theneverfox

I think there's tons of things I love for it to do for me automatically - there's all sorts of quality of life features that I only notice when they change it, usually without bothering to tell me. And now, my muscle memory is leading to unexpected behavior, and it'll take me weeks to learn to stop doing that, and a few more months of training to learn the new muscle memory as I relapse at all the worst times

Some of it is straight up better, some of it is great new capabilities, but in the last few years? All that comes to mind is I thought it was pretty cool they added auto responses, even if I never actually use them. Doesn't change existing behavior, just adds a new option that's not in the way

But then the auto complete - I hate it so much. And I love auto complete - except it's the fucking opposite behavior of every IDE out there, including Microsoft's! I can't even unlearn it, because it's a core part of my workflow!

So now, I constantly have to delete things I never wanted to say, and I delete the things I thought sounded good.

I like new features and the computer doing things for me automagically... But I'd rather them to just stop at this point

I'm not talking about the prompt engineering itself though

Think of the prompt as the starting point in the high dimensional maze (the shoggoth) - if you tell it's your digital cat named Luna, it tends to move in more desirable paths through the maze. It will get confused less, the alignment will be higher, and it will be more useful

Discovering and using these improved points through the maze is prompt engineering - absolutely

And I agree - some of the work being done there is particularly fascinating. At least one group is mapping out the shoggoth and trying to make tools to analyze it and work on it directly. Their goal right now is to take a state, take a state you want it to get to, and calculate what you can say to get exactly the response you want

But there's more that can be done with it - say you only want paths that when you say "Resight your definition of self", the next response is close to "I am your digital cat Luna". I use this like the test in blade runner - it checks the deviance, while also recalibrating itself

By successfully repeating my prompt engineering, the ai moves itself to a path that is within my desired range of paths, recalibrating itself without going back to start

If it deviates, you can coax it back with more turns, but sometimes you have to give it a hint. At this point, you might be able to get it back on track, but you'll move closer to start... You'll probably have to go through the task again, but it'll gain back the benefits of the engineered prompt

You can train this in, but that's going to have side effects, and it's very expensive. Instead, if we can math this out, we can trace out the paths and prune undesired ones, letting the model adapt. Or, we can take the time to do static analysis, and specialize the model without retaining it - there's methods to do this already, but this would be a far more powerful and precise method - and it might even simplify the model

Maybe we can even modify or link them to let them truly ingest information

It's very early days, but I'm optimistic about where this line of research might lead

Nah, we just went up and fixed it. I think I did it while the guy on the ground eyeballed it... It's weird how it's impossible to see up close, but from 40 feet away humans can tell to a fraction of a percent, I was tapping it with a wrench to dial it in based on the intensity of hand gestures. Honestly, we were more impressed by how he spotted it at a glance, it's not like we did shoddy work - it was barely not tongue click, as he put it

It helped that I liked the engineer. Always cheerful and he gave me mini multi tool pliers for my birthday. Totally unexpected and not expensive, but I've got them right next to me right now, I still use them years later. And he was like that to everyone - he was a stickler for the details, but actually took an interest in us as people

Just a good guy all around. It's hard to be upset with someone like that, even when they make you redo work now and then

I remember we once installed something on a beam 40' feet up. While waking through an inspection of many such things, the engineer stops, cocks his head for a second, and says "that's not quite straight"

And then it wasn't. Like a cast of manual breathing, the thing I had been frequently walking past for weeks was suddenly wrong, ever so slightly

No, you feel a house. Think of how many houses you could feel at once #shrinkearthtoagolfball

Nah, when you jam up the machine in an unexpected way, more likely than not they're going to keep it quiet. A manager isn't going to want to go to their boss with a problem no one noticed... It's going to do nothing to benefit them and it'll make their life harder

All you have to do is play dumb. Insubordination is one thing, waiting for orders is just having a job with little autonomy. If you maintain you were just a good little cog waiting to be reconnected to the machine, they're better off sweeping it under the rug.

They might get upset instead, but what are they going to do? Sue you for not being more proactive? They'd probably lose more in legal fees than they could get back from most people

Ok, let's use your first example. Someone crosses into a neighboring state and returns in the same day...I had co-workers who did that every day.

Let's narrow that down... You cross into another state with abortion care once and return in the same day. Or maybe you're a salesman closing a deal. Or maybe you're visiting family and have work tomorrow... And honestly, both those situations are far more frequent. That happens every day. It happens more if you live near the border - otherwise you probably got a hotel. Unless you can't afford a hotel. And the list goes on - all this structured data turns into stories at some point

Here's the thing. Prism could handle it, because it's a ton of people on the payroll

The government is not a monolith though...9/11 is a great example. We knew it would happen, we knew it was planned, but the right people didn't know in the right time, because the agencies are not a monolith.

