AI isn't for cutting costs, it's for multiplying impact | Super.com's Matt Culver

**Andrew Zigler** (0:06)
Okay, from the top. Welcome to Dev Interrupted. I'm your host, Andrew Zigler.

**Ben Lloyd Pearson** (0:11)
And I'm your host, Ben Lloyd Pearson. This week... Oh. Whoa. We're professional. Yeah. All right. This week, I'm sitting down with Super.com's Matt Culver, who joins the pod to discuss why the common view of AI is a tool to reduce head count, is a misguided accounting mindset that ultimately destroys trust. Instead, leaders should see AI as an opportunity to reinvest in their teams and improve the entire product development lifecycle to generate more value and empower developers. But first, we have some news. So what's up this week, Andrew?

**Andrew Zigler** (0:45)
Yes. In the news this week, we have an interesting roundup of both AI skepticism and excitement. So one of the things that came across our desk was Anthropix report about model poisoning also paired really well with Charlie Waugh's timely article on intelligence drift with these models and how folks are using them. Is that real? Is it fake? Is the phenomenon either of those? Doesn't matter. And speaking of real or fake, we're talking about AI usage and water consumption. We found a really interesting article that breaks down the context of this number. And in emerging trolling as a marketing playbook in Silicon Valley, we're going to look at a little closer at what this means. And for the web dev nerds out there and the lovers of things like semantic HTML, myself included, we're taking a deep dive on a fun little HTML tag that's maybe been left in the dust for a long time in the HTML spec, and finally, we'll have its moment in the spotlight. So a fun little round up this week. What do you want to dive into first, Ben?

**Ben Lloyd Pearson** (1:42)
Yeah, let's just start right at the top with Anthropic and how a small number of samples can poison LLMs of any size.

**Andrew Zigler** (1:50)
Yeah, so Anthropic is leading the charge on all things AI model research and security, and they release these really fantastic reports in combinations with AI research organizations. In this case, the UK AI Security Institute and the Allen Turing Institute, and they found in this joint study that in as few as 250 malicious documents can poison a large language model regardless of its size or its even training data volume. So what this means is that a huge model, like something with like a 600M model, can be backdoored with these articles with like only 250 of them. What this means is that there's only a small amount of documents in any model that an attacker has to control in order to poison or alter the model's behavior. So in their example, they used some really interesting approaches that generate some documents with gibberish and were able to poison all sorts of models. What did you think of this one, Ben?

**Ben Lloyd Pearson** (2:49)
Yeah, I think that example that they used of testing it where they convinced a model to output gibberish whenever it encountered a certain term was a really great illustration of how this works. And past research in this area has shown or in similar areas have shown us how you can do things like, for example, using a lower grade model to judge the output of a higher grade model to spot deception, for example.
So I'm almost wondering if there's some way that you could accomplish something similar here, like using a lower grade model to determine if the higher grade model might have been deceived by something that has been injected into it. But it is a little shocking just how easy it is to inject bad information into these models. And maybe this will, in one of our later stories, we'll maybe have some insights that tie to this. But I think really what we're seeing here is that just the fundamental nature of software security is changing. So now information itself is an attack vector. So if you're using models that ingest data from somewhere for their training, that is now an attack vector into your organization. And I think this is where stuff like AI observability is really going to become more and more important, just to make sure that models are successful at the things that you expect them to be successful. But I'm just gonna continue sitting in this chair here, preaching the good word of data provenance. You need good data input into your systems. You need to trust that data, and it needs to be enough to give adequate context to your models. And I honestly think that is probably the biggest challenge that practically anyone that has an AI initiative is facing right now, just getting good data into your models, whether it's your own data or it's a model that's been trained on data that you don't control.