Dan Shipper on Codex as daily driver: every agent needs a human, and AI is actually increasing demand for experts

for an inventor. It's the It's the scammy grifter.

Yes.

Who's the inventor?

It truly truly is. Well, uh we have someone who ships. We have Dan Shipper in the waiting room.

The shipator coming back to the show. Let's bring him in to the TV van Ultradome. Catch you up with Dan Shipper. How you doing?

Good. How are you? It's great to be back.

You're a founder. You're a shipper. You're an author, a writer. Do you think of yourself?

I mean, I'm sure you just get this like way too much now, but the nominative determinism is like a little too on the nose with this one.

Yeah, I I just try to live up to my name.

Yeah,

you do. You're an elite ship athlete. Uh, it's great to see you. It's been too long.

Do you think of yourself as an inventor?

Do we need to bring inventor back? I feel like that term, it's like what I aspired to be when I was a kid. So, there's like a little 11-year-old in me who's like, "Yeah, absolutely. I am an inventor." But I would never go around calling myself that. Cuz I feel like you have to have

a garage workshop and be like making things that like have springs and make weird noises in order to be an inventor, you know?

Gadget related.

Yeah.

Yeah.

Even though

Yeah. I don't know. Inventions can be patented. people that invent software. You should give people credit for those things, but we've moved away from that term for better or worse. Anyway, uh it's been maybe has it been like four months, maybe more since we caught up. It's been a while. It feels like years in this era. In this era,

the world's changing.

Uh yeah, give us give us the update on on all things every and and Dan Shipper.

I mean, there's there's there is a lot to say. Uh the the big update is I just wrote a piece called after automation. Yeah. Um because there's so much uh there's so much progress has happened over the last four months. I think the biggest thing has been there's a hu there was this huge step change in models um starting in November with Opus 45 and then

uh GBD53 and that has just continued and I think it's starting to like delegating work is starting to cross the chasm from something that you do if you're a coder to all of knowledge work with uh first cloud cloud code then cloud co-work and now codeex which codeex the big shift the biggest thing that has happened is codeex is my daily driver now and it's fantastic. It is like totally changing how I work. Um, but I think as that starts to happen, you're starting to get a lot of people um who maybe have not been that clued into agentic AI use it for the first time and go like, "Holy [ __ ] [ __ ] like everything is going to change. Am I gonna have

I saw I saw an interesting benchmark here where someone in sort of outside of technology was was sort of uh lamenting AI progress because the the the tricky prompt they had was uh was uh come up with a Pokemon that ends in the two letters er did you see this? Uh and and and the and like the the base models, the cheapest things, the older models sort of get confused by it because of the tokenization. And it's sort of like a new version of how many Rs are in Strawberry. But if you hit a coding model, like you can just watch the traces of Cloud Code or Codeex and it's like go to Pokemon wiki decks, download them all, put them into a CSV, write the Python that checks the actual last two characters of the Pokemon name and it just nails it perfectly and it's traceable and so you get the perfect answer. And I think people haven't realized that that's what's possible and it doesn't all need to be memorization anymore. That's like the big shift.

I totally agree. And and you're seeing people you know I mean Dario's out there being like uh you know uh AI could warn could wipe out half of white collar entry-le white collar jobs. Yeah.

Um but even even people outside the industry like um Citadel Griffin Citadel like yeah he had this whole quote and I I have it in my piece. He said like these are not mid-tiered white collar jobs. These are like extraordinarily high high skil jobs being I'm going to pick a word automated by a gentic AI.

Yeah.

And so you can just feel you can just smell when someone's like use an agent for the first time and they're like what the [ __ ] Um and if you've been in AI for you know we've been covering it since 2022 since the GPT3 days you know that like that meme where it's like he's got the noose around his head and he's like first time you know like that's that's sort of how I feel. Um,

