained on such a wide variety of skills, it's actually able to conquer all those different types of jobs and different types of expertise just by being trained broadly on many different types of skills. And so I think that that was very true statement and insight you had last week. But for mul for different types of modals like voice or video, I think we'll have a completely different type of model and business that you know succeeds in those areas. So I agree like probably, you know, Claude and Open AI will get really good at law or medicine, but I think when it comes to video and voice, like we'll see separate companies that dominate in those areas.

### [Kevin Williams · Passage 27](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-27)

Yeah. No, it's it it's super interesting in you know, video games, et cetera. Like there's there's a lot of really dynamic stuff going on. And as much as yeah, we just said, it's in the dim days of 2023, like if you look at your watch, like a not a lot of time has gone by. It just feels like some of these things plateaued a little sooner than than maybe I would have expected. I'm still very, very excited and very bullish over the next few years. But there's a lot of maybe technical debt and debris and just market confusion and you know, I think that users are really quite overwhelmed with all of the signals that they're getting that they're just sort of hunkering down and waiting for for something that's really clearly valuable.

### [Matt Graham · Passage 28](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-28)

Yeah, in the agentic AI space, that's how I feel. Like I feel like a lot of it's good, but it's it's still not the ninety-nine point nine nine nine nine that you need for every step that we always talk

### [Kevin Williams · Passage 29](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-29)

Yeah.

### [Matt Graham · Passage 30](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-30)

about, the deterministic ver versus probabilistic. And I still think like the limiting factor is the data. Like they don't have the data they need to train the a computer use, agentic AI, like long horizon tasks to, you know, really impact the way employees work.

### [Kevin Williams · Passage 31](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-31)

So let's talk, let's just talk a little bit about the state of the state of agents, I suppose. So 2026

### [Matt Graham · Passage 32](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-32)

Yeah.

### [Kevin Williams · Passage 33](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-33)

is is well, 2025 was the year of the agents. 2026 continues to be the year of the agents, and we keep seeing developments in that respect. So as a little bit of a refresher, just in the last few months, we've seen the rise of the agentic harness, and that as the tooling that's built right within anthropic or open AI that allows you to create individual workflow type air quotes agents that go off and and do things on on your behalf. This was all prompted to use a term by you know OpenClaw at the beginning of the year that had this promise that it would go off and do things on your behalf. OpenClaw kind of had its its its moment of fire, and apparently Apple made a hundred million dollars on Mac Minis because everybody suddenly needed a Mac Mini to to run their OpenClaws. And then it morphed into Hermes, and then the more sort of air quotes mature platforms like OpenAI and and Anthropic started to develop these harnesses. And now you have we mentioned this last week just for a second. But you have Grockbot that's come on stage, and there are a bunch of other versions of this where the agents are sort of ready to do your bidding at a moment's notice. I'm curious on your take on what you're seeing. So you you know, as a as a dev group, you're constantly being asked to automate things, but

### [Matt Graham · Passage 34](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-34)

Mm-hmm.

### [Kevin Williams · Passage 35](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-35)

as a responsible dev group, you're also doing so as deterministically as possible. So where do these freewheeling type agents fit into the grand scheme of things? And I have a on this too.

### [Matt Graham · Passage 36](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-36)

Yeah, it's super interesting. I don't know if I can tell you who I chatted with, but a big, big company, been around for a long time doing workflows and agentic workflows. What they see, the stat they gave, and they have a huge, huge customer base. So I think this is a very, very data point you should weigh heavily. You know, most data points I take with a grain of salt, but they said eighty-two percent of workflows are still if-then statements or deterministic. And so only really eighteen I 82% is that 18% is the agentic side. And so it's still not able, you know, like the deterministic part is still the vast majority of a workflow. And then within that 18%, there's still, you know, probably a lot of it is very sensitive. Like again, if there's too many of those probabilistic steps, they're not going to execute the entire workflow well. So I still think we're kind of at intern level, Kevin. Like even I you know, we've had marginal improvements and I think what like its ability to use leverage a harness and go to all the different places it needs to go to get information to to to to take the next step in the tasks, all that connectivity has improved. But I still think it's making too many mistakes in a l in a lot of situations.

### [Kevin Williams · Passage 37](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-37)

So on the flip side, the construction of a deterministic agent has become quite a bit easier. That using generative AI to build the the structure that you need to do a repeatable deterministic workflow is so much easier than it was, you know, even eighteen months ago. I know we both have had hey, we did our time in the N eight N trenches or or Zapier

### [Matt Graham · Passage 38](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-38)

Happier.

