Hacker News Reader

Comment by ReliantGuyZ | Hacker News Reader

I'm unclear on what people see in the current AI tech advancements that makes them think it will contribute to better manufacturing. The new feature of LLMs that makes them so interesting is their ability accept input and flexibly follow arbitrary instructions, meaning they're really good for varied work, especially when there are a wide range of acceptable answers ("creative work"). Everything I know about manufacturing at scale is that you want a person or machine that follows a tiny instruction set (at least in comparison to the potential flexibilities of an LLM) and nails the execution every time. This seems to me like the complete opposite of the strengths of an AI system like the ones that Wall Street are cheering.

View on HN

I've heard that the general transformer architecture (not specifically LLMs, which imply a language model, but applied to sensory perceptions and outputting motor commands) has actually been fairly successful when applied to robotics. You want your overall assembly line to have a tiny, repeatable instruction set, but inside each of those individual instructions is oftentimes a complex motion that's very dependent upon chaotic physical realities. Think of being able to orient a part or deal with a stuck bolt, for example. AI Transformers potentially would allow us to replace several steps in the assembly that currently require human workers with robots, and that in turn makes the rest of the assembly much more reproducible (and cheaper).

Training these models takes a bunch more time, because you first need to build special hardware that allows a human to do these motions while having a computer record all the sensor inputs and outputs, and then you need to have the human do them a few thousand times, while LLMs just scrape all the content on the Internet. But it's potentially a lot more impactful, because it allows robots to impact the physical world and not just the printed word.

And it's a nice problem to solve with AI of many kinds because you can forward-solve the kinematic solution and check for "hallucinations": collisions, exceeding acceleration limits, etc. If your solution doesn't "pass", generate another one until it does. Then grade according to "efficiency" metrics and feed it back in.

As long as you do that, the penalty for a a slop-based fuckup is just a less efficient toolpath.

I am not an expert in this, and don’t necessarily believe it. But the pitch is that existing manufacturing automation requires that specificity due to technical constraints. And that much of the factory automation that hasn’t happened is because it’s too costly to get to that level of specificity in that the existing automation requires higher scale to be cost effective. If you had more general purpose intelligence you could get around those constraints.

The video models are the ones that seem to be attracting the most attention in this area as it seems do similar to sight recognition.

> existing manufacturing automation requires that specificity due to technical constraints

Rather the opposite, I'd say: existing manufacturing automation is built around repetitive motions because an assembly line is making multiples of the same product. Having AI reinvent the wheel for every individual item is completely pointless.

One-off manufacturing can to a certain extent be automated. We're already seeing that with things like 3D printing and dirt-cheap basic PCB assembly. However, in most cases economies of scale prevent that from widespread generalization to entire products: ordering 100 or 1000 is always going to be have significantly lower per-unit costs than ordering 1, and if you're ordering 1000 you can probably afford a human spending some time on setting up robots or optimizing the design for existing setups.

There are undoubtedly some areas where the current AI boom can provide helpful tooling, but I don't expect it to lead to a manufacturing revolution.

I think you could make an analogy to the difference between ASICs and general-purpose CPUs. ASICs are great, but CPUs have flexibility and massive economies of scale. Similarly, a specialized machine might be more efficient than a humanoid robot at a particular task, but advanced humanoid robots could theoretically do all the tasks and as a result would likely end up being manufactured in very high volume.

Imagine a future where any hardware startup could design and provision an assembly line as easily and cheaply as software startups today use cloud computing. Maybe after a certain scale it becomes economical to consider replacing steps of the manufacturing process with "ASIC" solutions, but maybe there'd be a long tail of things which would continue to remain best served by general-purpose robots indefinitely.

What if everything in the economy could be efficiently produced through one-off manufacturing methods? This is something I have been thinking about a lot lately as i've been working on 3d printing/CNC/aluminum casting.

It avoids the need for any sort of parts shipping, and can be easily retooled to make war machines in a time of emergency (which is one of the motivations to bring back american manufacturing).

The main problem with this is that you still need surface-mount components if you are going to make PCBs, and you still need magnets to make motors, etc.

Manufacturing robotics is all about movement. All movement exists on a spectrum of difficulty and context needed to perform. For instance, welding the steel plates together in an empty and repeatable consistent 3d space is now on the lower end of difficulty. Navigating through a partially manufactured vehicle cab to install a complicated dash assembly requires a lot of context and is incredibly difficult for a robot to do.

The more we can bring down all the difficulty of all these processes, the more we can accelerate manufacturing locally.

That's at odds with everything I know about manufacturing robotics, having worked with people doing that work. The complexity of the environment is irrelevant because the robot is programmed to make a specific motion and to adjust that motion in predictable ways based on the appearance of specific features. That is by design, not because (or at least not just because) the robot is incapable of planning its own motion. The whole system is designed to be predictable instead of adaptable because that's what you need to do to do the same thing millions of times.

> The whole system is designed to be predictable instead of adaptable because that's what you need to do to do the same thing millions of times.

That final "millions" is the problem. Automation is great and easy when you will do the same thing millions of times. Sure it might cost half a million to program the robot (which itself cost half a million) - but that is $1.00 per part, and it goes down as you make more. When you are only building 10 though a million dollars is a lot of money and so you want humans - or robots that are "CAPABLE of plannings its own motion".

Costs have been going down. In high school I took the class on how to write g-code (I have one free period so I took shop for non-college bound kids for fun even though I was college bound - it was a great time that I highly recommend even though it was only for fun). These days almost everyone just uses their CAD/CAM and isn't even aware that the g-code is supposed to be a human readable programming language. (it probably isn't)

And yet China already has 100% automated vehicle production lines.

https://www.telegraph.co.uk/business/2025/10/12/why-western-...

> Navigating through a partially manufactured vehicle cab to install a complicated dash assembly requires a lot of context and is incredibly difficult for a robot to do.

Not really. The robots are programmed by having a human manually guide it, so the robot itself doesn't really have to do any navigation - it just has to follow a predefined path.

Want to install different variants of dash components? Split it up into methods and have the robot return to a neutral position after each method. You're literally programming it.

That's not how the LLMs should be used in manufacturing. It is still the current assembly lines robots that will do that. LLMs can be used by the humans who design the automation workflow, as coding assistants. That can lower the breakeven number of items that can be automated. Maybe if today it only makes sense to automate the manufacturing of a widget only if you can sell more than 100000 of those widgets, then with LLM assistance that number can be reduced to 1000. Whenever you have a 10x improvement of something, there's scope for a mini-revolution to happen.

I'm not even clear on what people mean when they say 'AI' anymore

Reply on news.ycombinator.com