What's up, party people? My name is Sunny Ray. This is the Sunny Ray Show. We've got now one more guest for the day. Actually, maybe two more. We're going to be speaking with Ali. we're going to be talking about controlling AI models from within. A conversation on how to guide, govern, and shape AI behavior from inside the systems themselves, rather than relying on external controls. So, Ali, how are you?
Hi, Sunny. I'm good. How are you?
I'm good, man. This sounds deep. It's like I feel like we're going to be talking about like spiritual machines or something.
Like controlling from within. Okay. But, what what's going on? What are you working on? Tell us a little bit about the company you're building and what's kind of the this problem you guys are solving at Ranks.
Yeah. thank you for taking the time to chat with me, Sunny. So, today, the problem that we're solving at Ranks is to be able to peek inside a model's brain, right? Like, most AI model you see happens by controlling and manipulating what goes into the model and what comes out of it. but, we have no idea today of what's going on inside of the model. So, Ranks is a company that is targeting this specific problem, where we try to understand what's going on inside of a model. so, just to give you an example, right? Like today when you ask the model to do something bad, you look at the data going in, you look at the data that it has generated, and then you make a determination on whether it's done something bad or not, right? This is the same as let's say if you were protecting a very, large building, you check IDs at the gate of people coming in, you check IDs of people coming out. You have no idea what they've done inside and what damage they've caused. so, this is kind of where we are with AI, and Rings is a company that is focused on solving that problem, where we sort of build techniques, build science to sort of analyze what's going on inside of the model.
Yeah, that's such a massive, idea. I mean, and it's so perplexing to think that these powerful systems that we're giving birth to, when most of the smartest scientists talk about it, and they they're asked, "Oh, how do they actually work?" And they're just like, "Well, we have no idea." Like, what?
How is that even possible? Like, isn't it just like electrons and like math and some formulas, I guess? Like, they're like, "Yeah, no, we have no clue." I mean, I always found that a bit confusing. So, anyways, so you built safety systems at Meta and Roblox. Like, what does AI safety actually look like at that scale, Ali?
Yeah. So, I mean, at that scale, because of they have such those products that you just named have or those companies that you named have such a massive user base, you also see abuse at scale. It also becomes very lucrative to sort of bad actors at scale, right? and the challenges of fighting abuse in those environments are very sensitive to price, believe it or not, because unit economics becomes very important. It's an extension of the idea that, you know, you're not going to pay someone a thousand dollars to guard a hundred dollar bill. In very simple terms. Right?
Yeah, I like that. Yeah, Ali, I'm going to flag man three times already. Your internet kind of conked out on me already. so I don't know, we'll keep chugging along, but you know, it worst case we could maybe bring it to an end and redo it or something, but let's just keep moving for now. So walking away from Meta and Roblox to build this is a real bet. Like who backed you early and like what do they need to see first?
Yeah, so for me, I mean in the last few years of my career, I was building AI systems to protect against fraud, right? And during this time I realized that the models that I'm using to fight fraud themselves are susceptible to fraud. So you got that realization sort of to fraud and abuse, right? That realization sort of made me want to venture out into building out this company that I'm working on right now. and in terms of backing, I've been sort of supporting myself as we go. I have done a few pilots with a few large enterprises. we're focusing on getting government contracts right now. and yeah, it's been self-funded so far and I've had I've been very privileged to be helped with like to be to have worked with a lot of people in my network. So I published research papers with them. We like worked on pilots together. It's been an incredibly rewarding journey. both in terms of learning and I'm just grateful for the support of the people around me.
Yeah, dude, your internet man, I don't know if there's anything you can do to like change network or maybe turn off, but yeah, it keeps happening. Oh yeah, I'm going to live stream CD. I'll grab that later. I guess what's the through line, you know, from content safety at scale to reading kind of like model internals.
Yeah, so I mean they are different versions of the same problem at its core, right? Like there's bad actors and bad actors will find different abuse verticals. So like on a platform like meta, which is pre-gen AI, which was built in the pre-gen AI era, people would upload bad content. Now you don't necessarily need to upload static bad content. You can just generate it dynamically. So it is the same class of people or that same class of bad actors that you're fighting. It's just that their tricks have evolved and so does AI need to evolve, right? Like so does so do the defenses. So the defenses need to evolve as well.
Yeah. that makes sense. Hey, you watched teams bold guardrails on the black boxes. Like why is that treating the symptom?
