Lumes

Lumos maxima for the dark corners of AI

[Lumes @Lumesssss]

@Flikore 100% broski

[Lumes @Lumesssss]

This is f*cking scary.

Jared Kaplan, co-founder and Chief Scientist at Anthropic (the company behind Claude), says humanity is running out of time on one decision. Somewhere between 2027 and 2030, we’ll have to choose whether to let AI train its own successor.

In an interview with The Guardian he called it “an extremely high-stakes decision.”

Today humans run the loop. We set the goals, train the model and check the results. AI already helps with parts of that work.

In Kaplan’s scenario, an AI about as capable as a human builds a stronger version of itself. That version builds the next one, and each step comes faster than the last. A model that took a year to build gets replaced in months, then in weeks.

Humans are slow. A safety review that takes us a month is useless if the AI ships five new versions during that month.

Kaplan named two risks if this runs without control: losing control of the process, and not knowing what the AIs are doing. In his words, “Do you even know what the AIs are doing?”

What worries me most is the speed. Once the loop runs fast enough, there may be no moment where a person can step in and hit stop. The systems would keep changing themselves in ways nobody fully understands, and every new version would be built on top of the last one.

Kaplan doesn’t call it inevitable. He says self-improving AI could also do a lot of good. But someone has to make that call, and soon.

If you had the switch, would you let AI train the next AI?

[CrazyShyyt @CrazyShyyt]

ChatGPT has now a big problem.

Researchers at Oxford exposed a massive threat to large language models.”

It’s called "Model Collapse."

And it explains why the internet is about to become a toxic waste dump for Large Language Models.

Right now, AI companies are scraping the entire web to train their next-generation models.

But there’s a massive problem.

The internet is no longer just human data. It is rapidly filling up with AI-generated text, images, and code.

What happens when an AI trains on data generated by another AI?

A new paper in Nature just gave us the answer. And it is terrifying.

When AI models are recursively trained on AI-generated data, they suffer from irreversible brain damage.

The researchers found that the model starts forgetting the "tails" of the original data distribution. The rare, weird, unique, and complex human nuances completely disappear.

It's like making a photocopy of a photocopy.

Every generation loses resolution. The output gets flatter, blander, and more generic, until the model completely collapses in on itself.

For years, we thought the biggest bottleneck for AI was compute, getting enough H100 GPUs from Nvidia.

This paper proves the real bottleneck is us.

If we keep flooding the internet with synthetic AI slop, we will poison the very data pool required to build smarter models.

The ultimate irony? We thought AI was going to make human creators obsolete.

Instead, this research proves that verified, genuine human data is about to become the most valuable commodity in the tech industry.

[Lumes @Lumesssss]

@0x_cognit true, broski
that's also the easiest way to get a team to adopt it

[Lumes @Lumesssss]

This is f*cking insane. Jared Friedman, partner at Y Combinator, the accelerator behind Airbnb, Stripe and Dropbox:

"Startup founders, especially young ones, are not fully appreciating just how big vertical AI agents are going to be"

"Here's my pitch for 300 vertical AI agent unicorns"

In plain English: pick one job, build an agent that does it better than anyone else, and own the company that comes out of it

His logic: "Literally every company that is a SaaS unicorn you could imagine, there's a vertical AI unicorn equivalent"

Prompts → Claude Code → Memory → Skills → MCP → One Agent For One Job → A Company You Own

Save this before it disappears from your feed

Everything below is free:

Claude Code in action
https://t.co/t3HYySzSVL

Prompt engineering
https://t.co/MC5hZbcbSI

Claude Code 101
https://t.co/SOrNK3oHR2

Claude.md + memory
https://t.co/541FrSTI75

Skills
https://t.co/yEKhflTiU1

MCP
https://t.co/mlcvJbxKfV

Claude Code ultimate guide
https://t.co/CL7wDpZK2q

Official Claude Code docs
https://t.co/lV1TgA5O2u

Cost? $0

YC is already funding the founders who build these agents. Everyone else is still watching podcasts about them

Bookmark it and start with the first link tonight

Next, learn how to build an agent that holds up in the real world. The harness engineering guide is below ↓

[rari🈀 @0xwhrrari]

https://t.co/3ANIWs48lJ

[Lumes @Lumesssss]

