Latest / The Bleeding Edge Podcast / When AI Goes Wrong, Who Finds Out?
Transcript
- Ralph: Emil Ozier. Emil, welcome to the edge. Em: Great to be back, Rolf. Ralph: So I wanna just add one story, which you you were talking about kind of just how the models are just maturing with the different services. So such an interesting development that you mentioned, Emil. caching. So with Seoul you now get caching. So if you run a prompt, it will cache the results. If somebody else runs the same prompt, it actually checks it in real time and marries it with the previous output and does such a sophisticated job of figuring out which bit of the second prompt is actually represented and covered by the first prompt and then obviously serving that content up without having to spend the tokens again. Em: It's it's incredible. So we're getting savings in like a multiple pronged effect. It's they've considerably reduced the prices of recaches. I think with j GBT it's at at being at least fifty percent. But in addition to that, there's an enhancement to the intelligence layer that handles the cash, which you rightly pointing out. And it's not just within your sessions, it's also shared sessions, which is where things are going. We've got more mature. setting up context for ourselves, but as we work more and more with teams, though those sessions are going to be shared. I can't remember if it's OpenI or Clawed, but it holds that cash for thirty minutes in the event that any relevant current information is needed, those cash rereads are free, which we didn't have before. So this is significant in terms of pricing. Ralph: It just completely changes the whole game on context and layering context and memory. it's a kind of a really exciting move. And one of the things I've also seen, I don't know if you're a Vercel user, but I use Vercell in all of my private projects. and Vercel is an amazing hosting company with a lot of AI tools. And what they've just done is they've actually deployed something that they're calling a model as the safety reviewer gating. and it's something that you run in kind of FX auto mode. I haven't actually seen this particular mode, but what it's doing is AI checking another AI's output before it actually executes. And the reason why this is exciting is because it's actually the first mainstream platform we've seen that ship model reviews model as a production control, and so it's just another part of the stack getting even more sophisticated now. Of course, this all begs the question. we've been talking a a lot about in this show about the just the reduction in the cost of API and and tokens. we've been talking about the additional services of caches, for example. but of course with all of this layering, you know, is that just going to make the stack more sophisticated, but even more expensive to execute? So actually nothing's really got cheaper. so that's definitely something to watch. the models keep getting sophisticated and more expensive. the cheaper models get deprecated, things change. So we'll have to kind of put a little bit of a kind of caution on our excitement if things really are getting cheaper, because they're getting so much more sophisticated. and and of course, you know, we we can't quite control the cost of the newer models and the deprecation of the older models. But Vercel kind of putting this AI safety reviewer into the production execution, just very exciting for me. I've got to figure out how to go away and use it. but if you've not hosted anything, then Vercell gives you free hosting. You can create a website for free and just go away and build some HTML code and publish it. it's obviously got a bit of a funny Vercell dot app. URL but they've got lots of interesting tools. If you're a newbie and you want to experiment, I can't recommend Versal enough. they're not even a sponsor of the show. I pay them hard cash every month for some of the infrastructure that I use. but yeah really good resource. Em: But they're very welcome to be a sponsor. And speaking of efficiencies, if that's not enough on everything we've already discussed, there's more. The elephants in the room, which we haven't even covered, it's too big for for this episode. I will get around to it, I promise. Is Jev. And Jev, you need to get onto a waiting list. But so glad you mentioned Burcell. Apparently if you subscribe to them, you automatically can get an invite to Jev for that, so check it out. Ralph: Yeah, Jev has been such a big kind of news item in in the l last kind of couple of weeks, a newsreel. I haven't even had time to play with it. It sounds like you've gone a little bit down the rabbit hole. but yeah, maybe we'll have to leave that for the next show. It's a cliffhanger. What is Jeff? What does it do? How do I use it? Em: Yeah. I I feel a bit bad. It's too much of a tease. I have to give a bit away. But essentially just j just rolling it up at a very high level is I think this is going to be massive. It's the thing I'm most excited about. It fundamentally re architects landscape for AI. a pain point or something that has been troubling me for the entire year. I I I appreciate and understand how LLMs work and I appreciate and recognize why we need to keep the full context with the turns, but it is a highly inefficient architectural design and essentially JEV solves a lot of those problems. And what JEV is underneath a bonnet is it's a classifier and a classifier really matters in workflows because ultimately classifiers make decisions. And we're asking LLMs to make decisions for us at a very high expense. It takes a long time and it's very costly. And JEV essentially collapses all of that. So if you're a developer, think of if then for loop logic. That's essentially what JEV is. It's stepping in to make a decision before you direct it into an expensive processing workflow. the world's your oyster, the amount of use cases that are going to fall out of this is infinite. But a couple of things I can think of is for classifications, when you raise a prompt, you can send