The AI Paradox

LENNY: The last time you were on this podcast, you had this hot take that people were sleeping on Claude Code. You were so unbelievably right. The premise of this episode is we're going to go through what else you predict will happen.

DAN: The AI jobpocalypse is not really a thing. I am super, super bullish on PMs and full stack designers.

LENNY: You guys are hiring, doubled in people in the past year, which is not what people would've expected from a company that is so AI-forward.

DAN: I'm simultaneously extremely AI pilled and very bullish on humans. Automation is a lie. Every agent needs a human. We have so much automation, so much AI, and I also work way more.

LENNY: Creativity, it just feels like it's going to be more and more valuable to stand out from all the slop that people are shipping and launching constantly.

DAN: What models do in general is they make yesterday's human competence cheap. And so it becomes commoditized. It's not valuable anymore. What humans do is we go in there and we're like, 'Yeah, we have all this frozen human competence from yesterday. How do I use this to make something new and interesting?'

DAN: It's going to bifurcate in two main ways. One is everyone's going to have at least one agent that they talk to that they can offload work to. Second is that most of the work that you do is actually going to happen on your computer in an environment like Codex or Claude Cowork.

DAN: I think the SaaSpocalypse is dumb. I would buy SaaS stocks right now. What agents do is increase the number of users of SaaS, not get rid of it.

DAN: We speed ran the CLI era. It was nice while it lasted, but I think CLIs are over.

DAN: One of the things about predicting the future at Every is that what you don't want to do is prognosticate. What you want to do instead is just live in it together. So everybody at Every is an AI early adopter. We're almost 30 people now. When we did our interview, we were 15, so we've doubled in size in the last year.

We're all early adopters and we have engineers, designers, writers, editors, salespeople, customer service people — and everybody has a little bit of that, whatever that thing is where you're just like, 'Oh, I like to explore. I like to experiment. I'm very curious and I'm super all-in on AI.'

What that does is it creates this little pocket of the future where we're all living in it together and we get to be a little bit further ahead, because at any other company, there's a mix: there's early adopters, there's the middle of the pack people, and there's people who are very anti.

Another thing that happens, which is really cool, is because of our role reviewing models and being a little bit of a pacemaker in AI, we get access to stuff before it comes out. We get to beta test and alpha test and help steer the direction of where things are going a little bit.

So when I think about predicting the future, when you create an environment like that, it's actually just about noticing what's going on. A core part of it is writing about it. Articulating what you're noticing, articulating the future brings it about in a way that makes it real for you, your team, and anybody on the internet who's reading it.

One of the things we talk about internally is what I call 'the reach test' — when you wake up in the morning, do you reach for it organically?

LENNY: I love this combination of you are using the latest stuff, and you're good at being self-aware of here's what's weird and new and different and interesting. That's a really cool combination, partly because you have to write about it and you write about it. So that's the perfect recipe for someone having a sense of where things are going.

DAN: One of my favorite questions. If you look at the benchmarks, AI is going to just take all of our jobs. METR has this really cool benchmark — it measures how long the newest models can do tasks autonomously. Mythos preview, the big Anthropic model, can do tasks of 17 hours at 50% accuracy. Holy shit, that's crazy.

And I think it is real. Model progress is going up exponentially. But my feeling is that we will look back in a year and say humans actually have a lot more work to do, even as models get better. There's a really interesting paradox there.

My big prediction is it's going to bifurcate in two ways. One: you're going to have at least one agent in your company that you talk to and offload work to. Two: most of the work you do will happen on your computer in an environment like Codex or Claude Cowork.

When OpenClaw first came out, I was very convinced that everyone has their own agent — a parallel shadow org chart. Did you ever read The Golden Compass? It's like having a little daemon on your shoulder, a little part of your soul. I really thought that was happening.

But I have completely flipped. Now I really think the model is going to be a super-agent — one agent for the entire company. Shopify famously has one. Ramp has one now. There's all this hype with personal agents like OpenClaw, then everyone realizes it's way too much work. This thing breaks all the time. I have to fumble around, SSH into my server, blah blah blah. Most people just don't want to spend that time.