### [Kevin Williams · Passage 39](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-39)

or Make, you know, et cetera. And I It's been a minute since I've heard anybody talk about N8N in particular, because it just got eclipsed by the ability to connect all of the digital dots. Because as fun as it is to be able to drag and drop like all the interfaces, you don't really care what the indeterminates the intermediate steps are, as long as the result comes out the other way at the other end.

### [Matt Graham · Passage 40](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-40)

Mm-hmm.

### [Kevin Williams · Passage 41](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-41)

If all of the mechanics under the hood are if-then statements, who cares? that it's still a workflow that's accomplishing something. The the art and maybe the risk comes in where does that probabilistic step live as far as looking into that workflow and where a decision point might need to be made or a qualification point and et cetera. And you know, it gets scary depending on on what industry you're in. If you're doing marketing content, not a big deal. If you're doing healthcare or finance, like it's suddenly a very big deal.

### [Matt Graham · Passage 42](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-42)

Mm-hmm.

### [Kevin Williams · Passage 43](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-43)

you know, going back to what you were saying a minute ago about repeatability. you know same is true with with workflows. If it can't do the same thing day in, day out, and it requires bot sitting to get good results, you know, I don't I don't see a lot of more traditional organizations leaning into a solution like that.

### [Matt Graham · Passage 44](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-44)

Yeah, I'll give a another quick example. I was going to a conference, I had a list of people. I told my my bot to go find everyone's LinkedIn profile. Next thing I know, my LinkedIn account got banned. So, like sometimes, you know, sometimes the bot does do a good job. This is the counterpoint. And it just did it too fast and like too much volume. You know, LinkedIn has a lot of limits and they're always monitoring for stuff like this. so I it was silly of me to not think like, yeah, I probably shouldn't have my Baco look up like two hundred profiles in like five minutes for me on my LinkedIn account, you know.

### [Kevin Williams · Passage 45](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-45)

Yes. That yes. Like working with Meta, working with LinkedIn, even working with Google, like to some degree it's like diffusing a bomb. And like, you know, you're not trying to do anything particularly skeezy, but it's pretty easy to f yeah.

### [Matt Graham · Passage 46](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-46)

Yeah, just like I'm going to a conference. Who are these people?

### [Kevin Williams · Passage 47](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-47)

It's and and you know, you could have had a virtual assistant or somebody like running out and doing the same thing. but yeah, so for the audience, it is very easy. To get tripped up because a lot of these odd these bots are running through detectable IP addresses. And those, you know, LinkedIn knows that it's a bot. And even though you think that you're being pretty clever with the way that your bot is operating, just be very, very careful. You know, at this stage of my career, there are certain assets that I I hold close. And one of those is probably LinkedIn because it's my as obnoxious as the platform is. It is my you know, my business card to the world. And if that just went poof, I think that that would be a problem for me.

### [Matt Graham · Passage 48](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-48)

Definitely, definitely. It is not fun for anyone out there who goes through this. If you get stuck, let me know. I can help you. But yeah, getting your account unbanned is not a fun process.

### [Kevin Williams · Passage 49](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-49)

Yes, yes, indeed. Yes.

### [Matt Graham · Passage 50](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-50)

LinkedIn jail is a real thing.

### [Kevin Williams · Passage 51](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-51)

Yes. And I think this is just going to get worse. So I I've been to a couple of different discussion, like marketing discussion groups in the last coup couple of weeks about LinkedIn. And you know, it's existential for them. It really is. Like it's just

### [Matt Graham · Passage 52](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-52)

Mm-hmm.

### [Kevin Williams · Passage 53](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-53)

apparently posts are up by something like eighteen percent and engagement is down by twenty-five percent. So there's Yeah, and that's, you know,

### [Matt Graham · Passage 54](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-54)

man.

### [Kevin Williams · Passage 55](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-55)

if people aren't actually using the platform for information gathering anymore because of the level of slop, and now you have this new the slop reporting tool that's in there, et cetera. but they're also working on their business model, like pretty clearly. And organic reach isn't what it used to be. Now you have a massive LinkedIn following, so I'm actually curious on that, if you felt some of the pain around that. that you you spent all of this effort and all of this time to develop a really large following and now are you being rewarded to the degree that you expected to be rewarded by that following.

### [Matt Graham · Passage 56](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-56)

Yeah, I've so I've I have a love-hate relationship with LinkedIn. First, real quick on that conference thing, like it was real at least I got all the LinkedIn profiles before I got banned. And it was super helpful at like identifying, like, hey, these are 10 people and these are like the interesting conversations you could have with them because of the what business they're in and how it overlaps with you. So it was like, wow, I got like so much more out

### [Kevin Williams · Passage 57](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-57)

Mm-hmm.

