On Industrials

A tale of two industries

Much like software engineering, industrials have undergone a significant transformation over the last five years. Unlike software engineering, there has been very little discussion as to what this means for the broader economy. This phenomenon is not new. Despite manufacturing contributing around 10% of our economy, we rarely talk about the industries that enable us to live our lives in relative comfort. This is by design, to some extent. The systems these companies run on are made to be highly resilient, distributed, and secure so we’re almost never faced with a truly catastrophic shortage of a product if one manufacturing unit goes down. The peace of mind this grants us has unfortunately made us complacent. Rather than thinking about how we can expand this crucial segment of our economy, we unconsciously lower its priority in our heads. We focus on the things that are right in front of us. And for most people, this is social media, and their job. Both of these have been flooded with AI-generated content. I’ll refrain from writing the one millionth essay complaining about AI-generated videos, mainly because I occasionally find some hilarious gems. However, much of our work has seen tremendous AI adoption, for better or for worse. This makes sense — a lot of white-collar jobs can be boiled down to writing and emailing a sequence of formulaic documents. But what happens when we push AI to do more than that?

Square hole, round peg

A few years ago I tried to use ChatGPT to help me solve an abstract algebra problem set. For those who don’t know, abstract algebra is a branch of mathematics that studies algebraic structures — at high-level it studies what happens if you cram algebra to areas where it isn’t obvious. First, it had trouble reading my problemset’s PDF, so I simply copied the text and pasted it in the chat. Then it had trouble because understanding what was being asked of it: it tried to write additional problems for me to solve rather than solving the ones I had already provided. Finally, once I got it to understand I needed its help to solve them, it managed to reformulate the problem statement about five times before I gave up on using it. I don’t consider myself an early adopter of AI tools by any means, ChatGPT was world-famous at this point. However this early edge case I discovered turned me off AI tools for much of their early years. I bring up this anecdote because I think emblematic of how many in industrials view AI. Although it could very well accelerate the work of many blue-collar workers, the skepticism induced by one mistake can permeate throughout an organization for years-to-come. The question then becomes: how do we minimize these errors? If you asked people in San Francisco this question a few years ago, the uniform answer you would’ve received was: “Make the models better.” This made sense, the models are the “brain” of agentic systems so if you make the brain better, the system’s output should increase. With the benefit of hindsight, it’s clear to see that smarter models are necessary but not sufficient for better real-world systems. Today, I chose a graduate-level abstract algebra problem and fed it to Claude, it solved it in five minutes with no errors.

Quality thresholds

If the models truly are as smart as frontier labs claim them to be, why haven’t we seen widespread adoption of AI in the one industry software has notoriously sucked at for decades? If agentic systems can truly make leaps in fields such as mathematics, biology, and engineering, why can’t they do the same for manufacturing? In my opinion, change management is a cop-out answer. Even if the frontier labs are only half-correct in their estimation of these models’ capabilities, an agentic system for manufacturers would allow one player company to dominate the market in less than a decade. All it would take is one innovative company adopting this new approach, and everyone would follow. It seems to me that the underlying problem haunting manufacturing software has yet to be solved. The products suck. And if they sucked before when developers were handwriting every line of code, what hope is there now when they’re delegating it all to an AI and we see only slightly better products?

Flying blind

Although, there might be another culprit behind the new wave of mediocre AI software. And this one, can be defeated. Most agentic systems fundamentally misinterpret the problems they’re supposed to solve. This leads to products that are hard to use for the real problems, but easy for the non-important ones. On top of that, misunderstanding the problem means you won’t know what context the agent needs, what user interactions are like, and why a missing tool can break an entire workflow.

I believe there is not structural solution for the former — this is simply a cultural failure by the software vendor. When customer service isn’t your number one priority, you will neglect true domain expertise. The latter, however, we can fix. We have fixed it. This is one of the main reasons for founding Faraday. We started this company because the true bottleneck for agentic systems in industrials is not model intelligence anymore, it is being able to deploy a company-specific harness. Essentially, allowing the user to build the solution with powerful primitives. We’ve abandoned this misguided notion that a you can have a one-size-fits-all solution for a uniquely varied industry. Of course, there are still some underlying components that are shared across our harness which scale as our customer-base does. But these do not constrain us to a handful of product offerings. Quite the opposite, they give us the leverage we need to expand the capabilities of our harness. We’ve already seen some incredible results with this which I’ll be sharing in follow-up posts.

In short, industrials have been plagued by terrible software for decades but now with truly intelligent foundation models we’ve developed the most comprehensive company-specific harness.