What AI CNC Programming Looks Like Inside Real CAM Workflows

[00:06] Michael Finocchiaro: We're live. This is IMTS Chicago, day 2. This is my second interview this morning. I caught the Siemens guys, partners of yours actually. This is Michael Finocchiaro, and I'm with my friend David Priev of Limitless Labs, formerly Limitless CNC. Want to tell us a little bit about what you do, and then the name change?

[00:27] David Priev: Sure. Limitless Labs is an AI startup based in Israel, now establishing its presence in the States, in New York. We're developing the first agent for the world of CNC precision manufacturing. An agent that can understand the CAD and geometry context and help you program a CNC part in an optimized and fast way.

[00:55] Michael Finocchiaro: So why Labs? You were called CNC for so long.

[01:00] David Priev: We started as Limitless CNC. We're still passionate about CNC, but our vision just got bigger.

[01:07] Michael Finocchiaro: Which we're going to talk about in the interview. So let's start with the problem. What is fundamentally broken about the way we do CNC programming today, that made you think we needed an AI-native approach rather than just automating the CAM process?

[01:23] David Priev: I don't think it's broken, but...

[01:25] Michael Finocchiaro: It's outdated. Like 20 years out of date.

[01:28] David Priev: Exactly. The way you program parts today hasn't changed in about 20 years. CAM technology was a breakthrough 25 years ago, and from that point on it stayed the same. The problem we have, especially in the West, is a gap of an entire generation in the middle. You cannot find skilled machinists in their 30s or 40s. You have the juniors and you have the experts, and the experts are retiring. The knowledge is going with them. A lot of manufacturing businesses find themselves without the expertise to run parts they already mastered. That's a huge problem. The second problem is that designs are getting more and more complex and lead time expectations are getting shorter. You need to do more with less. How do you do that? You cannot train a new generation in two years. You need to go to technology. That is exactly where Limitless is positioned: to help manufacturers be empowered with AI.

[02:28] Michael Finocchiaro: I remember when I interviewed you on my podcast, you told me the origin story, when you were in the military trying to do this stuff. So it's from real-world experience.

[02:46] David Priev: Yes. I had the opportunity to see firsthand how important those machine shops are in critical times, especially in wartime, and on the other side how neglected they are. It's hard for them to find, train and keep talent, and everything is still so manual. That's crazy.

[03:05] Michael Finocchiaro: Frustrating.

[03:06] David Priev: And they're frustrated as well.

[03:08] Michael Finocchiaro: You've described Limitless as a physical world model for CNC. What does that actually mean technically? What does your model understand about machining that an LLM wrapped around a CAM API doesn't?

[03:21] David Priev: I'll start with the end. If you ask an LLM today to explain CNC machining and the different phenomena, it will give you a great lecture. But when it comes to "tell me what's happening in this toolpath," the LLM will say: sorry, I don't know how to understand what's happening there. Two reasons. First, LLMs lack the ability to understand CAD in its native way, the topology of CAD. I'm not talking about screenshots. I'm talking about how the surface behaves at a micro level. That is one thing we solved at Limitless. We developed an AI technology that understands CAD natively.

[04:00] Michael Finocchiaro: In the context of manufacturing, for NX, Mastercam, Creo...

[04:04] David Priev: It's all B-rep in the end. A STEP file or a Parasolid file that represents the topology in an analytic way. That was one breakthrough. The second part is the physical part. We developed a GPU-based physics engine that describes machining in terms of physics. Running on the GPU, it can estimate tool deflection, tool wear, heat generated and material deflection. And it's not trying to replace finite element analysis. It's not finite element at all. It's trying to build the right intuition into our agent, so the agent makes the right decisions about machining strategy. Because that's what machinists do. They don't calculate the phenomena or run analysis. They have the right intuition about how things are going to behave.

[04:57] Michael Finocchiaro: So it's not a surrogate. It's still physics-based.

[05:00] David Priev: It's physics-based.

[05:03] Michael Finocchiaro: Take me through a real part. From the moment I open the CAD model in NX or Mastercam or Creo, how does Limitless do feature recognition, setup planning, tool selection, all the way to post-processing?

[05:20] David Priev: I could describe it the technical way, but let me tell you a story about one of our users at Blue Origin. A few weeks ago he told us he gave the agent a program in NX that was 30% programmed by him manually. He did the start on his own and then called the agent: continue from this point and finish the program. The agent picked up the program in the middle and continued. It understood the remaining stock, the target CAD geometry, the available tools, the machine kinematics, the physics of the material and the workholding, and finished the full program in a few minutes. The cool part: the machinist took that program straight to the machine and ran the part. 60% of the toolpaths were AI-generated, and it ran everything properly. And the part was made of Inconel.

