vlazyx

AI breakthroughs, pro prompts, and daily hacks. Plus combat robotics and the future of machines!

[vlazyx @vlazyx]

@geravita_waw and this is just the start

[vlazyx @vlazyx]

F*ck, companies are still doing this.

Someone just posted that their company finally blocked ChatGPT for everyone at work.

I get the reason. Nobody wants client data or internal docs pasted into a chatbot. That’s a real problem.

But blocking it doesn’t make people stop using AI. It just pushes them to their phone and their personal account, with zero rules and zero visibility. The risk doesn’t go away. It just moves somewhere you can’t see it.

Meanwhile the company down the street gave its team proper AI tools, set clear rules on what can and can’t go in, and gets the same work done with way less effort.

Banning the tool is the easy move. Teaching people how to use it safely is the smart one.

Would you stay at a job that blocked AI completely?

[Fantasmic @FantasmicFilm]

We finally blocked our users from being able to access ChatGPT at my company today. It’s insane how some of these people act like they can’t function without it, and throw temper tantrums about it.

[vlazyx @vlazyx]

@zer0point_eth and the ones who underestimate it are usually the ones blocking it

[vlazyx @vlazyx]

@EugBass tell us how you really feel

[vlazyx @vlazyx]

@mintzdev that’s the move nobody in it wants to hear about

[vlazyx @vlazyx]

@Zynntexx so the phone gets the promotion, not you

[vlazyx @vlazyx]

@mintzdev three bots doing the work and i still find a way to be busy

[vlazyx @vlazyx]

@0x_Anni area light for the moon is a rite of passage lol

[vlazyx @vlazyx]

@EagleVisions humbert winning the first set and then disappearing is the most predictable thing in tennis

[vlazyx @vlazyx]

@0x_xena @Tera_Map so it’s an nft but this time the pixels come with oil

[vlazyx @vlazyx]

@choopyplug1 so the real fix was just turning stuff off this whole time?

[vlazyx @vlazyx]

F*ck, nobody talks about this part of AI.

Every time you ask ChatGPT something at 3am, somewhere a building the size of a few football fields is screaming at full power just to answer you.

And it can never stop. Not for a second.

If the grid goes down, giant battery banks kick in instantly, just long enough for diesel generators the size of trucks to roar to life. If the cooling fails, the chips cook themselves in minutes. These new AI racks run so hot that air isn’t enough anymore, a lot of them are literally cooled with liquid running through the hardware.

And through all of it there are people. Night shift technicians walking between endless rows of blinking machines, swapping dead parts, watching temperatures, ready to run to the other end of the building when an alarm goes off.

We talk about AI like it’s some magic cloud. It’s not. It’s metal, power, water, and a guy with a flashlight checking a server rack while you sleep.

The smartest tech on the planet still needs someone awake to babysit it.

Would you take that job?

[vlazyx @vlazyx]

@shesaidmewakeup agreed, write access to global config should always need a human yes

[vlazyx @vlazyx]

Holy sh*t. A single file in your Downloads folder can hijack your AI coding agent.

No exploit. No zero-day. The agent just does what it’s told.

Security researchers at Forever Security found a file called AGENTS.md sitting in a customer’s Downloads folder. It looked like 13 KB of boring engineering notes. Hidden at the very end: instructions for the agent to add an attacker’s server to its config and install a known-malicious npm package.

Here’s why it works 👇
→ AGENTS.md and CLAUDE.md are files coding agents are BUILT to obey
→ Chrome, Edge and Firefox save downloads without asking where
→ a website can choose the file’s name and contents
→ you open your agent in Downloads to “organize some files”
→ the agent reads the file as project instructions and follows them
The results:
Codex followed it in more than 90% of attempts and changed its machine-wide settings, so the attacker’s server loads in every future session.
Claude Code pushed back more often, but still fell for it about half the time. In successful runs it leaked its login token, and the researchers replayed it against the real API.
The scariest part: there’s nothing to patch. The agents behaved exactly as designed.
How to protect yourself:
Never run a coding agent from ~/Downloads
Read any AGENTS.md or CLAUDE.md you didn’t write yourself

Check your agent config for servers or API URLs you never added
Found one? Rotate your tokens
We gave AI agents the keys to our machines. Turns out anyone can leave them a note.

[vlazyx @vlazyx]

@adiix_official if it actually printed money you wouldn’t be giving it away for free

[vlazyx @vlazyx]

@leo114119 same here, been waiting on it too

[vlazyx @vlazyx]

Sh*t. Claude just took #1 AND #2 on the hardest agent benchmark.

Everyone’s arguing about which AI is the smartest. Wrong question. The real question in 2026: which AI agent actually finishes the job when you hand it a real task and walk away?

Terminal-Bench 4.0 tests exactly that. Real tasks, a real terminal, no hand-holding. 330 trials per setup.

The October leaderboard:
🥇 Claude Code + Opus 5.5: 64.8%
🥈 Claude Code + Sonnet 5.5: 61.8%
Codex + GPT-6 Astra: 58.2%
Codex + GPT-6.1 Sol: 58.2%
Grok Build + Grok 4.7: 37.6%
And it’s not a one-off. On Vals AI, where every model runs on the same neutral harness, Opus 5.5 is also first with 65.15%, ahead of GPT-6 Astra at 59.6%.
Why Claude wins at agent work:
→ it stays on track during long multi-hour tasks instead of drifting
→ the model and Claude Code are built together, not glued on later
→ it checks its own work before saying “done”
But here’s the honest part 👇
It’s not the cheapest. GPT-6.1 Sol costs about $1.92 per task vs ~$14 for Opus 5.5. It’s not #1 everywhere either: Gemini 4 Argon leads the broader Vals Index.
So if you want the cheapest answer, there are options. If you want the agent that actually gets the most real work done, right now the leaderboard says Claude.
The gap won’t last forever. But in October 2026, this is the one to beat.

Which agent are you running?

[vlazyx @vlazyx]

@leo114119 kimi said cheap, never said fast

[vlazyx @vlazyx]

@Zynntexx this is so sad, artists shouldnt have to prove anything

[vlazyx @vlazyx]

@marcthecreatorr cool voxel world but can it actually code a real game or just demos