MC

Mike Chambers

Things Mike Says on Podcasts

Where to Find Them

Mike Chambers has been a guest on The New Stack Podcast .

Recently: “Beyond LLMs: How Small Decision Models Make AI Agents Smarter and Faster” on The New Stack Podcast (November 2026).

What They Said

“You don't want the agent to be marking its own homework. You don't want the agent to be deciding, yeah, the thing I decided to do, I definitely think I should do.” — Mike Chambers, The New Stack Podcast

Asked what rules should apply when an agent can spend money, Chambers says the guardrails should sit outside the agent's own loop. He sees cheap decision models as a way to afford more of those checks.

The New Stack Podcast · 2026-11-04 Permalink → Listen →
The New Stack Podcast Around 22:45 into the episode
Mike Chambers

well, yeah, maybe either

Nick Lucchesi

or, I suppose. I

Mike Chambers

…an agent and persuade it to send me some money, that would be great. But yeah, I guess, yeah, there's a place for a decider model in something like that. So, one of the things that you want to do when you're doing and implementing these guardrails is you want to have a system which is able, or it's a good idea to have a system which is out of loop of the agent that you're currently running. So, you don't want the agent to be marking its own homework. You don't want the agent to be deciding, yeah, the thing I decided to do, I definitely think I should do. And so, having something which is running somewhere else. So, this is where we have like LLM as a judge and all that kind of stuff comes into play. Those kinds of terms are used for different things, but this could be used there too. And so, a decider model might mean that we can have some decisions made around those kinds of processes maybe more frequently than we could before, because it's in…

Nick Lucchesi

important. This is not on my list of questions, but I want to ask. Like, how far away are we knowing what you mean? You've been covering, you know, in AI since it was called machine learning, right? So, or popularly called that. How far away are we from agents regularly spending, let's say…

Mike Chambers

Look, I guess there are a couple of different ends of the spectrum as was sort of intimated here. So, like, I am, I'm personally like interested and have played around with X402 and those types of things, where I think it's a super fascinating protocol full stop, but also as a protocol to basically…

Nick Lucchesi

I can see us going there. Yeah, it sounds like your risk tolerance is the same. Of course, yeah. I wonder where we'll be in a fear. I wanted to ask just kind of a general one. You switch gears a little bit. Strans Decider is open source. Yes. Yes. So developers can inspect, they can adapt it.…

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
“A decider model can't do anything that a large language model can't do. A large language model can do it, but it takes longer because it's reasoning and problem solving and token generating. With these newer models, they're making one forward pass through the weights of the model and it's bang and it's out. There is no autoregressive looping like you have with language models. That's why they're so much faster. Therefore, that's why they're so much cheaper.” — Mike Chambers, The New Stack Podcast

Asked to define a 'system one' model for agents, Chambers explains that small decision models get their speed and cost advantage from a single forward pass with no token-by-token generation.

The New Stack Podcast · 2026-11-04 Permalink → Listen →
The New Stack Podcast Around 10:27 into the episode
Speaker 4

wow,

Nick Lucchesi

that's two interesting use cases there. Oh, I just want to jump back a moment. System one model, what's your word definition of a system one model again?

Mike Chambers

Yeah, I guess so. So when the blog post came out, I think this was from TypeSafe when they launched Jerv, they sort of like coined this thing of system one versus system two. And this is sort of going back to the system one versus system two thinking that humans might do. So system one thinking being the things which are sort of like innate. I guess a really good example I read about this was: you know, if I see the brake lights of the car in front of me and I jump on the brakes in my own car, that's system one. It's just instinctive, I do it. So it's not quite as instinctive as just breathing, but it's just something that I've learned to do. I just jump on and do it. Whereas if I'm doing parallel parking, reverse parking, or whatever, that's definitely system two thinking. I have to think about it in depth. So it's that. And the way that that equates into these models then is that you, okay, important point. A decider model, with the decider models that we have today, they can't do anything that a large language model can't do. If I can put it that way. So a large language model can do it, but it takes longer to do it because it's reasoning and problem solving and token generating and doing all that kind of stuff. And maybe it's inside of…

Nick Lucchesi

I appreciate that. And my next question is sort of jumps off that point a little bit. Suppose that the decider model that the decider model needs to look up more information and act more like an LLM to piggyback off your previous answer. What would a developer give Decider to help it make that…

Mike Chambers

Yes, I guess, so how do we prompt it, essentially? I mean, like, I mean, yeah, again, with a large language model, you can kind of send in whatever text you want and let's see what happens and what comes out the other side. With the decider model, with a decision model, you also send in natural…

Nick Lucchesi

What does the, after it goes on this journey, what does it return to the agent?

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.

Collections They Appear In

Hands-On Engineering Podcasts