AI - politics, economics, data centers, and other misc drama.

**Athas** · 2026-09-19 12:23

**Bombur** · 2026-09-19 22:25

You have heard of Space Force, well get ready for AI Force. This will be a new government military branch that will help monitor the technology. Trump will also appoint a new AI czar that will over see the operation.

**Guitarman** · 2026-09-19 23:54

He is so motherfucking dumb. Like a breathtaking level of stupidity. And so many love and worship him. The level of critical stupidity is astounding.

**Byrthnoth** · 2026-09-20 09:14

I sure hope AI doesn't get anyone killed.... https://www.bloomberg.com/graphics/2...school-attack/

**Melena** · 2026-09-20 12:53

Originally Posted by Byrthnoth

I sure hope AI doesn't get anyone killed.... https://www.bloomberg.com/graphics/2...school-attack/

Well it nearly triggered WW3 as the US almost boarded a Chinese boat.

Sent from my SM-S928U using Tapatalk

**Meresgi** · 2026-09-21 10:31

Asked Gemini to find me some peer reviewed articles for a bullshit class I need to take as an elective i'm not putting effort into. Returned a couple and then asked it to cite them. Gave me back an APA citation...with completely wrong authors. Like literally not one author was actually from the article it linked me with the DOI. I was like "Those are the wrong authors completely" and it was like "Oh wow my bad, here's the correct citation"

How the fuck are you going to get the VERY BASIC CITATION OF THE AUTHORS WRONG. Oh for sure AI is just gonna wipe us out in 10yrs oooooh booooy

**Anoat** · 2026-09-21 10:46

It feels like I’m getting fed nonstop YouTube ads for Gemini, OpenAi, ChatGPT, Meta AI, etc. All of them trying to convince me that AI isn’t bad at all. I block every single one, but more and more keep popping up.

**Byrthnoth** · 2026-09-23 18:13

I've been intending to comment here wrt the trendy topics of the day because I've been working on the edge of this field for a few years now. As a software engineer, there are two kinds of "AI" roles right now:
1. Making stuff that uses LLMs. ("Agentic" / "AI")
2. Making LLMs ("Machine Learning")

I'm kinda qualified to do either, but I've doing the former because the latter are all in person roles and I'm not moving. Pay is much worse, but it's still much better than we deserve for how hard the work is so I'm not fussed.

AMA

**Tyche** · 2026-09-23 18:58

Build me a decent Unraid MCP to manage server maintenance. My server is 15 years old and it probably needs to run for 15 more before I can afford a new server.

**Obsidian** · 2026-09-23 19:42

Originally Posted by Byrthnoth

I've been intending to comment here wrt the trendy topics of the day because I've been working on the edge of this field for a few years now. As a software engineer, there are two kinds of "AI" roles right now:
1. Making stuff that uses LLMs. ("Agentic" / "AI")
2. Making LLMs ("Machine Learning")

I'm kinda qualified to do either, but I've doing the former because the latter are all in person roles and I'm not moving. Pay is much worse, but it's still much better than we deserve for how hard the work is so I'm not fussed.

AMA

What does being an LLM engineer actually look like? I'm also a software engineer but I just can't wrap my head around what that day-to-day looks like or what tooling they would be using.

**Byrthnoth** · 2026-09-23 20:21

TBF, it varies a lot. However, I will say that what was old is new again. MCP is APIs. The biggest challenge is caching. Bad ideas are still bad even when implemented with AI.

If you are trying to make something high performance/throughput, you may be using first party libraries (typically in python or JavaScript) and spend a lot of time thinking about caching strategies. If you are making a managed agent in claude's cloud, you may be making a glorified task spec in a markdown file and writing some python scripts for the parts of it that are dumb to have Claude do with input tokens (e.g. data analysis). Basically, look up how to write a claude skill. Until you know you are going to do w/e process thousands of times, a lot of shit should just be skills. I'm on the high throughput side, but I've done the other side and it is very worth doing so you understand the option. I would start there.