yeah. I mean, it's a big I mean, it's just it's a big it's a big shift because the fun Yeah. You're going from like for everyday knowledge work being like good at writing documents and good at collecting information which are valuable things but there you've never worked with someone in an organization that's like a superstar where all they did was where all they did was like gather information and like collect you know collect reports and things like that right it's like okay what can you do with that what can you what what kind of uh actions you can take Uh and yeah, it's it's it's been even even with how even with how kind of close we are to the action and following, you know, not not as long uh you know, we weren't doing the show in in the GPD3 days, but uh I still find myself like operating dayto-day thinking like, oh, I should reach out to that person to get their, you know, to see if they can help with this thing and then realizing like, oh, I can just ask codeex to do that or I can just I can get a model to do this. And it's a really it's a really it's like a fundamental shift and I'm still not routing enough kind of tasks through the models versus um or at least as much as I could. Um the other the other thing um I wanted to get your update on like how how you feel like memory is evolving. I've had some like pretty pleasant experiences recently

where I will I will do, you know, I'll I'll go to chatbt and and um ask for something and it's like actually able to pull in a bunch of really relevant kind of like context and memories in a way that is like helpful and and not um not uh not just kind of like distraction or annoying. and specifically like wanted to get your update because like let's call it 18 probably a year ago 18 months ago I think everyone was saying like memory memory is the moat like memory and then and then people just like kind of stopped talking about it for call it six or nine months or something like that but it feels like it might be having a moment too

so I I mean whether or not it's memory specifically I the the thing that has changed dramatically is that codecs and and also cloud code or co-work to some extent but codeex the reason it's so powerful for me is it has access to everything on my computer

um and so you know I when I published this article um after automation yesterday I was like who should I send this to and I was going to go through go through my texts and go through my emails and then you know whatever and I literally just said to Codex hey can you go figure out like a couple people I should I should send this to and it came up with like a bunch of investors a bunch of journalists a bunch of founders all people who I've talked to in the last like 3 months, but just going through my text messages, going through my emails, and it's just like a complete change in what is possible because it has access to all that stuff which includes memories.

Um, and and one of the things I'm trying to do with this piece is is to say as a company, we've automated everything that we can. Every single person has access to an agent. They have access to as many tokens as they possibly can use. And yet we've grown from four to almost 30 people since GBT3. So like what the [ __ ] is that about?

Yeah.

Um it it's very counter to the narrative I think especially like the popular narrative right now with like Meta is laying people off and Block is laying people off and the ClickUp guys like I'm going to fire everyone. It's only going to be 100x employees. So there's I think there's a really interesting question here about if you're actually on the frontier like we are. Um it seems like there's more human work to do than ever. Why is that? So I did a bunch of work

and the other thing is Ken Ken Griffin in the the quote you pulled earlier from in that talk he was saying uh my software engineers are more efficient than ever but there's no limit to how much software we need to build or that we want to build. And so he's not sitting there being like okay like uh a lot of this work is automated like we can we can become a lot more efficient here and reduce headcount. He's saying like great we can do a lot more as a team and I think that's like universally what companies that like have momentum are saying it's a companies that like don't have meaningful momentum or don't have clear narratives that are having to you know do these layoffs and and again there's still you know there's still a lot of the there's just still a lot of uh blaming you know sort of like crediting AI for layoffs that that would have needed to happen even if we weren't in this technology cycle.

Yeah, it's interesting because it's very popular now to say AI in the layoff announcement, but it's also very popular to say and our business is better than it's ever been.

Mhm.

Um what what do you guys think about that? I just feel like that's such a weird strategy or weird meme to be going around. Well, I think I think they're saying you can say our business is better than it's ever been, but that doesn't mean it's being valued better than it ever has, right? Like you can say like better than it's ever been. What does that mean? Revenue is up, right? Uh but but the example that I know you're thinking of, I can't imagine that that company is trading anywhere near it was trading uh you know, two a year ago, two years ago. And so when you have a massive correction uh in in the sort of valuation of your company, like you're gonna try to make moves to change that or free up resources to invest in new in new product areas and get that momentum again. So, I would say like when I see, oh, the bit the the business is bigger than it's ever been or better than it's ever been, um, I just don't necessarily I don't necessarily buy that because you have to understand like what what metric are they using to to to sort of like qualify that.

Yeah, I think that that's a reasonable frame. Um, and I wish that I wish that they would say that more because it gives it's just such a bad look for all all of technology for CEOs to be doing that.

Um,

did you see the standard chartered CEO got in hot water?