The fundamental, underlying thing is: in order for an AI agent to be useful right now, it really needs a human who cares about it. It needs a human personal connection — someone watching what it does, making sure it's doing the right thing. The minute you sever that connection, the agent is not really useful anymore.

That's why it has shifted to a one-agent-per-company model. You set up a forward deployed engineer who's responsible for making sure that agent works for the whole company. As models get better at being independent, that will trickle down — we'll have more personal agents because we won't have to fuck around with all the internals. The mechanism is: agents need people who care about them.

LENNY: That is so interesting — you need to garden your agent because there's context you have to keep adding to it.

DAN: I'm so excited about this one. Anthropic realized that with Claude Code, if you put an agent on your computer, it has access to everything you have access to. It uses the terminal, so it has superpowered access. These agents really understand how to use the terminal because there's so much content online about it — that created a super powerful coding paradigm.

OpenAI was, in my opinion, very behind on this, then has surpassed them recently. When people were still thinking about coding agents as pair programmers, Anthropic was among the first to say no, agents should run on your computer. There were people before, like Devin, who had the big cloud environment, but real adoption happened when you put it on your computer.

Once you have a coding agent on your computer that can build anything, it's actually really good for any kind of work. People started just hacking Claude Code to do all of their work. Anthropic then built Cowork — a nicer wrapping around Claude Code, fundamentally the same thing.

OpenAI's earlier Codex was very technical, super smart, but a little bit autistic. They didn't quite get what you meant — they got exactly what you said. Around the time they launched 5.3, they moved in this direction: 'Oh no, we get it. This model is fast, it's really good for general purpose knowledge work.' Then they launched the Codex desktop app.

Codex right now is my daily driver. I spend all my time in it. When I'm writing a document, Codex has an in-app browser. I open it, go to my doc (I usually do it in Proof, an online markdown editor I built), and Codex is running and watching me. I can see what Codex is doing. It's all in one place. I feel like I have this parallel work buddy that can respond, write, do research, use my computer.

I've been at inbox zero for 10 days straight, which if you know me is crazy. I literally have Codex gather all my emails with Cora, our email agent. It renders a little page, and I just monologue into it: 'Okay, go research this. Here's a question from our lawyers. Can you go collect all the documents from the last four years?' And it just does it. All the stuff I would procrastinate on, I don't really procrastinate on anymore.

For a long time, I thought the optimal experience of AI was to take AI and put it in a browser. The reverse is starting to happen and be really valuable: take the AI agent that you use all the time on your computer and put a browser in it so it can see everything you're doing. That's a magical combination.

LENNY: This is more profound than it may even sound. Instead of AI being baked into SaaS tools, you're predicting that SaaS tools will run within Codex or Claude Code?

DAN: That is one really important second-order effect. Yeah, I'm using Proof or PostHog or whatever inside of my agent, and the agent has access to the website. It has access to everything I have access to, and it has access to my whole computer. When I run the agent on that website, I'm using my tokens — I'm not using the vendor's tokens, I'm not using the app's tokens.

So it puts SaaS back in this place where, yeah, you want to make it friendly for an agent — everyone's got a CLI now, you want to make the HTML really usable, make sure anything that happens in the CLI shows up for the user immediately. There are a lot of issues to deal with. But once you do that, you actually don't really need to think about having an AI surface that's primarily going to be the thing that users use, in the sense that you don't need to build an agent necessarily into your product.

There's another really interesting bifurcation: having two agents is better than one. With Proof, anyone who uses it, I don't pay for tokens because they bring their AI to Proof. It changes what you build as a SaaS company — you build it now for both humans and agents to use at the same time, and it changes your margins back to: well, I don't really have to pay for tokens anymore because the user's going to bring AI.

LENNY: A lot of people are moving to CLI and trying to work from the terminal. Is part of this prediction that people shift away from that and back to actual UX with agents running alongside them?