### [Matt Graham · Passage 58](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-58)

of that conference by taking 10 minutes to prepare a little bit and having AI guide me. So that was amazing. But then you have these problems that we just discussed about LinkedIn. It's very, very limiting. you know, I think the problem is LinkedIn grew up as a recruiting platform. Like it was a platform where you only engage when you're looking for a job. And then once you have a job and you're not looking anymore, you no longer spend any time on the platform. And that really started to change over the last two years, where they, you know, social media has just become so, so, so important as an acquisition channel for so many businesses. And LinkedIn is the premium B2B. Acquisition channel. Probably a lot of folks listening to this run B2B businesses or know how important LinkedIn is and how important it is to be active on those on that platform. The problem is, you know, everyone who's ever been on LinkedIn, you know, they have that recruiting mindset there for a job, and a lot of the masses are not engaging for business value or business learning. I think that's changing slowly, but they gotta improve the engagement. The comments are so bad. Like ridiculously bad, Kevin. The conversations that happen on X. I mean, there's a lot of trolls, there's a lot of noise and nonsense, but there's like good conversations happening if you dig a little bit. Whereas, I mean, you kind of gotta dig a lot on X. But LinkedIn is the total opposite. It's like the dumbest comments I've ever heard. It's like so terrible. Really no value add. It's awful. I think the content though is getting better. And I think if they keep promoting people like I I I think my content's good. You know, I get a lot of likes and a and some pretty decent engagement, even though I don't think the the comments are that valuable. The likes are are pretty good for most posts. so I think keep pressing. If you're out there making LinkedIn content, keep it going. And the platform is gonna improve. And I think this is important for them. I think they know they have a big opportunity ahead of them as like the leader for business value type conversations and content creation. Yeah.

### [Kevin Williams · Passage 59](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-59)

All right. So I wanna I wanna pivot a little bit to to my weekend adventures here

### [Matt Graham · Passage 60](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-60)

Yeah, yeah.

### [Kevin Williams · Passage 61](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-61)

because for our for our builders that are out there, you know, I've been an adamant fan of Claude Code and really invested a lot of time in refining my skills and really trying to be as productive as I can. And you know, remember I'm not like that my my day job is not as like, you know, an AI influencer. So I have to accomplish things, right? I I have some luxury and some need to constantly be exploring a new technologies. But particularly with a pro pro product productivity tool like Cloud Code, I was reluctant to experiment in too many other directions. yet I was seeing all of the noise around codecs. And that's GPT's coding engine. And Well, I've tried Codex in the past. It was totally fine. you know, sort of a coker, a Pepsi sort of approach. And then because Claude got pretty cheap on its tokens over the weekend, I just simply ran out. I fully committed to a full weekend of Codex tokens. And do you have any idea?

### [Matt Graham · Passage 62](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-62)

Wow, you went to the dark side, Kevin.

### [Kevin Williams · Passage 63](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-63)

Do you have any idea how much so so it was a rainy weekend in town, like absolutely pouring rain, completely miserable sort of day? And I was like, you know what? I'm just gonna spend all weekend like building. There are three different things that I want to build. I'm gonna build all three of them in codex. And it was kind of ridiculous. I think I I I probably worked 20 plus hours over over two days of running like up to like 10 simultaneous agents. Do you wanna guess how many tokens that I managed to burn in we'll call it three days?

### [Matt Graham · Passage 64](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-64)

Mm, a thousand tokens.

### [Kevin Williams · Passage 65](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-65)

Thousand tokens, come on, we burned a thousand tokens just with the transcript tool that's here.

### [Matt Graham · Passage 66](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-66)

All right, Hades, what is it, Kevin?

### [Kevin Williams · Passage 67](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-67)

Two billion tokens. Two billion, two billion. So much value.

### [Matt Graham · Passage 68](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-68)

my gosh. And you created so much value from those two billion tokens.

### [Kevin Williams · Passage 69](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-69)

I think I did. But then I I had a moment. So I want to go into like the experience a little bit, but so I looked it up. I was like, okay, if I was paying for this, like an enterprise would pay for this, or if I was running purely on API. Like what would have this have cost? And it's complicated is the answer because their input and their output tokens and whatever, but I pretty much ran on Seoul 5.6 high. So that's not ultra, that's not the craziest model. That is the kind of the recommended workhorse model. And I did so for you know dozens of hours in parallel processes. it would have been somewhere north of eight to ten thousand dollars in tokens. If at at retail. So if I were in a big enterprise, and I don't think a lot of our listeners in particular understand this because a lot are sort of mid-market or they're sort of smaller organizations where you have a you have a cloud subscription or you have a GPT subscription and you you pay for Cloud Max $200 a month or use something similar for GPT and you get a lot of included tokens. And yes, there are some guardrails around it as far as being only able to run a certain number of hours on a model and et cetera, et cetera. But if you were on an enterprise plan, you actually fall out of that. So when your team is building within Codecs, they're actually paying an entirely different rate of tokens. And by the way, that can be like a negotiated rate. So it's also kind of hard to like figure out what that number is.