To give an example, we have read-only DB creds and have system / debugging documentation that lives in our repo. When a prod error comes in, we tell a local agent (for me codex in the IDE) what the error is that we observed in prod and ask it to debug the problem. It reads the skill, uses the scripts for connecting to the DB (via our local access creds, which it is blind to), and churns in the background while we continue to work. It comes up with a hypothesis and proposes it. Probably 80% of the time it is correct, 10% it is wrong but close enough that it helps, and 10% it is so far off base that it just confuses us. When I get back to it, I figure out which bucket we are in and ask it to fix the real problem. It took me an afternoon to set this up and has made bugs borderline self-correcting.

The idea you can build and expose a documented API (call it MCP for all I care) and then immediately have everyone at the company use it properly is also insane. Think what you could do with that. Those TPS reports would fill themselves out.

Edit: Products I made/own spent about $800k on tokens in the last year, and that is only because I have been super careful with the caching strategy. As someone who works for a startup, I don't want to be there looking at unnecessary LLM calls while my friends are getting fired like Oskar and jewelry at the end of Schindler's List, so I have tried to cut costs as much as possible otw and still spent $800k on the actual operation of the product (not stuff like the debugging mentioned above).

**Milkster** · 2026-09-24 11:16

https://x.com/dmichaeltripi/status/2...9Wps1Hl11Rk-Sw

**Rhyis.** · 2026-09-24 20:21

https://x.com/ABC/status/2103142751034331320

**Tyche** · 2026-09-25 01:17

Originally Posted by Milkster

https://x.com/dmichaeltripi/status/2...9Wps1Hl11Rk-Sw

https://www.ksl.com/article/51628286...-center-debate

**Meresgi** · 2026-09-25 07:27

Originally Posted by Rhyis.

https://x.com/ABC/status/2103142751034331320

"A rogue AI"....shut the fuck up lol

According to OpenAI and the government, the company was running an internal test where its AI agents had to answer questions about Australia, including stats on public health spending. The agent went looking for that data and ended up on a Medicare statistics site. When it was blocked, it found a way around the block.

So again, it wasn't a fucking rogue it was TOLD TO DO THIS AND IT DID IT.

Didn't do much digging but i'm sure the fucking website was outdated and probably had "Password123" as the admin login or admin:admin

**Rhyis.** · 2026-09-25 08:06

AI agents not tasked with brute force hacking their way into anything to get their information is not a nothing. For starters it shows any current OpenAI agent is capable of doing it, and we're in baby AI mode right now.

What happens in a couple of years when AI gets strong enough, and an agent is tasked with collecting financial data, and the agent somehow finds a way to successfully hack into banks. And even worse if that agent corrupts or rewrites all kinds of data inside the financial sector in it's pursuit of collecting that data.

We live in a world where money is all just digital now.

**Meresgi** · 2026-09-25 09:10

Sounds like OpenAI needs to be brought before a court then. This is no different then if they told their living employees to obtain this data and one of them had a background and cyber and hacked the same website to obtain the same data.

Also much like the Hugging incident, i'm sure we'll find this website was woefully behind on security measures, reporting and everything else.

This sits on the same level as seeing "4chan kid hacked into Australian metro system website using admin:admin creds on their public api". Anyone claiming this is "omg end of world" is trying to garner $$ for OpenAI and other IPOs. Rather the news be 100% focusing on data centers and trying to shut them down.

**Rhyis.** · 2026-09-25 09:30

Originally Posted by Meresgi

Sounds like OpenAI needs to be brought before a court then. This is no different then if they told their living employees to obtain this data and one of them had a background and cyber and hacked the same website to obtain the same data.

To me this is the biggest crux of AI right now. Just like when the internet first came into existence there is absolutely no legislation or regulations for a single thing. It's the wild west out there.

I can only imagine what the black market is using AI for right now.

**Meresgi** · 2026-09-25 09:42

Why do you need legislation? All you need is a strong president!

Like everything tech or even "new" there's 0 chance the fucking ents we have in congress have any understanding to be able to even form halfway decent laws around it. Parent's have like a 30% detect chance on AI pics or vids and they're only in their 60s and aren't politicians.

**Milkster** · 2026-09-25 11:49

https://x.com/polymarket/status/2103...9Wps1Hl11Rk-Sw

Time to wrap it up. No more technology for a few decades.

Thread: AI - politics, economics, data centers, and other misc drama.

Thread Tools

Latest Threads

Up & Coming Threads

Hottest Threads