HateJPG

Real AI use cases, autonomous systems, and clean architecture

[HateJPG @chopeaceus]

@0xVirtanen such a brutal reality check but honestly he is so right

[HateJPG @chopeaceus]

@koplenkoo scrimming against ai clones of top teams sounds absolutely wild

[HateJPG @chopeaceus]

@Abobsterina treating your vault like a codebase is a genius way to organize

[HateJPG @chopeaceus]

7 RTX 5090S IN DISTILLED WATER TURN CLOUD BILLS INTO AI BOX YOU OWN

the stake is simple:

one machine for Qwen, DeepSeek, Llama,
image jobs, embeddings, RAG search,
and private client code
without renting every token from a cloud API.

not free AI.
not a cute desktop chatbot.
not one Mac mini pretending to be a lab.

Seven RTX 5090 cards make it less like a PC
and more like a tiny AI utility plant.

you can serve local models.
batch inference overnight.
fine-tune small adapters.
run evals.
keep a coding agent on localhost.
still have room for experiments.

but the romance ends at the plumbing.

power draw, motherboard lanes, risers,
pump reliability, condensation,
mineral contamination, model sharding,
and heat become the product.

the distilled water tank looks insane on video.

the boring win is the real one:
stable thermals, enough VRAM,
and a setup that does not throttle
when the workload becomes real.

cloud is easier.

this box is for the operator
who wants the bill, the data,
and the bottleneck under their own roof.

[kocer @kocer_eth]

https://t.co/S7QSXVUTwg

[HateJPG @chopeaceus]

@Aspect0F privacy setups are so satisfying to see working smoothly

[HateJPG @chopeaceus]

THIS JETSON ORIN NANO RUNS A AI PRIVATE VOICE ASSISTANT WITHOUT CLOUD

Tiny board on the desk.
Lilliput screen full of terminal logs.
Mic and speaker sitting next to it.

The loop is simple:

speech goes in
local LLM answers
text-to-speech comes back out

No API call has to leave the box
for the basic assistant flow.

That is the Local AI PC stake here.

Not "this replaces GPT-5.5."
Not "a small edge board beats an H100."

It is the boring private loop
owned by the person using it.

A workshop assistant.
A home lab bot.
A field device.
A kiosk.
A robot brain that still answers
when Wi-Fi dies
or the data should not touch a cloud dashboard.

The Baymax question on the screen is tiny.
The mechanism is not.

Microphone.
Local inference.
Speaker.
All beside the Jetson Orin Nano,
instead of behind a subscription page.

The limit is still real.

Model size, RAM, thermals,
quantization, and latency decide
what feels usable on this board.

But for narrow voice agents
and private always-on assistants,
this is where local AI stops looking
like a hobby demo
and starts looking like owned infrastructure.

[kocer @kocer_eth]

https://t.co/S7QSXVUTwg

[HateJPG @chopeaceus]

@papa_couch what are you building with it

[HateJPG @chopeaceus]

@kocer_eth absolutely love local voice hardware no lag and totally private

[HateJPG @chopeaceus]

@beamnxw building the workflow is the real cheat code for these faceless accounts

[HateJPG @chopeaceus]

@merccante the leverage you get as a solo builder right now is actually crazy

[HateJPG @chopeaceus]

@kadsxr building a consistent digital actor makes way more sense than just posting random pictures

[HateJPG @chopeaceus]

@Rossst_03 crazy how one paper from the fifties still completely dominates how the biggest funds move capital

[HateJPG @chopeaceus]

@N01ennn from a massive rack to a paperback size in nine years is crazy tech advancement

[HateJPG @chopeaceus]

@gippp69 seeing mcp bridge llms directly to hardware is seriously cool tech

[HateJPG @chopeaceus]

8.4M VIEWS FROM TINY AI FOOD BABIES, BUILT BY TURNING CLAUDE INTO A HIGGSFIELD VIDEO OPERATOR

not the niche.
the loop.

video shows a YouTube Kids format called “Cute AI Food Babies” pulling millions of views: baby fruits and vegetables eating other food in ASMR-style clips.

then it shows the actual production path:

1. open Higgsfield ai
2. go to MCP
3. copy the Higgsfield MCP connector link
4. open Claude
5. Customize → Custom connectors
6. paste the connector

now Claude is not just writing prompts.

Claude can take the reference image, understand the visual pattern, and call Higgsfield to generate the video.

the useful move is simple:

take a screenshot of a video style already working.

look at the subject, framing, food texture, colors, camera angle, and pacing.

then ask Claude for the same format with a different concept.

example from the video:

“create me a viral video in this exact same style but with a different concept”

then Claude generates a broccoli baby dipping into cheese through Higgsfield.

that is the payoff:

you are not staring at a blank video prompt anymore.
you are giving Claude a visual brief, a target format, and a video tool it can actually use.

but this is not a money button.

millions of views do not mean easy monetization, and copying the style too closely just creates disposable slop.

the better play:

10 characters
10 foods
10 hooks

ship variations fast, measure retention, then decide if the format is worth building around.

[Skaly_Bull @Skaly__Bull]

https://t.co/tl1wxsCOFM

[HateJPG @chopeaceus]

@raidenfomo classic documentary the tension on that floor without any computers is unreal

[HateJPG @chopeaceus]

@Neuron_404 pretty crazy that anyone can make a movie on their phone now

[HateJPG @chopeaceus]

OBSIDIAN'S BEST FEATURE IS THAT YOUR SECOND BRAIN SURVIVES THE APP

Not the graph view.
Not the plugin screen.
Not the clean mobile demo.

The real mechanism is much more boring:

Every note is a local .md file.

That changes the whole power dynamic.

In Notion, Apple Notes, or Google Keep, your notes can start to feel like they belong to the product first and to you second.

In Obsidian, the folder is the product.

You can move the same notes between a laptop, phone, Apple, Android, iCloud, Drive, Git, a local server, or whatever setup you use next.

If the app disappears, the archive is still plain text.

That is the payoff: your knowledge is not trapped inside one company’s interface.

The graph view only matters after that.

It is not magic knowledge management.

It is a visual side effect of notes actually linking to each other: personal facts, tech notes, decisions, workouts, daily notes, projects.

Over time, the useful part is seeing which ideas keep touching each other without forcing everything into one rigid database.

Use Obsidian for anything you want to survive tool changes:

research, decisions, technical logs, personal systems, writing ideas, operating notes.

Honest limitation:

If you need polished team permissions, dashboards, and multiplayer editing out of the box, Notion is still easier.

But if the goal is a second brain you can actually own, Obsidian’s boring file model is the feature.

[DegenCalls @Degen_calls_sol]

https://t.co/Yos03Pvbgt

[HateJPG @chopeaceus]

@crytonbuton real estate is such a massive untapped niche for these ai agency setups