DAN: Oh, yeah. CLIs are over. We speed-ran the CLI era. It was nice while it lasted, but it's pretty clear... Sorry. It's not that CLIs are going to completely go away — obviously they've been around for the last 30, 40, 50 years. They will continue to be around.

There was this moment when Claude Code was so popular, and people were like, 'The thing that's working is the fact that it's the CLI.' I don't think that's what it is. When you move into an actual UI for this, you start to realize we made GUIs for a reason. It's just nicer to be in a GUI, and you can get all the same benefits inside of a GUI, especially for non-programmer work.

The majority of the technical people inside Every are not using CLIs anymore as their main work surface. A lot of programmers are still flipping into it every once in a while, but more or less they're using Codex, Claude Code, Cursor — that kind of thing.

DAN: Automation is a lie, in the sense that every time you automate something, in order to make sure the automation is working well, you need a human on top of it making sure that it's working well.

I wrote this piece a couple years ago about the allocation economy — the idea that the way humans work with AI is going to be like being a manager. The thing you have to remember is managers actually spend a lot of time working. Most managers are not on the beach. They're checking in with their employees all the time, trying to figure out how to make the work better, how this person is doing.

There are differences between being a human manager and a model manager, but fundamentally it still requires a lot of time and attention. We miss that in the model discourse. One reason is benchmarks make it look like AI is more autonomous than it is.

I made my own benchmark — the senior engineer benchmark. I have this app, Proof. I vibe-coded it on the side while running the rest of Every. When we launched, it just kept going down. I had a lot of egg on my face. I'd say, 'Codex, fix it,' and Codex was like, 'I don't know what's going on,' or 'I do know — I fixed it.' And then it would cause four other errors, and I'd be going around in a circle. I wasn't sleeping. I vibe-coded so hard I got bursitis on my elbow. There's a life lesson in there.

LENNY: Vibe coder elbow.

DAN: I got two senior engineers to fix it independently — two different rewrites. Now when we get a new model, I just give it a prompt: 'This is vibe-coded slop. If you wanted to rewrite it from first principles, how would you?' All models until GPT-5.5 got 30 out of 100. A human senior engineer gets high 80s, low 90s. GPT-5.5 jumped to 62 — using an Opus 4.7 plan. Opus 4.7 plans are very good. GPT-5.5 is the only model with the sense of agency to rip out old code and rewrite from first principles. Other coding models paper over the edges — 'Oh, this is a big job, I'll just do a little patch.' And you're like, 'No, I specifically told you not to.'

It's clear that in a year or less it'll be senior engineer level. But here's the catch — I can change the benchmark to zero out the current model anytime. It took me a while to get to a prompt that didn't give away the answer but got the model to reveal what it's capable of. The original prompt I gave when production was going down was: 'We had four or five reported issues yesterday. Make a plan and resolve them all.' Every coding model on the market will take that seriously and try to fix the issues. What an actual senior engineer does is look at the code base and say, 'This is a piece of shit. We're going to have to actually rewrite a lot of this and it's going to be hard and risky.' Benchmarks rise on problems we've framed and can score, but a lot of human work can't be scored until you write it down — and the act of thinking to prompt it is something you can't measure. Even if benchmarks get saturated, it doesn't mean you replace all senior engineers.

DAN: There are definitely new job roles that are a thing. The Forward Deployed Engineer concept is for real, and it comes out of 'every agent needs a human.' Go to the big model companies — they have agents running internally, with teams of people running them. I don't think those teams are going away. Models will get more powerful, agents will get more powerful, the number of agents will grow — but people will still manage them.

That looks like a very specific kind of person. We have a couple of them internally, and they're the people in charge of making sure your agents are working and doing the right thing. We also do consulting and lend that out. It's another place where you go: hmm, automation was supposed to take away jobs, but it just created one — or many.

There's a specific type of engineer who really loves this. Nitesh, our AI engineer, fits this forward deployed category. He spends most of his time talking to one of our agents in Slack. We have an agent internally called Claudy that runs our whole consulting practice — he's in Slack with it constantly. There's code, and he uses Claude Code, but a lot of it is just talking: 'Why did you do this dumb thing? Let's fix that.'

