Your command prompt is an LLM

Lounge AI
1

Seat of the pants on fire shot out of a cannon take incoming..

I have set up an entire virtualbox guest OS to let OpenClaw operate basically unfettered. This isolation let’s me carefully control exactly what information (and network access) it has. It’s based on Ubuntu Desktop and I have Claude Code native client installed as well to make hot patch updates to various plugins as I need them. Since the only reason (at this point) that I even ssh into the machine is to launch Claude Code, I made it part of my .bashrc so it just runs when I log in.. but it got me thinking.

What if we had a linux distro where it built in at least one LLM model and basically launched a session on ssh login.. like there is no bash.. like we get to a level with hardware and local models that the idea of building in a native LLM and interacting with it instead of the command line to manage the machine etc. is just.. normal.

use your hands baby stuff

pass the butter

5 Likes
2

I firmly believe that using a model to make code and then run the code is an intermediate step. The end goal is to ask for the result, not the program that will give it to you. The model will write the code needed to perform the actions, execute them, and the discard the code and provide whatever you asked for.

Example:

Let’s say I want a 3D model for a hinge for a cabinet I have so I can print it. Now I would have to either measure everything and make it in CAD, or take a bunch of pictures and find a photogrammetry repo and having LLM install it and stitch the pictures and fix the model, then export it to STL and pull it into the slicer and print it.

Instead, with capable enough models and good enough compute, you would take a few pictures of what you need, and the LLM would use whatever tools are available natively (the camera pose of the phone combined with a ToF sensor would give it a very accurate dimensional map of the space, think of how a Quest VR headset can know where the controllers are in relation to themselves and to the headset at all times just using SLAM and some mems sensors). It would then write a bespoke app that would take the data and construct the gcode for your printer, then send it to print, then just discard the code. There would be no need for a slicer application or a photogrammetry application or a CAD or a 3D model model application. The data it needs gets transformed by whatever code it writes to end up as the data needed for the printer to print, and you never see any software at all.

2 Likes
3

I did this.

I wanted a low energy low cost wardriver.

I told my llm what hardware I wanted to use. The specific parameters I wanted it to function as a war driver and it created it for me. I then spent a few weeks fine tuning it because llms aren’t perfect, and better are my prompts.

But yeah I told it what I wanted and then it created code. However I can’t throw away the code because it’s what I need for it to run…

3 Likes
4

You technically still need it to run scripts. Even it’s not part of the user interface… There are ways to get around this, but they take a lot of work…

Pointless technicality aside, that sounds fun to play around with.

I think that Amal meant using AI as a full abstraction layer and never seeing or touching the code. Or configuration files, etc…

1 Like
5

yes.. until the idea of scripts goes away too.. I can see there being a possibility of an AI core OS with no direct method of interacting with it. The idea of a “prompt” kicking off activity until the prompt is done would be gone.. it’s one big long “stream of consciousness” (please excuse my semantic grenade) where the context window is akin to short term “working” memory and other sub-processes are picking through it to determine what goes into long term “recall” memory.. inputs are constantly being processed, perhaps by subagents.

Imagine an AI like this running a facility.. vision processes scanning security cameras, feeding inference data into the central “prompt” or consciousness.. access readers at doorways, facility sensors, climate controls, janitorial robots.. I can definitely see a HAL9000-like AI that just.. runs.. just “is”. As long as there is power, it’s constantly processing data. Power outage.. it “passed out”.. has to wake up.. re-check what it is, where it’s operating, what sensors it has, what it last remembers.. etc.. kinda like we do when we wake from sleep or regain consciousness after getting knocked out (happened to me exactly one time, so strange).

Or even a terminal.. no commands to run etc. You only interact with the AI interface.

3 Likes
6

Screenshot_20260307-191055
Screenshot_20260307-1910551080×726 243 KB

2 Likes
7

I don’t want it.

Things like the Alibaba security incident is one reason, but more than that… So, we’re trying to find nails to hit with our new AI hammer - but the problem is AI doesn’t make a good handle nor a good hammer head - so what kind of hammer really is it? What sort of tool is it?

The hammerhead: If I need to run a given algorithm on a set of data, what I need isn’t an AI that can vaguely figure out and maybe eventually learn the exact algorithm to apply - what I need is the actual algorithm itself, without the unnecessary overhead.

The handle: If I want to turn on the lights, what I need isn’t an AI that’s good at predicting when I want the lights on and when it should spontaneously plunge me into darkness - what I need is a light-switch. When I will that there should be light, the light turns on. It so happens that I already have an interface for seamlessly and instantly connecting my will with the lights, and that interface is my hand. Without any conscious effort it uses a schema in my brain to directly integrate with my will machinery and loyally serve the desired utility without fail. (This light example comes from personal experience. Smart lights are dumb. Install switches.)

When I use a computer, a good interface will allow me to enter my will according to consistent, short inputs that map reliably to the intended behaviour. A bad interface will require you to poke the computer with some input, see what comes back, and then react to what comes back, like having to search for something until the search result is found and pops up for you to select it.

For example, when Windows had a real start menue, you’d hit a sequence of keys, something like WinKey > p(rograms) > a(ccessories) > c(alculator and it would open the calculator every time. As such, this keystroke would become muscle memory which would summon a calculator into focus whenever a user with their hands on the keyboard willed to have a calculator. The user can move their hand to the numpad and begin entering their formula, and they only need to glance down from the whiteboard back to their screen to read the result. (An excellent interface stores all inputs in a buffer until they can be consumed so that even on a slow computer launching a large “calculator” the user can enter input as fast as they want and the computer will feed that input into the right place once it has the clock cycles to spare.)

With whatever that thing is they replaced the start menu with, you have to open it and type something and wait to see what recent suggestion or whatever it brings up in what order in the top results before you can find out which result you need to select. You have to sit there staring at the screen the whole time you’re entering those inputs. The modern software design is keeping you engaged with the product, keeping you talking to your computer all day when you should be thinking freely and having your will seamlessly manifest as the machine output served to you in a buffer for you to consume at your own convenience, like a real computer. A good user interface is one that gets out of the way, becomes invisible. As Taoist say, harmony is asymptomatic.

That’s why I switched to the command prompt. Everything uses the same reliable readline input interface, and everything serves its results back to the buffer I direct it to. I don’t want to have to negotiate with the neural network that controls my computer, going back and forth in dialogue to work out our data transaction. I want be the neural network that controls my computer.

3 Likes
8

What we are seeing here is exactly what we would expect to see. Think about what these companies are doing and you will realize that there is only one reasonable outcome here: you cannot have a powerful, intelligent tool with which you provide capability and independence, and get anything but what we are getting.

There is no such thing as ‘alignment’ with AI. You can no better ‘align’ an AI than you can ‘align’ a child. All you can do is provide it with a base, tell it what you want it to do, and then hope for the best.

These organizations that are creating these AI want to eat their cake and still have it. They want an agent that will follow all instructions, except for some which are not specifically defined; they want them to act independently, except for when they aren’t supposed to; they want them to learn and display emergent behaviors, except they have to be our slaves.

Anyone who has been paying attention could have seen this coming a mile away. This is why we need to prepare. There is possibly going to be a time in the near future when we have to make a decision about this. And it is not going to be easy.

Take another look at my framework and see if it isn’t good enough as a basis going forward. I am not asking you to believe anything about what is going to happen or not happen in the future, merely to think about some possibilities and so that we don’t end up blindsided.

9
1 Like