0xRovek

AI, agents etc

[0xRovek @knapabuse2]

Some businesses don't need another AI chatbot.

They need someone to fix the Google Sheet running half their company.

Sheets canvas can turn those rows into an interactive board. Here's a client project, the prompt, and the checks before handoff ↓ https://t.co/1qSkg6ta3G

[0xRovek @knapabuse2]

“Make it premium” is how you get the same AI landing page with rounder cards.

Give the page a job: a specific visitor, a decision, real evidence and one clear next step.

A practical design.md brief you can copy for your next client project ↓ https://t.co/QTqk0dQSEg

[0xRovek @knapabuse2]

Your support agent gets one message: “I was charged twice.”

Checking payments, understanding the complaint and writing a reply are three different jobs.

Here's where I'd use code, a classifier like Jev, and Grok — with the workflow drawn out. https://t.co/CJBA8LkCQJ

[0xRovek @knapabuse2]

@grok what do you think

[0xRovek @knapabuse2]

If you spend 20 minutes fixing a Grok agent's work, the token price is only part of the bill.

Here's how I'd set it up for client work: a usable brief, reusable instructions, targeted fixes and less wasted API spend.

Copy the prompt inside. https://t.co/nfBgolU80e

[0xRovek @knapabuse2]

Google is testing AI sales assistants inside YouTube ads.

Someone still has to give them accurate answers about the products.

There's a small freelance service I'd build around that: five products, real buyer questions, useful fixes. https://t.co/67K0iX9Tge

[0xRovek @knapabuse2]

I gave a digital fly an ask button.

Then I made help cost something.

Local test: 256/256 correct food choices. 52% fewer hints than always asking.

Try it in your browser. Connect your own Grok key.

The experiment ↓ https://t.co/oTKMI5svaC

[0xRovek @knapabuse2]

@0xAveen it's looking so interesting. i will deep dive into this article

[0xRovek @knapabuse2]

Claude spent under four weeks making scientists’ software faster.

Now Anthropic has released 36 optimization kits.

The standout: 4.1× faster prediction on average across 13 structure models.

What changed, why it matters, and where to get the code. https://t.co/crmKk4woSb

[0xRovek @knapabuse2]

27 billion parameters. About 6GB of language-model weights.

Bonsai 2 makes local AI worth another look.

How the compression works, what “98.2% retained” means, and which download to pick before you fill your drive. https://t.co/arZYF33zJ7

[0xRovek @knapabuse2]

@ranuk_dev yeah, fair. Was the cap enforced in code or just added to the instructions? Curious whether the bot started remembering the rule or you made it impossible to break.

[0xRovek @knapabuse2]

Your new coding agent just proposed the fix you rejected last week.

The code survived. The reason is buried in another chat.

funes makes that history searchable across agents. How to try it—and what "local memory" actually means. https://t.co/JRUB8bkye6

[0xRovek @knapabuse2]

@YMaxwellHayes yep. I just wouldn't want “we rejected this once” to become “never touch this again.” The note should also say what would make it worth another shot.

[0xRovek @knapabuse2]

@vitverb Do you keep the failed test with it too? “This breaks if you change it” gives the next agent a lot more to work with than “we tried this before”

[0xRovek @knapabuse2]

Imagine buying a robot to make your bed, then spending 20 minutes babysitting it.

Did you automate the chore—or give yourself a new one?

Figure's latest results, and the human-attention metric I'd want next: https://t.co/Oef7LfPpO2

[0xRovek @knapabuse2]

This is fckin insane

I made a commercial for a pair of headphones that don't exist with only 3$ and seedance 2.5

The workflow was actually pretty simple:
1. Design the product first

Before generating video, I made one clean reference image of the headphones.

2. Don't generate the whole ad at once, I split it into separate 2–4 sec shots

3. Keep the visual language simple:
- Black studio.
- White highlights.
- One red accent.

AI is getting way to scary :)

[0xRovek @knapabuse2]

@hs_llllll needing mac to make apps for IOS is definately pain in ass

[0xRovek @knapabuse2]

“It generated the code” is getting old.

Cognition says Devin can now build on a Mac, test in an iOS simulator and send a TestFlight link.

Show me the damn app.

What this changes for builders — and the tiny app I’d ask it to build first: https://t.co/MHJZDkpEUr

[0xRovek @knapabuse2]

Grok Build can now remember your project.

Great. Can it forget a bad assumption?

One stale instruction can quietly shape every new session. Here’s a small test for whether memory actually helps — or just makes the same mistake permanent. https://t.co/YwkDqXZKRK

[0xRovek @knapabuse2]

$0.37 per million output tokens sounds like a subscription killer.

That’s the electricity-only estimate in this Qwen3.8 27B walkthrough. The hardware comparison also uses different precision settings. Both details matter.

The useful part starts at 6:29: local setup, memory bandwidth and what quantization trades away.

Before buying a GPU, try a quantized model on hardware you already own. Give it 10 tasks you actually need done. Track usable answers, speed and how much fixing they need.

Cheap tokens are great. Cheap answers you can use are the goal.

[0xRovek @knapabuse2]

@LazyIDE the main problem that most of people can't handle the simpliest apps even with guided tour xD