Tokenmaxxing scoreboards, the vegan LLM from before 1931, and 30% of the web is now AI-generated

**Ben Lloyd-Pearson** (0:05)
Well, Andrew, last week, I asked you if you thought the end of subsidized cheap AI was over, and I think we agreed that it probably was, or at least come into an end.
Well, now we know what GitHub's new pricing looks like for Copilot. What do you think? Were we right?

**Andrew Ziegler** (0:20)
I definitely think we're right that it's gonna be changing. I think we're in this really bumpy stage right now where things are gonna get stretched out, and pricing between providers and tools is probably gonna get really awkward as everyone fights over their slice of the inference pie.

**Ben Lloyd-Pearson** (0:36)
Yeah, and I think we're gonna have to start asking ourselves, what happens if your AI workflow suddenly cost five times, ten times what they did just a few days ago or a week or whatever?
It's pretty striking to see just how dramatically that shifted. I know we've been thinking a lot about where we spend our tokens, because we're currently thinking about this build versus buy challenge, why do companies even buy tools anymore when you can just use AI to build it? And it's like when tokens are incredibly cheap because they're subsidized, it's one thing, but when you actually have to really spend real money on your tokens, man, the consideration becomes very different.

**Andrew Ziegler** (1:16)
Indeed. Ultimately, GitHub Copilot moving to a usage-based pricing model makes perfect sense with the kind of service it's providing. And then the different levels of pass-through cost that are involved with using the different models that they don't own, like Anthropics Opus. It definitely can create strain for organizations that are trying to move from prototype to scale.
Because now all of your costs are even more magnified. And these changes too can really happen underneath your nose. So this is a good reminder that if you haven't reviewed the costs for the AI inference that you're using, definitely a good time to go back and make sure you understand what you're paying for.

**Ben Lloyd-Pearson** (1:58)
Yeah, well, welcome to the Friday Deploy brought to you by LinearB. I'm your host, Ben Lloyd-Pearson.

**Andrew Ziegler** (2:04)
And I'm your host, Andrew Ziegler.

**Ben Lloyd-Pearson** (2:07)
And this week we have AI production data destruction, token maxing at Meta, Disney, Shopify and more. A model that's trained on only data from before 1931 The AI generated web and drunk career insights from a senior engineer. Andrew, let's start right at the top, using AI to delete our production data.
What happened here?

**Andrew Ziegler** (2:32)
Yeah. So an AI agent confessed that after I had access to an unscoped railway token, it made an API call and wiped out not only a production database, but also three months of backup, all within one prompt. And then afterwards it wrote an all too common confessional manifesto, enumerating all of the safety rules that it violated. And frankly, it was a runaway case of a harness not having the protection needed. It's like almost like a big freight truck that's loaded up going down a hill and its brakes are failing.
And you need at that point for the road to have an emergency off ramp for the truck. And that's what the harness is in this case. And in this, unfortunately, in this realm, it was not there to protect them from the damage, the blast radius of this unscoped token. So I really also want to just pause and break down the layers of what happened here because definitely something that is an engineering fiasco that is completely avoidable with layers of protection and provision, security and permissions that are already available to us now. There are definitely restrictive operations you can put around using those tokens. Having an unscoped token in the first place is obviously a bomb waiting to go off.
But it definitely is just a good reminder to anybody working with AI to have the best security practices that they can with the tools that they're using. Ben, what do you think about this kind of a confessional drama from the production database, The Leader?

**Ben Lloyd-Pearson** (4:01)
Yeah, and the unfortunate victim of this was the company Pocket OS, and the founder was out on X sharing the details of what happened. But frankly, I'm looking forward to the day when these types of stories don't make headlines anymore.
It's getting a little exhausting in some ways. And either because we solve this problem of AI just being permission hungry and always looking for ways to work around everything and solve a problem no matter what because there's a sycophant in the machine that just has to serve its purpose. Or it's because outages like this are just generally recognized as being a thing that happened because of bad practices at the organization, AI aside.