The Last Summer, Instinct > Instinct, Chad Super App, Jev, DevDay Is all MCP, Cloudflare Birthday Week

SPEAKER_1 (0:00)

Go, we're back.

Wilhelm Klopp (0:02)

This is the Bad Agent Podcast.

Matt Carey (0:05)

Bro, it's been ages.

SPEAKER_1 (0:11)

Welcome to the Bad Agent Podcast with Matt and Wil.

Wilhelm Klopp (0:14)

Remember Prompt Ops? This is intros from spring 2025

Matt Carey (0:19)

I was actually thinking that, I was like, how good does AI have to get before we actually have to update the intro?

Wilhelm Klopp (0:26)

Yeah, everything is definitely changing.

We're in the singularity. Are we in the singularity, Matt? Are we in the takeoff?

Matt Carey (0:36)

Fast takeoff. I don't know.

I don't know.

Wilhelm Klopp (0:42)

Are you okay?

Obviously, we haven't recorded in a while. It's been a summer. We've taken our summer break, which feels very appropriate for two Euro bros summer off or whatever. But I keep saying to everyone in my life, this is the last summer. From here on, it's just going to be crazy. It's going to get crazier every week, every month.

So this is the last summer or last weekend, we went to a festival in SF Europe. It's called Portola. It was really fun. I was like, this is the last big weekend of the last summer. So like, let's go, you know?

SPEAKER_1 (1:23)

What are you saying, dude?

Wilhelm Klopp (1:25)

All right, I'll give you one example.

There was a chart on Twitter yesterday. That was like, it used to be that there were three months between model releases from OpenAI and Anthropic. Now, it's 11 days. So every 11 days, there's a new model, right? You plot that chart out, it'll be every day soon. Then will be it no two requests will hit the same model.

Every request will hit a newer model.

SPEAKER_1 (1:55)

No, but they're going to... No, okay.

Matt Carey (1:57)

No, no, no, no, no, no, no. You theoretically could have two people updating a website every second, but you don't because it makes you feel seasick, because the website keeps changing.

SPEAKER_1 (2:09)

How no one's going to...

Matt Carey (2:11)

Okay, I can see this for chat, right? Them doing like continual upgrades, but for like API access, they can't do that, like they can't continually upgrade shit because then everyone's stuff will just continually break in differently wonderful ways.

Wilhelm Klopp (2:26)

Sounds like a model skill issue. Sounds like your model's not aligned enough, bro.

SPEAKER_4 (2:30)

Oh, Jesus.

Wilhelm Klopp (2:32)

Align your model.

Matt Carey (2:32)

I also think, okay, this is my other take that we're seeing at the moment, so I don't think it's necessarily that spicy, but the model labs are not releasing their best models to the public anymore. And this started with Mythos from Anthropic. But now we're seeing this with, is it 6.1 Astra is not getting released?

At least not anytime. It wasn't released with the rest of 6.1 Sol from OpenAI. And that was supposedly due to deception or whatever in the model. But, yeah.

Wilhelm Klopp (3:10)

Oh, right. I'm very confused about this, to be honest. I feel like they're, yeah, yeah, yeah. Sorry, continue.

Matt Carey (3:16)

Well, I feel like as the models get better, AI safety becomes more of an issue for the labs releasing something that is going to be deceitful, can have larger and larger consequences.

So then it basically costs more money to secure better models. And so the same way that if you make some software that you don't know is that, you don't know how good it is, you're not going to release it, but you know you're going to have infinite demand. For instance, Cloudflare releases a new product. You see a lot of the new products are released under private beta. They're released basically with customer consultation. Not everyone can use it because they want to see the weird and wonderful ways that it breaks at a scale that is very hard to replicate. So, I think that process is going to happen more and more and more.

And so there is a limit to how fast you can continually release models while you're doing that process. I reckon it's going to be more like an S-curve. We're in this takeoff because the model capabilities are not shatteringly good, but then at some point they're going to get so good, like the Mythos models, that they get way too scary, and then they take longer to work out the guardrails for each of them. Or they don't release them at all.

Wilhelm Klopp (4:39)

That's my theory. Interesting.

One of the lines I liked recently was from, I forget his name. It was, I think, on one of the Dwarkesh things where there were a bunch of different researchers talking about the challenges with recurrent self-improvement, and was it John Schulman? Recursive self-improvement. Yeah, John Schulman.