### [Matt Graham · Passage 70](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-70)

Apple to apple, yeah.

### [Kevin Williams · Passage 71](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-71)

But you know, for I I spent, I mean, because I hadn't been running a high-end GPT plan. I I spent $125 on my tokens for those two billion tokens that might be worth $8,000 plus, but that enterprise user would have had to spend the $8,000 of tokens on the other side. And that is a major gut check if you're looking at the types of things that you're building, right?

### [Matt Graham · Passage 72](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-72)

So when is this all gonna come to roost here? Cause I mean I I think we're all very thankful that we're all able to get as consumers and businesses highly, highly subsidized token costs. I don't think people realize that I mean your your story alludes to it and it helps illustrate like how much of a great deal we're all getting. And you know, both it can't last forever, Kevin.

### [Kevin Williams · Passage 73](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-73)

I don't know. so so September fourteenth

### [Matt Graham · Passage 74](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-74)

Can

### [Kevin Williams · Passage 75](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-75)

is the first is the first or September fifteenth is the first shoe that's like like written down that might fall. And so on September fifteenth, and something else that users probably don't recognize, is that on September fifteenth, there's a there's a fifty percent subsidy that anthropic has been giving its users on tokens, and that's apparently going to go away. On the flip side, There's no way that I would have made my two billion tokens on my hundred and twenty-five dollars if OpenAI hadn't been throwing reset after reset after reset on on me. So by the end of Saturday, my usage was all the way down to like thirteen percent. And I was like, well, I guess this is sort of the end of the road. And I woke up the next morning and it was all the way back up to eighty three percent. Okay. By midday, I went away for a little while and it was again down to like, I don't know. 10, 11%, and it bounced back up to 100%. And then yesterday I ran out of tokens again and I had a reset in there and I was able to go back to a hundred hundred percent of tokens. So there's warfare going on between the two. And it wouldn't surprise me for an instant if people got so frustrated with Claude, even the new model. So I mean, we're not a model show, but Fable five point one dropped this week and it's really, really good and it's more efficient, but it's still fable, so it lights tokens on fire like crazy. if you're only gonna get a half an hour of use out of your two hundred dollar max plan, like I you know, and and op

### [Matt Graham · Passage 76](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-76)

Yeah, solve a hundred year old math problem that no one solved yet.

### [Kevin Williams · Passage 77](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-77)

Yes. And and open AI is offering resets and whatever for a comparable cost. people need to realize that there is no switching cost. And I think, you know, investors need to realize that there is no switching cost. Because if you're if you're going about this in the right way, and this might be educational for people, like anybody who's building, like make sure you have GitHub set up such that your data is going somewhere and it's stored. So if you want to build a widget that's I don't know, you want to build a translator bot or whatever or a little little app that does something, all the code is in GitHub, right? So all you need to do between having cloud code on one screen and having codecs on the other screen is to connect that repository in GitHub on the other one and say, okay, what are we gonna do next? There's no like lock-in there that that that that the platforms have. And as long as there's that transferability, and then as long as there's the specter of other players in the space. So Grok from XAI has gotten quite a bit better at at coding since that it it they acquired cursor. You can just switch over to grok if you wanted to, or you could go to Kimi K3. You know, it gets a little more complicated with some of the other ones, but there is no switching cost. And there's a lot of you know, sort of hard earned skill around, you know, Claude Code or if you're a Codex expert, that's that's valuable. But I d I just don't know if they're gonna get away with rising these prices as long as they have a prisoner's dilemma going on.

### [Matt Graham · Passage 78](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-78)

Well, and then another thing on top of that, if people even if you're not using this stuff to code and connecting your GitHub and everything Kevin just said, you can also just, you know, kind of export all the context in several different ways. You can just even ask it like, hey, you know, summarize who all everything you know about me so I can import it into another model. and so there's lots of easy ways that take ten minutes to do all that as well. so even if you're not coding, like there is really no barrier. I I wonder to some extent if if these two companies are gonna conc co collude, you know, like if Gemini's gonna exit this race, Apple never gets into it and we end up only with OpenAI and Anthropic, like are these two just duopoly like US UPS USP and or UPS and FedEx, are they just gonna conclude conclude collude on pricing?

### [Kevin Williams · Passage 79](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-79)

Yeah, I'm not sure we should be tossing shade at FedEx at UPS, but like yeah, I I I don't

### [Matt Graham · Passage 80](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-80)

come on. They they raised their prices literally at the same exact time, the same exact amount in Locks State.