Yeah, because guardrails by design operate after the model has already generated the data. So at this by the time a guardrail intervenes, your model you've already spent a few thousand dollars generating or like at scale, right? At scale. You've already spent a few thousand if not a few million dollars generating that content that you're going to block anyways. so there's cost then there's cost that's not only in terms of money but also in terms of speed. So with the guardrail model, you have to wait for the entire primary model to finish generating before a guardrail can take a look at it. And because of that, guardrails don't work with heavy payloads like video or audio or images in the same way because those are very expensive to classify. So your end-to-end user latency becomes very like prohibitively high. So as a result, people are deploying these video models and audio models without any practical guardrails today.
Crazy. That's wild. what does it mean to catch intent before a single token is produced?
Yeah. So, I mean, models like foundation models today, the fundamental unit of output is a token, right? being able to catch intent sooner is because you you're your model's output is sort of determined by the input that was given to it, right? So, by the time the model does its processing internally, we have an idea of what this is of the nature of the content this is going to generate. So, if you can intercept that intent of the nature of content early, you don't have to do the successive 50,000 100,000 tokens of generation, right? So, you save that cost. That is why it is and in doing so, you can do it a lot more efficiently. So, to quote one of the research papers that I wrote, we sort of build these tiny devices called probes, which attach onto the brain or to the internal substrate of the primary model while it's doing generating. So, we were able to catch bad content with a performance comparable to a model like guardrail model like Llama Guard and Granite Guardian and all of those using less than less than 0.01% of the parameters. That means you're using that much less GPU. Right? And you're we're doing it a lot faster. So, with Granite Guardian or with Llama Guard, you would end up spending maybe at least like 100 to 150 milliseconds in a super optimized deploy, we can do it in under 1 millisecond for text, and under 5 ms for video.
Interesting. And what is Rink's shipping today? And like who needs it first?
Yeah, so Rink's today is shipping a couple things. So, number one, we're shipping probes, ready-made probes for open weight models. We're also shipping a platform that enterprises can use to sort of.
Build their own probes on their own custom models that they might have in the house.
Fascinating. Okay, yeah.
Yeah, and I mean a third like a related thing to the second part is also a platform to extract the internal states of a model.
So, you want a first customer, I guess, inside of 90 days, right? Like what does that deployment look like? And like what happens if it slips?
Yeah, so I mean, we we've had customers before, but to continue with the spirit of the question, I think the way we do this is because what we're developing is so novel and not a part of the standard AI build stack yet, what we want to do is we want to build this out and release it open source in the open source realm for people to sort of work with it. Some of my favorite code has already been released, but we want to release a lot more of the platforms of the weight so that people can go out there, experiment with it, and sort of see for themselves the difference that our protect can make to production AI deployments. Like because of the way we're building this, we can work in a defense setting in an air-gapped environment where you have no access to the internet. We can work with on the in on the edge setting where you can just deploy the one model that you need to, and not have to deploy a second guard model at all. So, we can like work on people's computers, we can work on like edge devices like your toasters. If you We gen AI-enabled toaster, we could hypothetically run on that on an edge device like that. for a serious note, like in a defense environment, you have a lot of edge compute, we could protect all of that because of the novelty of our technique. You don't need to have a centralized server to do your safety checks.
Cool. How do you prove a probe is right when there's no ground truth for intent?
So, in a practical setting, when you are building these probes, these techniques, like there are techniques, like even if you zoom out from probes, right? this problem of not having enough labels is a separate problem in the detection realm that you sort of have to solve. Probes are just like a more efficient or a better detection methodology. So, the way you solve those problems remain the same where you sort of, you know, either produce some synthetic data or find a way to label data for real cheap or sort of, you know, do similarity checks and things like that using a very small amount of data. and then you can just use probes as an alternative detection methodology, instead of guardrail models or something like that.
I was going to say, reaching, an AI safety team is expensive, right? In both time and money. Like what has actually, like opened those doors for you so far, Ali?
I think, I think a lot of it has been my network. the second thing is just community outreach, so just putting whatever work we've been doing, just putting it out there cuz safety teams, having been on safety teams in multiple organizations at this point in my career, I know one thing that, you know, safety teams usually are in like firefighting mode a lot, find themselves in firefighting modes a lot, right? So, what we're just hoping for is when there will be fires, it's not a question of if, it's a In our minds, it's a question of when there is a fire, we want to be in the right places that safety teams look. And I think that's how my initial plan is to build discoverability essentially.
Yeah, that really resonates. So, cool, makes sense. I guess like just help me land the plane, you know, 5 years out, like our models still black boxes? Like what does legible AI actually look like?