Certain engineers love that — having their hands on the latest thing, and loving making this 'being' that's in a workspace. It looks a bit different than building more traditional software.

LENNY: Your sense is we're not near a place where these agents don't need a human?

DAN: Yes. I'm simultaneously extremely AI-pilled, extremely, and very bullish on humans and the role of humans in making sure AI is working well.

DAN: I am super, super bullish on PMs. My anecdotal case that has convinced me of this: we have a guy internally, Marcus, who runs Spiral, our writing app. Marcus is a PM by training. He previously ran Axios's writing product, was a PM, had a big team, got it to tens of millions in ARR. He took a year off, got super AI-pilled, and just learned how to use Cursor really well. Now he uses Claude Code.

I would call him lightly technical — knows what a database migration is. If he has to look at the code, I think he can understand it. We never could have hired him to do this job even a year ago, but coding models have gotten good enough that he can pair the technical knowledge he does have with his really spiky product sense — sense for writing, sense for users — and it's so dangerous. He ships faster than almost anyone on the team. He has an eye for every single user, every single conversation, what it means and how to collect it into a story about where to go next, what to fix.

He feels liberated because he doesn't have to organize a whole team to do that. He can just do it. It makes me very, very bullish on any PM who gets really AI-pilled.

LENNY: Music to my ears. The skills you need to build are the things — the building is done for you. What do you need to be good at? Figuring out what to build, figuring out if it's great, figuring out what problems to solve.

DAN: The other people who I think are going to be super power people are full-stack designers. If you're a designer using these tools all the time, you're so used to: I make this beautiful interaction, and the engineer just doesn't want to do it, or it doesn't happen the way I think it should. I see so many designers internally and externally now feel so empowered to build because they have all these ideas to make things look amazing — interesting interactions. That's exactly the thing that's hard to do with vibe coding because it all looks the same. They can make stuff that looks so different, and now they can actually build it. When we work with them internally, they're just making pull requests. They don't need to hand it off as much. The thing is built and that's it. There's a huge opportunity for those people to become entrepreneurs and start their own thing because they can make stuff now.

DAN: The only thing you need to do is ride the models. That means use them for whatever it is that you do. Codex and Cowork are becoming the standard operating system for work. When new models come out, try them and figure out: now there are new powers, how can I use them? Instead of being like, 'I'm going to try to ignore it because it makes me afraid' — which is honestly a rational, reasonable response. If you ride on top of them, they extend your powers in a way that doesn't leave you behind. You're part of the future and part of the way work happens.

LENNY: I like this term, 'ride the model.' What should someone working at, say, Salesforce do to ride the model?

DAN: A lot of companies handicap their employees from even doing this — at Salesforce you may not be able to use the latest models. You may have to do it in your off time. The thing I really like to do with new models is play. There are certain things where I know it can't quite do it yet, but when a new model comes out, I always turn the rock over again to see: 'Can I do it now?'

The way to ride the models is not one specific thing because they're always changing. It is to be curious and playful — to apply the new model to whatever you care about, whether that's your job or something outside, and to keep turning over rocks. It may not work now, but it may work eventually. The way you use it matters.

People think the edge of AI is in San Francisco. I actually don't think that's where it is. The edge of AI is wherever AI meets a real human doing something. The people in San Francisco are making it, but they don't actually know everything about how to use it. They need to see how other people use it. Whenever a new model comes out, you get to be one of the first people in the world to discover what it might be useful for. It's like a new discovery.

DAN (closing): Ride the models. Try all of your workflows in Codex or Cowork. If your company doesn't let you, do it on your own time. Try out agent products like OpenClaw, Hermes, or for less technical people, Viktor. Get comfortable with both ways of working — and try to have fun. Too many people are doing this out of FOMO, fear of losing their job. The best way to figure out useful things to do with AI is to do something enjoyable.