So, uh, CEO of standard charter uh came out and said like AI is working so much we're going to reduce our workforce and he used a really spicy phrase. He said like low value human work and we're going to replace it with capital that we're investing. It was a very quantitative analysis of something that's uh deeply human. Uh but I dug into his actual projection and he said 15% headcount reduction by 2030 which is like I would assume that attrition would get you there. Like is this like the most retentive company in history? Like I feel like 4% of the workforce probably turns over annually anyway. You can just like slow your hiring plan and get there if you want to shrink. And that's what a lot of companies have done in previous eras where there's been uh hiccups or disconnects between valuation. Like there were some there were some big private companies that uh during the COVID era or ZERP era, they just sort of froze hiring and grew into their valuations because the business model was sound. They didn't need to hire as much. They refocused, raised the hiring standard, allowed attrition to play its part, and they got to become more lean like more lean and mean organizations without like actually doing a layoff, which at that time would have been optically very bad because if you were doing a layoff during ZERP or COVID, it was like an indictment of you because it was like, oh, well, you're a victim of COVID or you're a victim of ZERP. Like you your business doesn't work in a high interest rate environment. your business doesn't work in a remote in in a more remote world. Whereas now the the uh the vibe is oh if you're doing layoff that's a sign that your business is is ready for AI, ready for automation, ready for for more uh for more efficiency. Uh and so I think people are sort of playing both sides. It's very very

how are you uh how are you uh thinking about you know this has been said you know talked about plenty at this point but I'm curious in the context of after automation this uh difference between like a job and like a task right so like AI right now very clearly can do a lot of tasks but it cannot do very many jobs or really

any jobs or like there's a there's a subsection jobs that AI can do well 90% of the time, but even that is not enough to replace the role or function entirely.

Yeah. And I I think this I think this all change changes and the and the lines between those terms are all going to change because they're all sort of relative to the capabilities of technology and and that's all changing too. But I can tell you that what we find internally at every and you know again we're one of those places if you at someone in Slack it's like toss up whether or not it's a bot you know um and uh everyone's using codeex and cloud code to doing do all their work every day in the engineering or like for writing for editing for design all that kind of stuff

and I can tell you that um every agent needs a human the further away an agent is from a human who's managing it the worse it does Uh and you see this too in like even in this even in the scaled AI companies like inside of OpenAI or inside of Anthropic uh they do they do have uh companywide bots that you can you can at but they're run by teams of people and I think that that's actually a really interesting and pretty stable phenomena based on how how these agents work. Um obviously they do like more and more complex work but what we see is that um one of the someone on our team Kieran Clson uh calls it the the the human sandwich is is AI collapses uh collapses tasks that used to take hours into like you know a few minutes. Uh but the human is still kind of like the sandwich on either end or the book end on either end who's like framing the task or um evaluating it when it's done. And what's really interesting is even though it can do expert even though AI can do expert human work my experience internally at every and I think you find this across the like AI industry is it actually uh even though AI can do expert human work it actually increases the demand for human experts because what happens is you can get expert human work out of an AI you can get like pretty good writing or pretty good images or pretty good code, but it's all based on yesterday's competence. It's all based on what is in the training data from yesterday. And what that does is it floods the market with PRs or images or writing that's like kind of good but not quite right for the situation. And what you need are human experts to come in and and and take take the the cheap competence that AI enables anyone to have and turn it into actual really really good valuable differentiated work because otherwise it's it's just the same thing as everybody else is doing and it's not valuable.

And um I think that's a I think that's a dynamic inside of how the models work and uh and how they how they function in the economy that is kind of lost when people are worried about oh yeah it can do all this great stuff. It can. And in order for that great stuff to be valuable, it needs to be done by a human.

Excellent.

What do you think about the the idea of like the YouTubification of software? Like Hollywood uh has been on a decline as creation and content creation has been commoditized and democratized. Uh but you've seen a lot like many more small creators pop up, many more lifestyle businesses, many more small businesses get off the ground. And uh it feels like that that is one possible direction that this goes like more niche software instead of needing you used to need to raise a whole series A from a venture capital firm to rack servers. Then with AWS, you needed to uh you know uh go through YC, be technical, hire a couple engineers. Now a few people can sort of get something off the ground. Are you actually seeing any movement towards that do you think?

I think that's definitely a thing.

Um and it's actually one of the one of the things that makes working at a place like every really appealing because if you're someone who wants to make make software, everybody else can make software now. So having having some sort of distribution and trust trusted brand makes a lot more sense than it used to.