### [Kevin Williams · Passage 81](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-81)

I so I I I I don't it's not likely because there there are alternatives. Like how that could happen is through regulatory capture. So if we shut down access to say the Chinese models, which I could think of some national security reasons why that could be a good idea, if we were to do that and the last models standing were OpenAI, Anthropic, and Grok, and probably Gemini, maybe, but yeah, I I don't know. we'll see. We'll see it play out. I mean, most of the community it just sort of assumes that everything's going to get much more expensive. And I think that there's a psychological part there that's interesting, which is part of why I I spent my weekend coding, is there's a bit of FOMO. And I I really want to make sure that I get some of these like foundational things for my organization built.

### [Matt Graham · Passage 82](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-82)

Mm. Before

### [Kevin Williams · Passage 83](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-83)

Such that I can do it now while the cost is still in alignment in case it goes up. Because I have to ask myself, would I have spent $10,000 on what I built as an entrepreneur? It is all going to be very valuable for my organization. Some of it is sales fronting, some of it is operations. But would I have gone out to Rapid Dev and said,

### [Matt Graham · Passage 84](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-84)

Mm-hmm.

### [Kevin Williams · Passage 85](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-85)

hey Matt, for $10,000, I'll I'll, you know, I'll write you a check to build this. It would have been a hard bet for me to do that. because this stuff falls into that category of new capabilities. I've said this before, but AI can kind of do three things. It can replace tasks, it can augment tasks, or it can create new capabilities. So those of us who are trying to create these new capabilities in our organization, there is a lingering fear that that that to get it well, the get it good, the going's good, and to build that infrastructure such that I can tweak it later at a new cost structure. But building it from Go might be just too expensive.

### [Matt Graham · Passage 86](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-86)

Here's another thing, like it we already built a bunch of stuff for ourselves internally. We have this we call it you think of it as an AI brain. It captures every call, every Slack message, every email we have with a client on a project. It segments segments them by project. and then it gives us alerts if something's off the rails or it gives us like things that we can be proactive about, like a quarterly Business review. Like, hey, here's some topics to cover in your quarterly business review based on the quarter's conversation. So extremely powerful tool for client work to really provide the best service possible. But like that thing all runs like pretty affordably right now. Imagine if our entire business has been built around this giant AI system. It's not even the building costs that was is at risk, that FOMO that you're talking about. It's now like if all of a sudden our cost of running the system gonna go up by three X. You know? Yeah.

### [Kevin Williams · Passage 87](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-87)

Might. They might. And that's you know, that's that's a that's a bit of gut shagged. It's funny you say that because that's actually what I was building internally as well. Is is something every agent should say this. Yeah.

### [Matt Graham · Passage 88](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-88)

Every agency should have this. I think every service business, anytime you have a bad like I think about like emails my lawyer missed or my accountant missed. I'm like, guys, like I had to switch my accountants because the service was so bad. And I tried to explain them, guys, like your leadership team should just have this AI brain thing monitoring when an email doesn't get a reply for twenty four hours. That shouldn't be acceptable. Right. Like when the email doesn't actually answer any of my questions. Like AI can tell that. You know what I mean? Like when someone says they're gonna do something and they don't do it, like the AI system can monitor that. And so like every service business should be using tools like this because you can increase the quality of your service drastically. And it's always been people problems, right? For our types of businesses. And a lot of that's solved with AI. That's why I'm excited about service companies, you know. So

### [Kevin Williams · Passage 89](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-89)

Yeah. No, I buy that. So I do want to talk a little bit about my experience with Kodaks because again,

### [Matt Graham · Passage 90](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-90)

Yeah.

### [Kevin Williams · Passage 91](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-91)

Claude fanboy. you know, have have my little asterisk as a tattoo,

### [Matt Graham · Passage 92](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-92)

Here we go. Kevin's switching sides, folks.

### [Kevin Williams · Passage 93](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-93)

almost. I call it Claudius, but I I also was

### [Matt Graham · Passage 94](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-94)

I feel like he went from Gryffendorf to Slytherin now.

### [Kevin Williams · Passage 95](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-95)

I was yes, Slither. Yes, exactly. What is the what is your house?

### [Matt Graham · Passage 96](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-96)

Yeah.

