Sancho

Wizards don't FOMO. Wizards conjure.

[Sancho @Sancho_Wizard]

https://t.co/pKJPxIEsRr

[Sancho @Sancho_Wizard]

HE MADE 300 UGC ADS FOR $2.57. THE AGENCY QUOTED HIM $2,000.

No editor. No app. No timeline.

He hasn't opened a video editor once.

Here's the entire workflow.

Fable 5 connected to a video rendering tool through MCP.

He types one sentence: "make 10 UGC videos."

Then walks away.

The model picks the hooks, writes the scripts, sends the render jobs, and pulls back finished videos — while he does literally anything else.

Your own face or licensed creator faces. Batch after batch.

The craziest part?

The math.

An agency quoted $2,000 for what his setup produces for $25.

That's not a discount. That's an 80x collapse in the cost of content.

His words: "My ads make themselves now. This used to take my whole week."

While competitors are still booking editors and waiting on revisions...

His renders are already live and his next batch is already queued.

The cheat code for UGC isn't a better editor.

It's removing the editor from the equation entirely.

[Sancho @Sancho_Wizard]

https://t.co/pKJPxIEsRr

[Sancho @Sancho_Wizard]

@rugikkk The realism jump in the last few months is insane. Condensation dripping down the iced coffee is the kind of detail that fools people instantly. Clean workflow.

[Sancho @Sancho_Wizard]

@crimson_moti Wild breakdown. The frozen-smile AI accounts always fall apart the second there's a reaction shot. Live mocap driving the face is the missing piece nobody talks about.

[Sancho @Sancho_Wizard]

@Dexonfxf Solid proof that niche B2B solutions often beat "trendy" startups on profit. Curious how he landed those first clients — cold outreach or word of mouth?

[Sancho @Sancho_Wizard]

AN AI STARTUP CO-FOUNDED BY A FORMULA 1 DRIVER JUST HIT $9M ARR IN 8 WEEKS.

Eight weeks.

Most startups don't even finish their pitch deck in eight weeks.

And it's not a general AI tool. That's the entire point.

It's vertical AI — one model, one industry, trained on that domain's data, terminology, and workflows out of the box.

ChatGPT knows a little about everything.

Vertical AI knows everything about one thing.

That's what people pay for.

Here's why this window is open right now.

Unstructured data — the messy documents, calls, and records every industry drowns in — just became workable. LLMs finally read it like a native.

And the business model flipped: you don't sell software anymore. You sell the outcome.

Even YC's official request for startups this summer is pointing the same direction — vertical agents for medicine, diagnostics, personalized care.

The craziest part?

The horizontal AI race is over. A few giants won it.

But every single industry vertical is still wide open.

Legal. Logistics. Dental. Construction. Insurance.

The gold rush isn't in building general AI.

It's in picking one boring industry and owning it completely.

[Sancho @Sancho_Wizard]

https://t.co/pKJPxIEsRr

[Sancho @Sancho_Wizard]

SOMEONE MAPPED FABLE 5'S ENTIRE AGENT WORKFLOW INTO 9 DIAGRAMS. AND YOU CAN FEED THEM TO CLAUDE WITH A SCREENSHOT.

One whiteboard. Nine diagrams. The whole agentic OS, drawn out.

The Heartbeat. Orchestrator workers. The execution advisor. Trust ledgers. Standing goals. Quorum. Spawning. Compose. Ratchet.

Every pattern the top 1% of agent builders use — laid out visually, box by box, arrow by arrow.

Here's the part most people miss.

Claude reads diagrams.

You don't need to explain your architecture in three paragraphs of text anymore.

Screenshot the diagram. Paste it in. Say "build this."

The model sees the boxes, the arrows, the flow — and turns the picture into working code.

The craziest part?

Everyone's been writing longer and longer prompts trying to describe agent systems in words.

The answer was to stop describing and start drawing.

Nine screenshots is the entire playbook.

Most people will scroll past this.

The ones who don't will have Fable 5's whole workflow running by the weekend.