@EvgeniyIolld Yeah, I agree

[Lumes @Lumesssss]

@zuvexxx curious about cost

[Lumes @Lumesssss]

@0xVeyyro saving this for later

[Lumes @Lumesssss]

@0x_stat1x curious how scoring works

[Lumes @Lumesssss]

WTF, HOW ARE PEOPLE STILL UNEMPLOYED IN TECH WHEN THIS EXISTS

3.1M people saw this. 8.8K saved it.
maybe 1% will actually start.

that’s your edge.

the whole 6 month plan in 6 lines:

month 1 → electronics, soldering, python + git
month 2 → arduino, esp32, motors, sensors. first moving robot
month 3 → CAD + 3D printing. build your own robot arm
month 4 → ROS 2 + simulation. how real companies build robots
month 5 → PID, kinematics, computer vision
month 6 → LeRobot, VLA models, portfolio, interviews

no degree
no $3,000 bootcamp
no “I don’t know where to start”

everyone is fighting over AI engineer roles.
robotics is the least crowded high value skill in tech right now.

bookmark this. start monday. thank me in 6 months

[Ronin @DeRonin_]

https://t.co/b4fTEzmv6K

[Lumes @Lumesssss]

@0xCortiq The worst part is usually finding every PHP box, not patching it. Once you have the list, it goes quick.

[Lumes @Lumesssss]

NOBODY IS DOING THIS MATH.
one sysadmin costs $95,000 a year. Claude costs $1,200 a year. same 6 AM question

"how many PHP servers do we have?"

this company needed 11 hours and a whole IT team to answer it

Claude Code can answer it on your own network while you make coffee. it scans every host, finds every server running PHP and its version, checks it against your inventory, shows what each forgotten box talks to before you touch it, and gives you a patch plan with the scariest ones first

the patch took 4 hours. the searching took 11. that's the part you're paying $95k for and it's the part AI is already better at

full prompt in the reply

[Derek Devicemanager @IT_unhinged]

Security called me at 6:12 AM today.

"Patch every PHP server immediately."

I asked what happened.

"Two critical vulnerabilities. Remote code execution. No authentication required."

Cool.

How many PHP servers do we have?

According to inventory: 12

According to reality:

Nobody knows.

We found one running WordPress.

Three running old internal tools.

One under someone's desk.

Two inside vendor appliances.

A forgotten employee directory from 2016.

And a server named:

PHP-OLD-DO-NOT-TOUCH

Naturally, that one runs payroll exports.

Then we found a PHP 5.6 box.

I asked who owned it.

The documentation said:

BRIAN.

Brian left in 2019.

We called Brian.

He said:

"Oh wow, that's still running?"

Excellent.

By 8 AM, leadership joined the incident call.

CEO:

"Can hackers actually get in?"

Security:

"Potentially yes."

CEO:

"Without a password?"

Security:

"Yes."

CEO:

"Then why do WE need passwords?"

Not the lesson.

Finance asked if we could wait until the weekend to patch because downtime affects revenue.

Security explained that the alternative was letting strangers execute code on our servers.

Finance asked for a cost comparison.

At 9:17, someone suggested temporarily taking the vulnerable websites offline.

Marketing refused.

"Our campaign is live."

So for the next hour, the company officially decided maintaining a landing page for 20% off webinar registration was more important than preventing remote server takeover.

We finally patched everything we could find.

Then our scanner discovered another PHP server.

Hostname:

TEMP-PHP-2014

Uptime:

4,512 days.

Nobody knows what it does.

Nobody owns it.

Nobody can shut it down.

But every time we block network access, the warehouse stops printing shipping labels.

Security has classified it as critical infrastructure.

It is running PHP 5.4.

The vulnerability has existed longer than half our employees.

The emergency patch took four hours.

Finding out what PHP actually runs took eleven.

The internet isn't held together with code.

It's held together with servers everyone is afraid to turn off.

[Lumes @Lumesssss]

@Frelzu The 10 hours before it went public is the part I can’t stop thinking about.

[Lumes @Lumesssss]

here's the prompt. run it in Claude Code from a machine inside your network, and only on a network you own or are allowed to scan

I'm the admin of this network and I'm allowed to scan it. We just got hit with critical PHP bugs that let anyone run code on a server without logging in. Before we patch anything I need to know exactly what we run.