Um I think that I I'm I'm very bullish on people like a person or a small group of people making software for a niche. I think that's totally happening. But I'm very I think the the the SAS apocalypse is like totally overblown. 99% of people are not going to be vibe coding their own apps. they might do it once but actually maintaining software is really really hard and it's a particular skill set that most people don't want to have.

Um and I would be buying SAS stocks because I I actually you know if you look at every we can vibe go whatever we want.

It is a marriage violation.

Yeah. The only thing I think the only thing I think about is um oftent times you're buying software for a certain group of people to sort of manage work and as certain workflows become agents then I I do I like like another way to put it is um a lot of people that are outside of tech that have been vibe coding are sharing apps apps with me that they've built that like the LMS can just do natively like pretty well already.

And so that's been that's that's the kind of the the potential bare case for this sort of like super super long tail of software is like as agents get get more uh competent and people learn how to uh they're sort of like unhobbled there's a lot of this longtail that can just be like a thread effectively.

Oh, you're saying you don't need an app, you just need to talk to people.

Yeah. Yeah, it's basically like you're talking to an agent. It's like, "Hey, I want to get my Let's use the most broy example pro possible. I want to get my bench press to uh two plates by uh eight weeks from now. I'm currently at, you know, a plate and give me a plan to get there." Right. So, instead of needing like a

you could vibe code bench press.

Yeah. You could do like the Yeah. at distribution or you could or the customer could just wind up going to any

and they're like cool and and they they generate you a workout you're saying great I did it I I was failing after three sets or whatever and then it learns that and it gives you a new workout and it and it uh just goes and goes and goes and you functionally get the what a vertical product could do. So I just think like there is an opportunity right now for this like long tale of apps but part of it is that people don't realize what the models themselves.

I have a few thoughts there. The first the first thought is I actually look at that as training a customer

and a customer that's going to really stick around with that is going to end up wanting things that just the bare thread in codeex is not going to do for them very well. actually chat is like not a very good um medium for for a lot of app interactions. So in the same way that Excel was you don't get the SAS boom without Excel uh Excel is like teaching people how to use computers in a way that then becomes enterprise SAS. I really think a lot of these um a lot of these uh codecs or chbt or cloud use cases are actually training potential customers who be who are power users to want to encounter problems that they want to buy software to fix. But I have a very specific prediction for what that's going to look like. Um, and I I I'm currently obsessed with what I'm calling codeex native apps.

And the the basic insight is for all of these tools. So like cloud code desktop and codeex when they're built primarily for developers for now. And when developers are working in them, if you're changing your app, it has an inapp browser that you can use to like, you know, the agent's in there with you. you can like look at your app in in local hosts and the agents in there and you're you're going back and forth and it's very good a very good collaboration environment for developer. I think that is it's incredible for any kind of knowledge work. I spend all day just in codecs and when I open up a thread, I just open up a a browser tab and I'm in my documents, I'm in my emails and it's me and Codeex going back and forth on a SAS app that's running inside of the browser of Codeex. And it is the most powerful thing I've ever used. And I I really think that is going to be a um a significant user experience type thing that we're going to see across all of the uh all all of these uh provi pro providers of agent orchestration platforms for knowledge work.

Love it.

Uh do you think your agents are planning a surprise birthday party for you?

Oh,

happy birthday. Do we hit the gong?

Yeah, we're hitting the gong. We're hitting the doing an early gong for your birthday. Wow.

A little a little birdie. Dave Chuck in the chat said that said it's tomorrow. Great great timing for the 3-day weekend. Great timing for you to just lock in at your computer and just grind all weekend.

Absolutely. Me and Codex are going to have the best birthday ever.

Um, awesome. What uh what can people expect uh from every over the next uh call it month because that's like a year in in uh AI years. We've got a lot of good stuff. There's there's some interesting uh stuff coming on the vibe check front. Every time a new model comes out, we do some we do some good vibe checks. So, there's going to be some really good stuff happening in the next couple weeks. Uh and then we've got we have our agent product plus one and that should be in beta probably by the end of June. And I think that's going to be really really cool for people.

Amazing.

Uh great to catch up. Uh have a wonderful weekend. Enjoy your birthday