[Sancho @Sancho_Wizard]

https://t.co/pKJPxIEsRr

[Sancho @Sancho_Wizard]

@ajs6888 Exactly 💯

[Sancho @Sancho_Wizard]

99% OF PEOPLE ARE USING CLAUDE COMPLETELY WRONG.

They're chatting with it.

This guy turned Claude into a complete Agent OS instead.

One AI became a planner, a team of builders, a verifier, and a memory that gets smarter after every successful run.

No custom model.

No enterprise infrastructure.

Just the architecture Claude should have had from day one.

Read the article below to see how it works.

[Sancho @Sancho_Wizard]

https://t.co/pKJPxIEsRr

[Sancho @Sancho_Wizard]

A CHINESE DEVELOPER RAN A 70-BILLION PARAMETER MODEL ON A MACBOOK. AT 35,000 FEET. FOR 11 HOURS.

No WiFi. No cloud. No API.

A transatlantic flight — the one place on Earth where the internet truly dies.

And his AI never stopped working.

Llama 70B, running fully offline, on a laptop sitting on a tray table.

Coding. Writing. Reasoning. The entire flight.

While everyone around him watched movies, he had a frontier-class model running on battery power.

The craziest part?

Two years ago, running a model this size meant a server rack and a five-figure cloud bill.

Now it fits in a backpack and survives airplane mode.

Think about what that actually means.

No rate limits. No subscription. No servers going down. No company deciding your access.

The model is just... yours. Anywhere. Even over the Atlantic.

Everyone keeps arguing about which AI lab will win.

Meanwhile, the models quietly became small enough to walk out the door.

The cloud isn't the future of AI.

It's the training wheels.

[Sancho @Sancho_Wizard]

https://t.co/pKJPxIEsRr

[Sancho @Sancho_Wizard]

NVIDIA AND MICROSOFT JUST BUILT A COMPUTER THAT RUNS EVERYTHING. LITERALLY EVERYTHING.

One petaflop of AI performance.

128GB of unified memory.

Blackwell GPU. 20-core Grace CPU. Custom-built with MediaTek.

And here's the line that matters:

100% of NVIDIA's software stack runs on it.

Every application NVIDIA has ever created. All of CUDA. Ray tracing, DLSS, TensorRT - no exceptions, no compromises.

Microsoft and NVIDIA meticulously optimized the whole thing together.

But the real headline isn't the specs.

It's the last thing they said.

It runs agents.

Not chatbots. Not autocomplete.

Autonomous AI agents, running locally, on a machine that sits on your desk.

The craziest part?

A petaflop used to mean a supercomputer facility with its own power plant.

Now it's a box with 128GB of memory and zero cloud dependency.

While everyone's still renting intelligence by the token...

The two biggest names in computing just shipped a machine designed to own it outright.

The cloud was never the endgame.

It was the rental period before hardware caught up.

[Sancho @Sancho_Wizard]

https://t.co/JpYHp5OPoV

[Sancho @Sancho_Wizard]

@gerardsans Seems like I'll have an interesting evening ahead, reading your articles.

[Sancho @Sancho_Wizard]

YOU'VE BEEN BUILDING A SECOND BRAIN. ANTHROPIC JUST FOUND ONE GROWING INSIDE CLAUDE.

They call it the J-space.

A hidden workspace inside the model's neural activations where it thinks about concepts without writing them down.

Not the output. Not the chain of thought.

A layer underneath both.

Your second brain stores what you've thought.

This one is where thinking happens before it becomes words.

Researchers watched Claude silently notice bugs in code. Identify images in its head. Hold ideas that never made it into the output.

The craziest part?

Delete the J-space and Claude still talks fluently. Still recalls facts. Sounds completely fine.

But multi-step reasoning collapses.

Automatic thinking survives. Deliberate thinking dies.

Everyone's been linking notes, trying to build a brain outside their head...

The models quietly grew one of their own.