1. Scan [subnet] for every live machine and every open web port. Install nmap if it's not there.
2. For each web server, figure out the system, the web server and the PHP version from what it reports. Where I give you SSH access, check the PHP version directly.
3. Compare everything with [inventory.csv] and list every machine that's missing from it or has a different PHP version.
4. Look through DNS names, hostnames and our docs in [folder] to find who owns each machine. If nobody does, mark it UNKNOWN.
5. For every unknown or old PHP machine, show me which other machines talked to it in the last 24 hours using the firewall logs in [folder], so I know what breaks if I turn it off.
6. Sort everything by how old the PHP is, whether it's open to the internet and how much depends on it. Worst first.
7. For each machine write the upgrade steps, how to roll it back and the quietest time to do it.
8. Save the list as php_inventory.csv and the plan as php_patch_plan.md.
9. Give me a script I can run every week that tells me when a new PHP server shows up.

Only look, never change anything. Ask me before you run anything that touches a server.

[Lumes @Lumesssss]

@Plutkar the real estate cycle part is what stuck with me. 5 to 8 good years then oversupply means whoever builds last gets wrecked, not whoever builds first

[Lumes @Lumesssss]

@kivoran claude code sorted the log and skipped anything it wasn't sure about. i just blocked in small batches and checked party chat every time, because yeah, it dies with zero error

[Lumes @Lumesssss]

Your PS5 is spying on you and selling to you. Mine isn't anymore. One night with Claude Code

No jailbreak, no firmware. It wakes up from my phone now, the news feed is empty, and 0 promo and tracking domains are blocked for the whole house

It streams to my laptop when someone takes the TV, and Home Assistant puts it to sleep at 1am

Every change sits in an undo log, one line to reverse

It was super easy. Claude Code plus the right order of steps

The full prompt is in the reply

[goldtooth.exe @0xGoldTooth]

Hacked my TV and my PS5 with Claude Code in one night

The TV finally stopped calling home. Developer mode on, 37 tracking domains blocked, ads and bloatware gone, my own launcher instead of the laggy stock one

The PS5 boots straight into my games now, 9 seconds instead of 41. 14 preinstalled apps hidden, 190 GB freed for GTA 6, and the store stays out of my face

Also made a screensaver for both screens with my calendar, the weather, a ping for every new Stripe payment and a GTA 6 countdown

It was super easy. Claude Code plus the right order of steps

Follow and save this, in a few days I'll post exactly what I asked Claude Code

[Lumes @Lumesssss]

NOBODY SAW THIS. it's the difference between using AI at your job and being the person your boss asks to set it up.

a loop is when you give AI a task and let it redo it until it looks right. most people stop there.

a graph is when you decide who does what. one agent writes, another one checks it, and nobody grades their own work.

loops save you an hour a day. graphs are how one person ends up doing a whole team's job, and people notice that at review time.

there's no course for this yet. the people building AI are still figuring it out in public, under posts like this one.

[Anatoli Kopadze @AnatoliKopadze]

https://t.co/fyBGWQbsvn

[Lumes @Lumesssss]

@kivoran dns level. AdGuard Home on a Pi, router points every device to it. Claude Code sorted the query log, i blocked in batches and tested PSN after each one

[Lumes @Lumesssss]

How to Make Your PS5 Less Annoying With Claude Code (Full Guide)
You paid $500 for a PS5 and it still opens on a store.

There's no switch to turn that off. And unlike an Android TV, the PS5 has no developer mode, so nobody can "clean" the console itself. No jailbreak, no modded firmware, no risk to your account.

What you can do is build everything around it. One evening with Claude Code gets you:

a PS5 that wakes up from your phone
your games streaming to your laptop when someone takes the TV
Home Assistant showing the console status and putting it to sleep at 1am
promo and tracking domains blocked for the whole house
a much quieter news feed on the home screen
an undo log where every change is one line to reverse
Menu names can differ a little between firmware versions. If something doesn't match, describe what you see to Claude Code.

Step 0. What you need
a PS5 on the latest system update
a laptop (Windows, Mac or Linux) on the same Wi-Fi
your PSN account
ideally an always-on device: a Raspberry Pi, an old laptop or a mini PC. It runs the DNS blocker and Home Assistant.
One warning before you start: if the DNS blocker runs on a laptop and the laptop goes to sleep, your PS5 loses internet. Put it on an always-on device, or switch the PS5 back to automatic DNS when you're done testing.