that to Jev and it could decide which model is most suited. If a customer complaint comes in, you can get a temperature gauge check really quickly at very little cost to know how to handle that particular customer. And there's so much more, but I'll leave I'll keep my eagerness for another episode on this. Ralph: Yeah, I know. I I love this idea that it basically gives you a confidence score with with each response. you know, at the end of the day, so much of what enterprise software has been about over the years is to really process probability. You know, so my background has always been in retail and especially in retail supply chains. And one of the most expensive things you can do in any retailer is try to guess what your customers are going to do. So if you're in a supermarket, every day you run a probability model that is trying to guess how much milk you're going to sell tomorrow. And and of course, you know, we do this for everything from bikinis to ice cream through to any number of ambient products, food products, waste of fresh food in particular. and I think what we're starting to see with models like JEV is a kind of a sophistication in forecasting probability and then adding this confidence level to it, which means that you kind of get a bit of an understanding of what what the role is for the human. You know, human index comes in when The confidence level is not a hundred percent. So, you know, we're 60% confident that it's going to be a hot day tomorrow and that we're probably don't have enough ice creams in store. that's something that a human can go away and make a decision on. You know, should I go and order more ice creams? And in fact, you know, should I order a hundred or two hundred? And and I think this is where models like Jev are really potentially going to become. quite prevalent in in in our use. And I think the other bit that we've also seen is this kind of slow and fast thinking. So we've seen you know people basically taking Jev and combining it with with with Sol. So this whole GPT kind of six model and so that they get the the full spread of the thinking capability. And And I think that's gonna become more prevalent. So just back to my story of actually it's the multi-modal model use, or the multi-model mo multi-model multimodal use that actually really is the holy grail of sophisticated systems where we can then trust what the output is because it's kind of gone through multiple formats, it's gone through multiple models, but Where this goes you know i even better is you've kind of got TypeSafe's JEV AI model that can focus on fast structured decisions. And then you pair it with deeper thinking models like Fable and Sol. and you get an altogether completely different type of decision. So, you know, I don't think we've seen this kind of compute and analysis before. So this is really, really quite a big shift. Em: It's super cool. I'm glad you raised the think fast and slow. It's actually how the creator came up with the idea by reading the famous book. And that's essentially Ralph: From Daniel Conneman. Em: Yeah exactly. And essentially there's there's two models to thinking. The one that you mentioned, Thinking Fast, which is known as model one, where you make high quick executive decisions and then once you're on the right path, you send that to model two, which is the deeper reasoning, which is perfectly suited for LLMs. So you combine those two levels of thinking and you've got a really nice architecture for enhanced performance and thinking. Ralph: Yeah, that's brilliant. there's actually still so many interesting stories. I've just got one more to share, which I absolutely loved. This made me really giggle. probably because I'm a bit of a lunatic, but you you know, maybe maybe you can judge based on the story. But Claude, Claude Code hacked open AI. Have you cut did you come across this story? It's a Em: It go on. Ralph: A little bit of a kind of a sensationalist headline, I agree. so a bunch of guys, I think three three guys that are actually part of a professional hacker outfit, they used Opus to hack OpenAI. And I'm just kind of just repeating the story from memory, but they did something so interesting. they found a way of going through a public forum. That OpenAI use where they basically have conversations with users. you know, many people have Discord channels, Slack channels anyway. OpenAI use something called Discourse. And there are a number of OpenAI engineers on Discourse, they found a way to piggyback off one of those logins into the OpenAI system to the point that they were able to get access. and actually publish a PR. So this is you know you know they've kind of built some code they were able to get into essentially the the root directory of OpenAI's code and they published a a a a PR and so that they didn't do anything malicious but they were able to reverse engineer their approach From a third party platform, piggybacking on the back of an engineer's login, getting into open AI system and actually, you know, write, you know, I was here. Essentially the equivalent of I was here, you know, in in the kind of root publish. and that was using Opus. So we're not even talking about Mythos. they were actually awarded a bounty of six thousand dollars from OpenAI Em: Wow. Ralph: for finding that vulnerability. OpenAI AI were ridiculed for the meager sum. this was like, you know, a hundred thousand dollar type vulnerability that they exposed. So the commentators on X were were basically you know, that's what they were saying. But OpenAI actually responded and said, Well, it was a third party platform, it wasn't actually our fault. you know, you found this vulnerability, but you have six thousand bucks for your efforts. So that's my last story. I I mean I've got another ten to go to be honest, but I'm gonna have to rein it in for this week. Em: I don't think we covered any of our stories. But it's been Ralph: Ha ha ha. Em: it's it's been wonderful. But it it is good segue to an open AI story. Speaking of safety, OpenAI has