Link to original article

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: It's time for a self-reproducing machine, published by Carl Feynman on August 8, 2024 on LessWrong.
I've wanted to build a self-reproducing machine since I was 17. It's forty-five years later, and it has finally become feasible. (I've done a few other things along the way.) I'm going to describe one such device, and speculate as to its larger implications. It's a pretty detailed design, which I had to come up with to convince myself that it is feasible. No doubt there are better designs than this.
The Autofac
Here's a top-level description of the device I'm thinking of. It's called an Autofac, which is what they were called in the earliest story about them. It looks like a little metal shed, about a meter cubed. It weighs about 50 kilograms. There's a little gnome-sized door on each end. It's full of robot arms and automated machine tools. It's connected to electricity and by WiFi to a data center somewhere.
It has a front door, where it accepts material, and a back door, where it outputs useful objects, and cans of neatly packaged waste. You can communicate with it, to tell it to make parts and assemble them into useful shapes. It can do all the metalworking operations available to a machinist with a good shop at their disposal. In return, it occasionally asks for help or clarification.
One particular thing it can be told to make is another one of itself. This is of course the case we're all interested in. Here's what that looks like. You feed a 60kg package of steel castings, electronics, and other parts, into the door at one end. It starts by building another shed, next to the other end. The two sheds are butted up next to each other, so the rain can't get in. Once it's enclosed, there is no visible progress for about a month, but it makes various metalworking noises.
Then it announces that it's done. The second shed is now another Autofac, and can be carried away to start the process elsewhere. There's also a can full of metal scrap and used lubricant, which has to be disposed of responsibly. This process can be repeated a number of times, at least seven, to produce more offspring. Eventually the original Autofac wears out, but by then it has hundreds of descendants.
The software
The key part of the Autofac, the part that kept it from being built before, is the AI that runs it. Present-day VLMs (vision-language models) are capable of performing short-deadline manual tasks like folding laundry or simple tool use. But they are deficient at arithmetic, long term planning and precisely controlling operations. Fortunately we already have software for these three purposes.
First, of course, we have calculators for doing arithmetic. LLMs can be taught to use these. In the real world, machinists constantly use calculators. The Autofac will be no different.
Second, there is project planning software that lets a human break down an engineering project into tasks and subtasks, and accommodate changes of plan as things go wrong. We can provide the data structures of this software, initially constructed by humans, as a resource for the AI to use. The AI only has to choose the next task, accomplish it or fail, and either remove it from the queue or add a new task to fix the problem.
There are thousands of tasks in the life of an Autofac; fortunately the AI doesn't need to remember them all. The project planning software keeps track of what has been done and what needs to be done.
Third, there are programs that go from the design of a part to a sequence of machine tool movements that will make that part, and then controls the machine tool motors to do the job. These are called Computer Aided Manufacturing, or CAM. Using CAM relieves the AI of the lowest level responsibilities of controlling motor positions and monitoring position sensors. This software doesn't do everything, of...

#146 — RESUMIDO Entrevista: Sandra Ávila / Inteligência Artificial