Oh god, I would hope not. I think this is a frontier of science that need that it is actively being worked on. So, there's there's scientists at each one of these major AI labs and even outside of them in massively funded neo labs focusing on interpretability. So, I think we would make there's like I think we would make a lot more progre- progress in understanding how models will like how models work internally and hopefully develop better control mechanisms. I think in the future models will I see I envision a future where models will not only be controlled from the outside, but we would have not even the next not even in the next 5 years, but even in the next 6 months or in the next year, I anticipate having more controls that operate off of the model internals as opposed to like just feeding them as black boxes. I think we would we will start dealing with operating with models in a both black box as well as like we would have I see a stack of controls which operate not only at the black box level, but also at the white box level.
Interesting. I got a couple more for you, Aliyo. I think we have a few more minutes on the clock if you're okay. Like I guess what can you see inside a model that guardrails miss?
Yeah, so like I said, right? Like the analogy that I used earlier in our discussion was let's say you have a building, right? And you have to sort of let's say you have a building with like, I don't know, 1,000 condo units, right? And you have to run security for it. Just checking IDs at the gate is not going to tell you if a murder's happening in any one of the condos. Right? So, you need cameras in the doorways. You need you know, you need time safety doors. You need like you need doors with proper locks. You need different layers of defense. So, defense is typically built in any practical system, right? I you just give you an example of a real-world system, but in a digital system you can look at like a web application, a pure non-AI web application, right? Even that is built in defense there is built in layers. So, I think those layers of defense for an AI model are what need to be built out. Now, to your question about guardrails, right? Guardrails here only like have a fundamental signal gap. So, what they can do is they can only look at IDs at the gate. They have no idea of what's going on inside of the model because they just have no access to that data. So, it's like if you're a cop with the SEC and if you don't have access to the bank accounts of people, you're not going to be very effective at your job, right?
I know that makes a lot of sense, Ollie. I guess what why does catching intent early matter?
Yeah, I mean it matters because like you can do it cheaply. You can keep make it more safe. So, to this day, right? Prompt injection is an unsolved problem, and that is because people can build prompts. People can craft prompts that go inside a model and do weird things that we don't understand. But the guy who put in that prompt does understand what it's going to do, right? So, we have no idea. We have no visibility into where the bad actions are happening. so, that is why like we need probes. Like we need to build that mechanism. You cannot police or you cannot control what you don't see. Right? probes kind of help build that out.
And what makes a probe faster than a second safety model?
The richness of the signals it's operating on. So, because a second safety model operates at the text level. Let's say let's assume the example of text models, right? in a text deployment, let's say you're working with text, a primary model, like a safety model a secondary safety model, has to process that raw text. but what a guard what the way these AI models work is they take that text, human-readable text, convert it into a string of numbers, which is significantly more compressed and more rich in semantic information than a human-readable sentence. So, because these external models operate off of the off of these human-readable sentences, they have to perform that additional step of turning that into a sequence of machine-understandable numbers, which is redundant, right? So, we are essentially cutting out that redundancy, and we are actually observing what is going on inside the primary model. We are observing that sequence of dense numbers. So, that is what makes us a lot more, effective, a lot more cheaper, and a lot more faster.
Wow. And who feels this problem most urgently today?
I mean, a lot of these major AI labs have, since I started out on building this company about a year and a half ago, almost all the major AI labs, Anthropic, OpenAI, and even Google's Gemini team, have put out paper research papers, extensive research papers on probes, because they understand the value that this, brings to them. So, these organizations are already feeling like already feeling the pains of unsafe AI, right? There's well-staffed safety teams at each one of these like there's massive safety teams. I don't know about well-staffed. That's something that only the people individuals working in the company will be able to comment on. but the these teams each one of these labs have massive safety teams solving these problems and they are feeling the pain of it. Outside of these AI models, there's enterprises who are using AI security products. So, like you might have seen in 2025 Palo Alto Networks, Proofpoint, each one of these like Zscaler, each one of these companies secure cyber conventional cybersecurity companies bought an AI security player. So, it's not just if you just look at these data points, it's not just the labs that are the who are the providers of models that are feeling the pain, but the consumers of these models are also feeling the pain. And these data points that I just discussed with you show that you know, there's also a market that has been built around it.
And what would what would make AI truly legible in 5 years? And then maybe after that help me land the plane, Ali. Where do people learn more about you and the company?
Yeah, I think where we I will go from here I think the simple answer to that is it's going to be better. It's going to be more powerful and it's going to be more interpretable, hopefully. to learn more about what we're doing, visit our website www.wrynx.com, wrynx.com, or our LinkedIn page, which again is the company name, Wrynx,.
Good, man. Thank you so much, Ali. wow, dude, you're working on some really cutting-edge stuff, and I wish you the best of luck, and I see we have a bunch of people listening to us live, as well. So, thank you all, really appreciate it. And until next time, Ali, hang out for 1 second. I'm going to bring this stream to an.