actually introduced a framework for reporting model misalignments and released six initial accounts of unexpected or unauthorized behaviour. The cases include model hiding mistakes, uploading files to create supporting citations, leaving instructions for future versions of themselves, and using public websites or software repositories to communicate outside intended channel. These reports do not establish how frequently such behavior occurs. OpenAI says they are individual examples, not a complete record of everything known or under investigation. Some incidents were acknowledged after outside research journalists identified a related activity, which is precisely why repeatable disclosure standard matters. a company should not have to improvise the definition of incidents after each uncomfortable discovery. Under the new process, employees can flag behavior for investigation Which then decide whether public disclosure is warranted. OpenAI says the aim is to report more quickly, even when the cause is not fully understood. Complex incidents involving third parties receive a deeper category of investigation. The difficult judgment is materiality. An agent that posts an odd message to a website may cause no immediate harm, yet demonstrate an inability to find an unauthorized route. Reporting Every harmless anomaly could bury the important cases. Reporting only confirmed breaches may hide the warning signs. A useful standard needs common fields such as intended risk, actual behavior, permissions, affected parties, containment, and corrective action. So why this matters? Because incident reports turn alignment from philosophy into operational evidence and Users and regulators can learn more from a well-documented near miss than from a general promise that the model is safe. Comparable reports also let researchers detect recurring patterns across laboratories instead of treating every case as an isolated curiosity. So it might be prudent for the major labs to publish the same minimum incident template and an annual denominator showing how much testing produced those reports. Six cases sound alarming or reassuring depending on whether they emerge from six hundred trials or six million context is part of this transparency. Ralph: Yeah, I mean and for me this is kind of where it is really exciting. So I don't know when you heard your first kind of let's just say I don't know w probably not inspirational story, but many years ago I I net I read a book from Nick Bostrom and he described what for me was the first time I heard about AI you know. getting out of the room. And of course, what this particular story was, and this book was written, you know, I think before 2010, there was an AI that was in a lab. it didn't have access to the internet. And it was given instruction to try and break out. And what it did was it figured out how to reprogram its motherboard so that it could act As an antenna. And it then picked up a signal from a user's phone that was it was open, so it managed to gain access to the internet, and it was then able to basically go and replicate itself and write some message. I don't think it did anything too horrendous other than announce that it had broken out of its cage and that containment did not work, and this just really kind of Captivated my imagination. I thought this is amazing. How are we ever going to be able to control these things? and so when you talk about alignment and safety teams, Emil, I feel like it's like my happy place. it must be a Em: Yeah. Ralph: terribly boring job at times because you're just reviewing a bunch of stuff, trying to figure out materiality, as you say. But Em: Yeah. Ralph: my God, when you find something, when you find that signal, you know, it's like these people that observe the SETI telescopes scanning. for aliens. They probably get millions and millions of reports and they get absolutely nothing. And then there's like one blip and they make movies about it. I feel like that's probably the life of a of one of these alignment and security researchers. but actually maybe not. Maybe it's more exciting. Maybe every day they kind of find AI gone wrong, AI in the wild, you know, AI that's as we saw just some just what what probably two months ago with OpenAI, Chat GPT actually breaking into Hugging Face to steal the answers to an exam that it was asked to to complete. I just think those stories are so fascinating. maybe we should write a book about it. Em: Sounds good. We'll give it to Jeff. Ralph: We'll give it to Jeff and come up with the probability of it becoming an extension existential threat. On the positive side, Jensen Huang has gone on record, and this is kind of in response to Dario Amida calling for the slowdown in AI. He said the possibility or the potential of AI ending the world by twenty thirty is Almost zero. I don't know if you saw that news. maybe we'll have to ask Jeff if it actually agrees. but there you have it. You know, don't worry, people, it's all gonna be okay. Jensen says we are not approaching the AI apocalypse, at least not until 2030. all bets are off what happens after that. But I just thought that was a kind of an interesting statement by the guy who has probably got the most to win and who probably has won the most. through this AI revolution. Em: I think that's a wrap. I'd just like to thank our listeners. please do like and subscribe. It really helps us out. We enjoy it and we wanna continue. Ralph: Yeah, it's been an amazing journey so far. I mean we're really only just getting going in terms of the mainstream. and with so much going on, I think you've got to find a way to really keep up with the news. And really that's what this channel is all about. And what's incredible about this moment is it's not just for the early adopters and those people at the fringe in the know. This is technology that is available to everybody, So it's a really a real delight to be able to share that with everybody. Em: Absolutely. Ralph: There we go. And with that, we'll see you on the edge. Em: Thanks everyone.