Being realistic about automation.

Bagger 288 is 315 feet tall, it digs 240,000 cubic meters a day.

It is operated by 5 people, I bet they like their job. An apartment building such as “Watergate East” holds 240 units, each unit is appox a “2 and half” and is the same height.

It was built in 1978, we will have to check the history, but I don’t find any evidence Anthropic or Openai played ANY role in this massive highly automated system.

As far as I know, this is as successful automation as we have achieved.

99.99% automation is possible you just need to make a big ass machine. Hows the 100% automation attempts with artificial “brains” and reading the whole internet going?

Hmmm, maybe flying to close to the sun, lets continue working on the 99% automation solutions. You could import a dozen or so 3rd world slaves to pick your crops or have 1 man who knows how to fix a machine in an air conditioned tractor.

I don’t get whats so hard for people to understand here, plan for human oversight in real processes. While your designing your grand visions of the future, put a well trained human in an air conditioned box on top of the machine; this human will need tools, so there should be space for a tool box, the machine will be made of parts so imagine how pieces of it can be swapped out.

I dont fucking want an ai written power point or podcast. I don’t want a 5 financial middle men in between to drive(or god help me, bike) to me a cold burger with a bunch of cancer chemicals. The wheat needs to get out of the ground, then put in a bag, that bag needs to end up in my possession; these were all automated long before the Internet and by scaling up pass a single time sensitive meal can be a much high automation percentage then fucking ubereats.

You all know that Mr. “bird” started freezing vegetables before people had freezers right? And you still can buy a mixed bag for 99 cents right?

These silicone valley types are claiming to increase the automation of society, but these app’s are adding human steps to an otherwise fairly high quality process. When I was younger “the swan man” would bring you frozen food you ordered via catalog. Moving hundreds of meals at a time in a big truck. Rather then 1, subsidized, meal at a time. Every time you order via an app, your regressing our society. I hate your stupid subsidized apps silicone valley. Stop “capturing market share” and start decreasing work by building longterm pro-humanity machines.

They are “planning” on it being self driving soon, but then they are hiring “Actually Indians”, so what ratio of drives will they actually pull off? Whats more realistic, a truck with 1000 frozen meals inside or a self driving car that does 1000 1 off trips successfully in a row?

——

So where should these ai’s be used.

Its extremely important to note these are not brains. they are our best black box approximator. Like all black box approximators they simplify what they are looking at.

Consider the function:
```
int f(int x){
if(i==88092187){
return 3;
else {
return x;
}
```

Every approximator and humans, merely seeing the input-output of this function would quickly assume that the function is the identity function. You see f(3)=3, f(1000)=1000, your just going to assume f(x)=x.

We should employ this tool where it can “self-play” Simplifying the chess ai’s are treating chess as a function and estimating whats the best move. Likewise the dna folding, its estimating the xray machines output given a dna string. The games are more complicated, but these fields are doing well because self-play is on the table. When ai’s “self-play” at reading the internet, well you get the worthless reddit clone.

The strongest near universal self-play game is “style transfer”. Again the mechanisms are more complex then this, but imagine you define the game like this: The database picks 2 images that each have a single subject, lets say a cartoon bar and a real bear, and you the player given the 1st image must produce a real bear, and a 2nd player given the real bear must produce a cartoon bear. Then a judge gets all 4 images and is asked to produce “bear”, “cartoon”, “real” and a prediction on which images was the original ones.

Like chess, estimating playing it well is the same thing was doing well at this game. We can just not turn on the 2nd player and the judge and just ask it to “style transfer” starry sky to “bio-punk masterpiece hi-res no-mistakes”.

And there is an art style that easier for a human to edit, line drawing, black and white; etc. whatever makes it as outline only. You do a quick edit and then style transfer it back. That is how it should work. And there’s a type of code that’s easier for humans to edit, rants in markdown docs.

Back and forth from the blueprint to the artifact. Making quick edits with the tools we already know.

—-

Your all stealing code and style transferring it to your preferred language, or getting fragments of an idea for an image and feeling in control because the is good. The tool is fantastic but it wont get to 100% automation.

We need better control over the style transfer game and to pick the “genre” we target our tools for to the least effort. And for the silicone valley culture to fuck off.