Controlling AI Models From Within

What we talked about.

In this episode of the Sunny Ray Show, host Sunny Ray talks with Ali, founder of Rinkx, a company working to make AI models interpretable from the inside out. Ali explains that most AI safety today works by inspecting inputs and outputs while the model's internal reasoning stays a black box, comparable to guarding a building by checking IDs at the gate with no cameras inside. Drawing on experience building fraud and safety systems at Meta and Roblox, Ali describes how Rinkx builds lightweight probes that attach to a model's internal states to detect harmful intent before generation completes, using a fraction of the compute and latency of traditional guardrail models. The conversation covers why guardrails are reactive and expensive at scale, how probes can run on edge devices and in airgapped environments, and why major labs and cybersecurity companies are already investing in interpretability. Ali closes with a vision of AI systems controlled through both external guardrails and internal white box techniques within the next year, moving the industry toward more legible and trustworthy models.

Rinkx founder Ali explains how probing a model's internal states, not just its inputs and outputs, could make AI safer and faster to govern.

The questions, and the answers.

What are you working on at Rinkx and what problem are you solving?

At Rinkx we're trying to peek inside a model's brain. Today most AI safety works by controlling what goes into a model and what comes out, but nobody knows what happens inside. It's like guarding a huge building by checking IDs at the entrance and exit with no idea what happens inside. We build the techniques and science to actually analyze what's going on inside a model as it generates.

You built safety systems at Meta and Roblox. What does AI safety look like at that scale?

At that scale you have such a massive user base that abuse also happens at scale, and it becomes very lucrative for bad actors. Fighting abuse there is also extremely sensitive to price because unit economics matters a lot. It's like the idea that you're not going to pay someone a thousand dollars to guard a hundred dollar bill.

What's the through line from content safety at scale to reading model internals?

They're different versions of the same problem. Bad actors just find different abuse verticals. On a platform like pre-Gen AI Meta, people uploaded bad static content. Now you don't need to upload it, you can generate it dynamically. It's the same class of bad actors, their tricks have just evolved, so the defenses need to evolve too.

You've watched teams bolt guardrails onto black boxes. Why is that treating the symptom rather than the cause?

Guardrails only operate after the model has already generated the content, so you've already spent the compute, sometimes millions of dollars at scale, on content you're about to block anyway. You also have to wait for the full generation before a guardrail can check it, which is why heavy payloads like video and audio often ship without practical guardrails today.

What does it mean to catch intent before a single token is produced?

A model's output is shaped by its internal processing before any token appears, so if you can read that intent early you avoid generating thousands of unnecessary tokens. We built tiny probes that attach to a model's internal substrate during generation and match guardrail models like Llama Guard in accuracy while using under 0.01 percent of the parameters, running in under one millisecond for text and under five milliseconds for video.

How do you prove a probe is right when there's no ground truth for intent?

That's really a labeling problem that exists across detection methods generally, not just probes. You solve it the same way you would elsewhere, by producing synthetic data, finding cheap ways to label real data, or doing similarity checks with small amounts of data. Probes are just a more efficient detection methodology sitting on top of that same underlying challenge.

Five years out, are models still black boxes? What does legible AI actually look like?

I hope not. Interpretability is a real frontier being worked on at every major AI lab and well funded nonprofits outside them. I think within the next six months to a year, not even five years, we'll see more controls operating off model internals rather than treating them as pure black boxes, so we end up with a stack of both black box and white box defenses.