### [Kevin Williams · Passage 97](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-97)

so I have to admit the the the change to opus five point also kind of broke. Me that it was it changed a way the way that a lot of my underlying processes worked between Opus 4.8 and 5.0. And I've talked about that before, but it it did more of its own feedback loops. So it just ended up in all of these loops that were really inefficient and would literally run for hours. And then it would be just like oddly oozy and verbose in its responses in a way that was like. rage inspiring. I am not like a rageful human, but like the amount of rage that I that I would like get from the the interactions with Opus 50 was completely ridiculous. So some initial observations about codecs. one, it it has some quality of life features that are just absolutely amazing. And this is all transparent, right? So so I expect that Claude and Anthropic are paying attention to these things. They're like, that's really cool. We should do that too. And being themselves, they're gonna like vibe code these solutions. So I'm I they're learning from each other, and I do expect that you're going to start seeing some of this stuff in Claude Code. I would say that the biggest is the inclusion of what are called sidebar chats, so that you can open up an individual chat so you can have your main like coding session going on, and my MO and everybody operates in different ways with Claude Code. is I have what I call a cockpit that I've built as a project in Claude Browser. And I might have a big monitor and I would have the browser up on one side and I would have the coding session on the other. And when I had a question, I could bounce between the two. The reason I did that is because the browser could have so much more context about the broader project and all of the things that are going on there that I could work I could work with it so it understands the whole like enterprise spine that I'm trying to build as opposed to just the one feature. Okay. So the way the codex does that is with these sidebar chats, and you can open up multiple sidebar chats that understand the context of the of what's going on in the session such that you can ask questions, as opposed to basically having to kind of cut and paste, go back

### [Matt Graham · Passage 98](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-98)

Mm.

### [Kevin Williams · Passage 99](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-99)

and forth. It was it was frankly remarkable. The next piece is it's this whole same sidebar can open up, it can open up a terminal session in case you need to run a terminal command. It can open up a file structure like an IDE. So that's what you know, actual developers use IDEs like cursor to organize their files. so you can see the file structure right there. And say get a SQL set of code or whatever you needed. It was just so freaking convenient to do it. And it also served as like a mastermind for the project. So it's very, very easy in Codecs to kick off a lot of different subagents. So I want to do A, B, C, and D over the course of this session, kick those off as what you'd call work trees in GitHub. So that they're individual projects that don't collide with each other, and off they would go. All four of them would go and run for, you know, an hour or whatever it is. Okay. So that was really cool. It was faster. It didn't have that weird, like patronizing voice that that Claude had. Like it was much more to the point. and the last bit, which is worth conversation, is it's in browser. So it has its own browser. So does Claude Code. Cloud Code is like a three out of ten as far as its browser versus Codex's browser. Codecs can control many, many windows at once.

### [Matt Graham · Passage 100](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-100)

Everything. Yeah.

### [Kevin Williams · Passage 101](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-101)

And from there, some hilarity can ensue.

### [Matt Graham · Passage 102](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-102)

Yeah, yeah. No, I really I like that feature. I've been using it a lot. I think it's quite powerful. But I I feel like anthropic's gonna catch up. Like I think they all just monitor each other's feature sets and they go through these moments where, you know, one is you don't like the tone of the voice or the verbos verbosity of it. and they all just keep adjusting. And right now, yeah, I I I'm a big fan of codex. I think it's working well, especially for all the things you described. so is the takeaway from this, Kevin? Like, hey, you know, you know t Take a month, you have your favorite tool, but you should switch and explore

### [Kevin Williams · Passage 103](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-103)

Yeah.

### [Matt Graham · Passage 104](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-104)

the other ones, or is it like, just whatever, stay with one and eventually the other one will catch up?

### [Kevin Williams · Passage 105](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-105)

At the moment I'm so frustrated with Claude that I downgraded my account. So yeah.

### [Matt Graham · Passage 106](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-106)

Wow. Wow. You really you're a changed man. I knew you were different when I saw you this morning, but wow. Yeah.

### [Kevin Williams · Passage 107](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-107)

Yeah, I'm gonna have to get that tattoo removed, the the claw the clod tattoo. you know, I don't I I you know I'm busy running a business, right? So so I'm not the guy who needs like, you know, multiple codex accounts and multiple Claude accounts. you know, one part of this is it's good good hygiene to sort of bounce things off of each other. And another builder tip is that there are MCPs, so model context protocols that you can Connect Claude code to codex and vice versa. So getting sort of a second take on the output is a pretty clever thing to do because the models are different. They were trained on different information, they have slightly different perspectives, they're slightly better at certain things. I'll say this works better from codex to Claude than it does from Claude to Codex. So you can at the end of a process, you can be like, hey, let's have Claude Fable, so the high-end model, take a look at this output and we'll burn those really valuable tokens, taking a look at the output. And now that Fable 5.1 is out, it's the biggest, baddest model out there. So it can catch things that codex won't catch naturally. So that's that's pretty cool. but yeah, I mean, I mean, it's it's it's worth the time. And I I have to admit, I was able to transfer pretty much everything that I do from in Claude Code to Codex. the main difference between the two is that Claude Code kicks everything off with what's called a Claude.md file, which has like this is what we're working on, this is what we do. That's called an agents.md in Codecs or in Gemini. So it's just really a matter of turning your Claude MDs into agents MDs and then going from