Step 1. Install Claude Code
Install Node.js (the LTS version from https://t.co/Sjb1Wi6wrK), then in your terminal:

npm install -g @anthropic-ai/claude-codemkdir ps5-cleanupcd ps5-cleanupclaude
First prompt, paste it as is:
I want to make my PS5 less annoying without touching the console itself. No jailbreak, no firmware, nothing irreversible.Plan: 1) Remote Play with chiaki-ng, 2) Home Assistant + ps5-mqtt, 3) AdGuard Home DNS blocking, 4) PS5 settings cleanup.Keep an undo log in ps5-undo.md: every change, why, and the exact step to reverse it.Go one step at a time and wait for me to confirm each step https://t.co/vU1MebZ3rP setup: [Windows/Mac/Linux laptop], always-on device: [Raspberry Pi / none].
Step 2. Prepare the PS5 (5 minutes)
On the console:

Settings → System → Remote Play → turn on Enable Remote Play
Settings → System → Power Saving → Features Available in Rest Mode:Stay Connected to the Internet: on
Enable Turning On PS5 from Network: on

Settings → Network → View Connection Status: write down the console's IP address
Step 3. Remote Play and wake from anywhere (chiaki-ng)
chiaki-ng is a free, open-source Remote Play client.
Download it from the official releases page of the streetpea/chiaki-ng project on GitHub. Claude Code will tell you which file fits your system.
Open chiaki-ng and sign in to PSN inside the app. It fetches your account ID for you.

On the PS5: Settings → System → Remote Play → Link Device. You get an 8-digit PIN.
Enter the PIN in chiaki-ng. The console is registered.
Test it: put the PS5 in rest mode, wake it from chiaki-ng, start a stream.
Tell Claude Code: chiaki-ng registered, wake and stream work. log it.
Step 4. Home Assistant + ps5-mqtt
No Home Assistant yet? The easiest way is Home Assistant OS on a Raspberry Pi. Claude Code walks you through it.
Help me add my PS5 to Home Assistant with ps5-mqtt and a Mosquitto MQTT broker.PS5 IP: [IP]. I want: a power switch (on / rest mode), the console status, and an automation that puts the PS5 into rest mode at 1:00 if it's on.Add every change to ps5-undo.md.

ps5-mqtt asks you to sign in to PSN once. After that, a PS5 switch shows up in the Home Assistant app on your phone.
Step 5. DNS blocking (AdGuard Home)
This is the part that quiets the home screen.
Install it:
Install AdGuard Home on [always-on device / this laptop] and give me its IP.Don't add any blocklists yet.
On the PS5: Settings → Network → Settings → Set Up Internet Connection → your network → Advanced Settings → DNS Settings: Manual. Primary DNS: the AdGuard Home IP.

Use the PS5 normally for 10 minutes: home screen, store, a couple of games, party chat.
Let Claude Code read what the console asked for:

Read the AdGuard Home query log for the PS5 ([IP]) from the last 15 https://t.co/A62v5KX8mG the domains into: news/promo, analytics, ads/recommendations, and REQUIRED (PSN sign-in, system updates, game downloads, party chat, store purchases).Never propose blocking anything in REQUIRED. If unsure, mark it REQUIRED.Then block the candidates in batches of 5. After each batch, stop and wait for me to test.
After every batch, check on the PS5: PSN sign-in, a game download, party chat, the store. If something breaks: batch N broke [what], undo it.
The 3 rules:
block in small batches, test after each one

never block update servers. a PS5 that can't update can't play online
keep the undo log. every block is one line to remove

Expect a quieter home screen, not a perfect one. Some images come from the same servers as the content you need.
Step 6. The settings Sony buries (10 minutes)
Settings → Saved Data and Game/App Settings → Followed Games: unfollow the games you don't want news from. This is what empties the news feed.
Settings → Notifications: turn off what you don't need.
Settings → Users and Accounts → Privacy: check what's visible to others and what you share with Sony.
Tell Claude Code what you changed so it goes into ps5-undo.md.

[Lumes @Lumesssss]

@AnatoliKopadze Really interesting and unusual info