[06:30] Michael Finocchiaro: Inconel is a superalloy...

[06:33] David Priev: To know how to machine Inconel you need at least a few years of running Inconel parts. Otherwise you break your tool, break your machine or scrap the part. Because the agent has this physics engine, it understands the physics of different materials, and it got it right. We had never tested the agent on Inconel before, but it had the right intuition. That's a great example of how this technology can help.

[06:57] Michael Finocchiaro: You mentioned the workforce is aging, which means there's a lot of tribal knowledge locked in people's heads. How do we get those years of experience out of their heads without retraining the next generation for years?

[07:16] David Priev: We focus on two approaches. The first is data-based. Customers pick their historical CAD/CAM projects that represent their best practices, hand them to us, and we train the agent for them on those projects. Eventually users will be able to onboard their data from the PLM and the agent will train itself automatically, so it won't need to be a project for enterprise users. The other approach is training on the fly. The same way you build skills today with Claude or ChatGPT, you can build skills with our agent. You can tell it: from now on, on titanium parts, I want you to use this machining strategy and pick these tools for this specific machine. You can encode all of it in skills and free text, and the agent will use it. The latest thing, which we're trying experimentally, is a teach mode. You tell the agent "I want to teach you something new" and you do your own work in the CAM. It records everything and extracts the patterns from what it saw on your screen.

[08:39] Michael Finocchiaro: And creates another skill.

[08:40] David Priev: Yes. It tells you: I identified these patterns, are these what you want me to learn? Then users can turn the skills on.

[08:49] Michael Finocchiaro: Very cool. Where is the human still better? Where is AI not going to help?

[08:57] David Priev: The human stays the reviewer. There are so many nuances in the process that AI just doesn't understand. You know your machine better than the AI. If this machine has a problem with the B-axis, you know how to work around it. Those are things that will take AI a long time to understand. Another one is internal stress in your stock material. You receive stock from a vendor and you know it's going to be problematic, it's going to bend in a specific way, because you're familiar with that stock. Those things are hard for AI. But as we keep doing this closed-loop iteration, bringing machine data and CMM reports back in, we'll eventually figure out how to help with that too.

[09:51] Michael Finocchiaro: Is the longer-term opportunity to become the intelligence layer underneath every CAM system?

[09:59] David Priev: We see the opportunity to become the AI intelligence that drives precision manufacturing processes. At first we thought: let's empower CAM software by providing our API. Now we see that the direct experience we created, with our agent integrated on top of the CAM, is much more powerful. That's where we're heading.

[10:28] Michael Finocchiaro: Before the future question, there were some recent announcements around Sandvik. Do you want to talk about those?

[10:36] David Priev: We're working with Sandvik and ISCAR. With ISCAR, we're about to launch an upgraded tool advisor on the ISCAR side. They have an existing tool advisor built on their machining knowledge base that recommends the right ISCAR tools for each application. The missing part is the ability to understand CAD geometry, and that's exactly what Limitless provides. We're embedding our core technology into their web application.

[11:13] Michael Finocchiaro: Where users can upload CAD. Limitless inside.

[11:16] David Priev: Exactly. Users upload the CAD file, we analyze it, and ISCAR takes it from there to recommend the right cutting tool based on their knowledge.

[11:24] Michael Finocchiaro: Something similar with Sandvik?

[11:25] David Priev: With Sandvik it's a different collaboration. It's with Cimatron, one of their CAM software products, where we're empowering them with our AI technology through an API.

[11:35] Michael Finocchiaro: To wrap up: what's coming down the road? More productive programmers, or programming becoming something machines do autonomously after an engineer defines it?

[11:47] David Priev: If you allow me, I'll rephrase it into the more direct question people ask us a lot: is AI going to replace me as a machinist?

[11:57] Michael Finocchiaro: Of course.

[11:57] David Priev: My answer is: AI will replace you as a machinist, but not the way you think. If you're the machinist who leverages AI tools and becomes 10 times more productive, you won't be replaced. Actually, every machine shop will be trying to hire you. But if you're a machinist who ignores these new technologies and insists on working the old way, you will be replaced by the machinist who leverages AI. That's how I see it.

[12:29] Michael Finocchiaro: It's been a great interview. Thanks, David. Thanks, everybody.