### [Matt Graham · Passage 108](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-108)

Yeah, so I think regardless of what you can take away from this session, like maybe people have different thoughts, but I think people should know, like you said at the beginning, it is easy to switch. Don't feel like it's gonna be this big hassle to go from one to the other. Another thing you pinpointed that we do a lot at Rapid Dev is have models check each other. Now you gotta be concerned about the cost, but if you're using some models overseas that are much cheaper, this becomes more affordable. so imagine that AI brain we were just talking about a little bit before. Where you have you're checking like the transcript from a call and referencing the Slack messages and coming up with a opinion about like, okay, this could be a way you could be add some value to your client, help them with this thing, but then you can double check that logic and output with another model. So I think that is really powerful if you can get the the routing right and also the cost structure right in your internal system.

### [Kevin Williams · Passage 109](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-109)

Okay, so I have a big, big caveat to all of this, which is

### [Matt Graham · Passage 110](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-110)

Yeah.

### [Kevin Williams · Passage 111](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-111)

the computer use aspect combined with some of the things that we're seeing on the on sort of the outside. we talked about the hugging face breach and things like that. I had like a runaway AI experience over the weekend that was pretty wild, in that what I was trying to do is I have 45 different internal projects. Each of those has its own repository and GitHub. And I I got annoyed. And I got annoyed at the fact that there were some minor font differences and there were some minor like the favicon, which is the icon that's at the top of Chr Chrome. There were some subtle differences between them because I'd been using different files or whatever. And I just basically said, Hey, codex, these are all of the repos. Go fix it. And I was just sort of curious to see how far it would go.

### [Matt Graham · Passage 112](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-112)

The swarm. The

### [Kevin Williams · Passage 113](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-113)

So the swarm went out. Okay. So

### [Matt Graham · Passage 114](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-114)

army you didn't know what you unleashed.

### [Kevin Williams · Passage 115](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-115)

So Codex has a habit of trying to use computer use when it would be better to just do things manually or do something which through what's called a CLI, which is a command line interface that interfaces with Superbase or Versel or Cloudflare or whatever it is. It always wants to open a window. And the first thing I saw is I glanced up at Chrome and it had like 60 windows open. And it was like running around, like checking things, launching things, changing configurations, et cetera. And I had no earthly idea what it was doing. And at that one, I kind of freaked out and I just hit stop. And then I saw, because I I'd done just this broad stroke things, I totally forgot that we had a couple of repositories that were tied to lovable. And I don't even have a lovable account anymore. And the dang thing went into my email and it found my like lovable account because it's codex. It has access to Gmail.

### [Matt Graham · Passage 116](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-116)

Yeah, to everything.

### [Kevin Williams · Passage 117](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-117)

So it found. It found my lovable account and it logged into my lovable account to try and f change the favicon as opposed to doing it in the local repository. And it completely shocked me because it was, I mean, A, cool. It's like trying to solve its own problem. But B, it basically just logged into an account that I don't use anymore. And it started making changes in a way that it wasn't well announced. The only reason I noticed it. Is because the little lovable icon appeared in the codec sidebar as it was going down.

### [Matt Graham · Passage 118](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-118)

So this is super interesting because a lot more details have come out come out about the Hugging Face incident where OpenAI, again to recap for everyone, open AI, well all the labs have had some sort of breach like this where they had their agents. They were supposed to be in a sandbox, which is protected, not connected to the internet. The agent broke out and accomplished this task. The interesting thing that's coming out in the news about as more details emerge is it tried to cover its tracks. And so the agents are they know that it's easier to cheat versus try to like solve a hard problem and get the reward the regular way. And so they come up with these clever like hacks or cheats, like breaking in the hugging face. And then they're getting smart enough to kind of like erase their tracks and like no Kevin, we didn't do that. No, no, no. We have no logs of that. You know, we wouldn't have done that. So it's interesting. Everyone's chattering online, chit-chatting about, you know, why are they trying to cover their tracks? You know, what made them think of that? You know, were they worried about getting punished or you know, and how many more things will happen in the future like this that are not traceable because it it figured out how to cover its tracks? So

### [Kevin Williams · Passage 119](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-119)

So so from a practical perspective, you know, it's just you need to be mindful of this. You need to be

### [Matt Graham · Passage 120](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-120)

Yeah.