And now we can finally look inside it.👇

[Anthropic @AnthropicAI]

New Anthropic research: A global workspace in language models.

Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with.

We found a strikingly similar divide inside Claude.

[Sancho @Sancho_Wizard]

@gerardsans Calling it "just next token sampling" is the same move in reverse - flattening real research into a slogan. The paper ships methods, a public demo, and outside expert review. That's falsifiable science. Critique the methodology, not the framing.

[Sancho @Sancho_Wizard]

THE CEO OF THE COMPANY BEHIND CLAUDE JUST ADMITTED HIS ENGINEERS DON'T WRITE CODE ANYMORE.

The AI does.

His people just look it over.

Let that sink in.

The company building the world's smartest AI... isn't typing its own code.

Claude writes it. Humans review it. That's the whole workflow now.

Which means something nobody wants to say out loud:

Claude is helping build the next version of Claude.

The loop is already running. Inside the one company that would know first.

The craziest part?

This was supposed to be years away.

"AI writing most of the code" was the sci-fi scenario. The thing skeptics laughed at.

It's now just... Tuesday at that office.

While Twitter argues about whether AI can really code...

The founder building it already handed over the keyboard.

You won't get a warning when this hits your industry.

You'll just wake up one day and realize the review is the job now.

[Dexonx @Dexonfxf]

https://t.co/rGwTnoqFHP

[Sancho @Sancho_Wizard]

@Dexonfxf Yep . Sounds like Karpathy was right

[Sancho @Sancho_Wizard]

@AnthropicAI Wild timing - everyone's building second brains in their notes apps, and Anthropic just found Claude already has one built in. The J-space is literally thinking before words. Karpathy was right: the interesting layer was never the output.

[Sancho @Sancho_Wizard]

@Dexonfxf

[Sancho @Sancho_Wizard]

EVERYONE'S WRONG ABOUT THIS ONE FILE. AND KARPATHY QUIETLY POSTED THE FIX.

11,000+ people saw the notes.

Almost none of them understood what they were actually looking at.

It's called CLAUDE..md.

Most developers treat it like a settings file. A place to dump preferences and hope the AI listens.

That's not what it is.

AI agents fail when they lose context. The model quietly ignores its own blind spots. The fix isn't a longer prompt — it's a tighter loop and a smaller state.

Someone took Karpathy's original notes and turned them into an actual protocol.

Four rules. That's it.

Read first. State your assumptions. Make the smallest possible change. Then verify it actually worked.

The difference this makes is brutal to watch.

Messy AI diffs look like genius work when you don't check them.

Surgical diffs look boring — until you realize they're the only ones that actually hold up.

Even Anthropic's own engineers describe their job differently now.

"I write loops," not prompts.

This file isn't documentation.

It's the rulebook that decides whether Claude Code writes like a senior engineer or a chaotic intern.

Most people never opened it.

The ones who did stopped rewriting everything the AI touched.

[Sancho @Sancho_Wizard]

https://t.co/JpYHp5OPoV

[Sancho @Sancho_Wizard]

THIS VS CODE EXTENSION PAYS YOU MONEY WHILE CLAUDE CODE IS THINKING.

Every time you run a prompt, there's dead time.

The spinner. The "thinking" animation. Those few seconds where you're just waiting.

Someone realized that dead time is worth money.

The extension is called Kickbacks.

It installs directly into Claude Code and Codex. While the AI is processing your request, it shows a small sponsored ad in that idle loading space.

You get up to 60% of the ad revenue.

Just for coding like you normally would.

No extra clicks. No extra work. The exact same workflow you already run every day.

The craziest part?

The time you spent staring at a loading spinner, annoyed, waiting for output — that's now the part of your day quietly making you money.

Thousands of developers run Claude Code for hours daily.

Every one of those thinking spinners was empty space.

Someone just turned empty space into a revenue stream.

You're not being paid to code.

You're being paid for the seconds you were already wasting.

[Sancho @Sancho_Wizard]

https://t.co/JpYHp5OPoV