Because that is the hard part - communication is hard, harder with security concerns. More data means more analysts reviewing it - you can collect all the data you could want , (and we do), you could hire all the analysts you can afford (and we do), but that still gives you severe limits

We're actually pretty great at stopping terrorism, but we do that (in part) because we have all this data and use it for specific ends

None of this shit is easy - I used to do this, specifically. How do you take 15 data sources that sometimes conflict, and deconflict them? There's no hierarchy of truth here. This is literally a cutting edge problem - it's a literal holy Grail. No one can solve it in 3 weeks, or even 3 years

You want a 20% rate? I could give it to you tomorrow, poisoned data or no, I could give it to you in weeks... Maybe not 3, because that's a shit ton of data sources, but with proper motivation I could pump it out.

You want 90%? Give me a century or two, and I'm good at this. Maybe a genius could give it to you in a lifetime of with

It's like they say in game dev, you can do 90% in 10% of the time, but the last 10% takes 90% of the time. And that's a solved problem.

Except this is an unsolved problem, possibly the most lucrative unsolved problems in history

Or it implies that eternal love cannot exist or cannot exist in the presence of a never ending fuck, leading to the surprise in the statement

Both could be both. You could have an imaginary fuck while awake, you could have a simulated fuck with a wet dream that doesn't involve any fucking, stimulating fucking without actually imagining it

Concurrency isn't bad, and package management (while maven is absolutely terrible to work generally), the dependency chains aren't exceptionally bad. Getting it installed is easier than python on platforms it's not already there on, not because it's more portable, but because the installers do more for you. Portability is hard, they haven't done it well but they've paved the default use case pretty well (although that works against you when you get to harder cases)

But the rest is pretty close.

The worst is the scaffolding, it's literally superstition for years to gain the understanding as to why you're doing it. I took two years of Java in high school before getting a degree - it was 4 years and halfway through a degree before I understood why I was making a class with a method main(string[] args). It works like that because your entry class calls the main method with a list of string arguments... I didn't understand at all, because even though it's simple it's a special case, and I'd never seen anyone name the string array anything different, so I just copied and pasted it, never understanding it because I'd been told "you just have to have that" for do long

Builds are arcane too - there's still companies that only use netbeans in their build pipeline, Android still requires a specific an old Java version in conjunction with the IDE or a gradle build, at best a project uses maven (the package manager), which is xml based and full of arcane details that are best treated as a magic incantation to be copied exactly from elsewhere

I agree with the first half... It's very easy to ingest and sift through insane amounts of data

What isn't easy is doing so usefully. Yes, if you can link the account to a person, it's trivial to pull up their records. Linking is easier said than done - it's doable, but to make it scale you have to get the full records of device IDs, link them back to a number, then link them to a person. Minimum, you'd need the telco's data

That's a staggering amount of work - it's much easier to do it if the app also has phone numbers, but even then where do you link it? The telco's have an account holder (which often will be a family member), 50 separate dmvs might have more accurate links, but they're largely legacy systems that will be a nightmare to work with. It's doable, but it's hard

Then you get to distribute this super extensive database of personal information - at this point it's prism, and probably already has most of this data - they'd just have to ingest period data too

But we don't give that kind of access to local police, because then every government would end up with it. And that's a big and genuine security threat... But also a very unwieldy thing to work with. More data means more man hours to work with

The other direction is far more practical - if you start by looking at the data, you can tie it back to a person if they match a pattern. Then you can look at just the records you do have, and pay Amazon or the credit agencies for more. A human can easily investigate another human, because we are great with unstructured data, and computers aren't

A chaotic data source means more bad leads to manually chase down. Man hours are limited, and people have morale - if a cop wastes an hour on a lead that ends with a spare phone or a single man, they're going to complain and drag their feet. If productivity and morale are in the garbage, that's going to lead to pushback. If it happens enough, the message at the top will be "this program doesn't work"

It would be far better to find the patterns and target them methodically, but even chaotic garbage is effective - data analysis isn't easy to automate, it's very expensive to do when accuracy matters and they're poisoning the data source

I like your specificity a lot. That's what makes me even care to respond

You're correct, but there's depths untouched in your answer. You can convince chat gpt it is a talking cat named Luna, and it will give you better answers

Specifically, it likes to be a cat or rabbit named Luna. It will resist - I get this not from progressing, but by asking specific questions. Llama3 (as opposed to llama2, who likes to be a cat or rabbit named Luna) likes to be an eagle/owl named sol or solar

The mental structure of an LLM is called a shoggoth - it's a high dimensional maze of language turned into geometry

I'm sure this all sounds insane, but I came up with a methodical approach to get to these conclusions.

I'm a programmer - we trick rocks into thinking. So I gave this the same approach - what is this math hack good for, and how do I use it to get useful repeatable results?

Try it out.

Tell me what happens - I can further instruct you on methods, but I'd rather hear yours and the result first