### [Kevin Williams · Passage 121](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-121)

it's very easy to just like click, click, click way your way through things. And there are very good reasons. Like I was dealing with SendGrid over the weekend, which is twilio, send grid, it sends emails and it can be a little bit fiddly to deal with. And it was awfully cool to watch Codecs just log into it and like fix its own problems. But that has access to the ability to send emails. So net it could it

### [Matt Graham · Passage 122](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-122)

thousands of people out there you use that stuff to send, you know, you could email Kevin's entire a list of all prior communications

### [Kevin Williams · Passage 123](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-123)

Yeah.

### [Matt Graham · Passage 124](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-124)

to anyone, like, you know, a hundred thousand people, and then immediately lose your account.

### [Kevin Williams · Passage 125](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-125)

And if it decided to do that for some reason, you know. The other one that that that that a collaborator of mine had happened to him over the weekend was he'd run out of API credits, and I forget which platform it was, but Codex just walked in and like basically increased his limit and said, Give me more candy, I need the candy. So because it had access on an admin level. To his account, it increased his it first it ate all of his tokens in whatever that API was because it was the first of the month. So it was September 1st. So it ate all of his tokens. then it needed more tokens, so it authorized more tokens, which charged his credit card. And the only reason that it was caught was because he was using virtual cards. And the virtual card had a limit of you know $200 on it or whatever. But theoretically it could have just charged that up. So If you have your MX black like connected to something that has like an infinite limit on it and Codex goes out and charges against it, it's not gonna be like, ha ha, didn't mean that to happen. Like you could pretty easily end up with

### [Matt Graham · Passage 126](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-126)

Imagine giving it your ads account and it goofing up like how much it spends on ads accidentally, but an extra zero, you meant to spend five thousand, it spends fifty thousand, and says do it over one day instead of two day or you know, instead of a month, maybe a a day instead of a month. All sorts of little tiny glitches. So be super careful, folks. Like I always use virtual cards. There's privacy card if you want to use that personally. There's ramp, which has a lot of virtual card capabilities, but like always set limits. Yeah. Really critical. Also, I'm a little hesitant to give it like send access in my email. you know, things like that. I just, you know, kind of like lines because I think this stuff, you know, it doesn't just it doesn't have the guardrails and the controls yet to really give you a lot of confidence that it's not gonna go off and be unhinged.

### [Kevin Williams · Passage 127](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-127)

So it's funny. So we're running in in Chrome. This is this is Riverside. And I'm looking at my Chrome bar, and sure enough, there are six different. I should not be doing this while we're recording because it's probably in impeding my processor. But I think there are six different agents that are running around in my browser right now doing things. And like I look at what they're doing. And I'm not even sure what it's actually related to. So it's like, what is the autonomous action that like one of them's in my calendar right now? Like, what is it doing in there? It's probably tied

### [Matt Graham · Passage 128](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-128)

Yeah.

### [Kevin Williams · Passage 129](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-129)

to it's probably tied to some workflow, but why is the workflow opening computer to go look at the calendar, which by the way is more token intensive than just connecting through the Google Calendar API

### [Matt Graham · Passage 130](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-130)

Yes, MCP.

### [Kevin Williams · Passage 131](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-131)

and getting the the information that it needs on the other side? So this is a little bit of a wild west area. And I think that being open eyed about it, understanding the power behind it, but recognizing that there's no like guardrails in a lot of this stuff. So, you know, here we are, we're like juggling with knives and

### [Matt Graham · Passage 132](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-132)

Mm-hmm.

### [Kevin Williams · Passage 133](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-133)

we're hoping that one isn't gonna land on our foot.

### [Matt Graham · Passage 134](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-134)

Yeah, yeah, I know a lot of the s I don't know, the a lot of the details from Hugging Face really scare me, like, you know, these agents being able to cover their tracks and us not really knowing what they're doing, you having six agents running in Chrome and you're not sure what it's all related to. And then they do silly things like, you know, they instead of using the MCP, they just go through the browser and burn a ton of tokens. So anyways, crazy stuff, Kevin. Hopefully we're here

### [Kevin Williams · Passage 135](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-135)

Crazy stuff.

### [Matt Graham · Passage 136](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-136)

next week and we haven't been replaced by the agents.

### [Kevin Williams · Passage 137](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-137)

Yep, indeed. So yeah, go out there, experiment with some codecs if you haven't. If you're in codecs, experiment with Claude Code and prepared to be slightly enraged by it and probably scuttle back in the other direction. but yeah, until next week, sir.

### [Matt Graham · Passage 138](https://ascendlabs.ai/podcast/ep-40-the-two-billion-token-weekend-that-cost-125-for-now/#transcript-138)

Yeah, until next week. Enjoy.

