iskrrra

🟧

[iskrrra @iskrrra1]

Dear @OlafScholz @EmmanuelMacron @JoeBiden @BorisJohnson Please #BanRussiafromSwift #CloseTheSky #SendNatoToUkraine

[plutos @plutos_eth]

The creator of Claude Code said something wild:

"I don't prompt anymore. I build loops, and the loops do the work."

That's the entire shift.

A prompt is you doing one step.

A loop is you designing the machine that does every step, on its own, until it's done.

/goal → runs to a finish line (some run for days)
/loop → fires on a schedule while you sleep

Your job stops being "write the prompt."

It becomes "design the loop."

[plutos @plutos_eth]

A $10,000 website. Built before dinner. For about $60.

not a template > not Wix > custom code written from scratch while you watch

one prompt > two free design skills > Claude asks YOU the questions, then builds

the whole trick is one line: "ask me clarifying questions" — so it designs on purpose instead of guessing

most people are still paying a freelancer $10k for what a $20 plan does in an afternoon

[plutos @plutos_eth]

Everyone's dunking on the $1,000 Steam Machine.

This guy's buying one anyway — and his reasons are hard to argue with.

He knows a PS5 or a cheaper PC beats it on raw power. he doesn't care.

1,000+ games already in his Steam library.

SteamOS that just works. one tiny box that unifies his modern library and four years of emulation under his TV.

He's not buying benchmarks. He's buying "turn it on and go."

[plutos @plutos_eth]

Farhan Thawar, Head of Engineering at Shopify:

"The learning is the collateral, not the code."

His team once spent 18 months building something — then their CEO asked, "if you could start over, how would you build it?"

They rebuilt a simpler version in three months.

Now AI writes the code, so the code was never the point.

What you keep is what you learned making it.

Don't measure your team by what they ship. Measure how fast they're learning.

[plutos @plutos_eth]

Valve wants you to feel bad for them about the price

$1,049 for a box that performs like a $499 PS5, and their excuse is "the AI boom ate the RAM supply."

Translation: you're paying for Big Tech's GPU hoarding, not for anything Valve actually built better

And the wildest part?

You can match it for $250 using a literal PS5 chip that got dumped on eBay after the crypto crash because nobody wanted it anymore

So no, this isn't a shortage problem

It's a "nobody bothered to give you the alternative" problem

So stop paying the AI tax on hardware you can build yourself in an afternoon

[plutos @plutos_eth]

"Obviously this is a stupid comparison." He ran it anyway — and the Steam Machine barely survived.

Desktop vs. Laptop, different use cases entirely.

But a cheaper gaming laptop landing this close to Valve's box says something.

Matched VRAM and RAM, a built-in display, double the storage — and only a modest frame-rate deficit across the test suite.

The Steam Machine is better.

For $400 more and no controller in the box, "better" has to work harder than that.

[plutos @plutos_eth]

You don't need Valve's Steam Machine. You need a parts list.

six Zen 4 cores (Ryzen 5 7500F), 16GB of DDR5-6000, a desktop GPU with twice the VRAM — motherboard picked by sorting on price, with a part thrown in free.

the result: ~60fps in Expedition 33 while the Steam Machine crawls at 35–40 — and it holds across a wide range of games.

Valve's box is slick.

Slick isn't the same as fast.

[plutos @plutos_eth]

The real estate investor:

"Productivity is going down. People are trying to fix something that's not broken."

He's not a skeptic. He's actively deploying AI — just on the one bottleneck that actually costs him money

That's the whole game

Not "are you using AI." But: which exact problem, with a number attached to it, does your AI solve?

If they can't name the number, they still in the exploration phase

The job is isn't to sell them AI

It's to find the one process that's quietly costing them $500/month or 20 hours/month in labor

And replace it with something that runs on its own

[plutos @plutos_eth]

We've spent 5 years talking about AI replacing knowledge workers

Meanwhile a florist is contorting their body to hold a phone while trying to make a bouquet

That's the actual market

The person who learned a craft and now spends 40% of their day being an accidental receptionist

An AI phone agent isn't expensive to run

It's answers custom questions, manages orders, handles the back-and-forth a web form can't

And that's what gives the florist their hands back

[plutos @plutos_eth]

A realtor starting to run her business on Claude:

"I'm taking everything from ChatGPT — like two years of stuff — and downloading it into Claude so it can start working fresh from all that data."

Her stack is deliberately simple:

Export the ChatGPT archive into Claude, add the Chrome extension for live sites, run the desktop app for chat and Cowork.

No code yet — that's "down the road."

Just the everyday setup that lets one person automate the busywork and spend more time with clients.

[plutos @plutos_eth]

Rahul, AI automation agency founder:

"Capture every bit of information. Decide what to do with it. Then take action. Jarvis does the whole sequence."

He built it with Claude Code — first for his own agency, now for real estate clients.

A new inquiry lands at midnight: it qualifies the lead, books the viewing against your calendar, updates the CRM, and chases the missing paperwork before anyone on the team wakes up.

Not one automated workflow.

One system running the whole business.

[plutos @plutos_eth]

Most "AI for small business" pitches end in the demo

Here's why:

They're selling a tool to people who don't want a tool

They want the problem gone

The dentist doesn't want "an AI scheduling assistant."

He wants to stop losing $2,000 every time a chair sits empty for an hour

Sell the filled chair, not the assistant

The people who get it aren't building more sophisticated systems

They walk in with one number — what the problem costs — and leave with a check

[plutos @plutos_eth]

https://t.co/FchOTmpz1d

[plutos @plutos_eth]

Stop selling the AI. Sell the pain point.

The agencies closing $15k deals aren't pitching features or "latest technology."

They ask one question:

How many calls a day are you missing — and what's that costing you a year?

A receptionist runs ~$60k/year.

The AI doesn't miss calls, doesn't take sick days, doesn't put four of five callers on hold.

Framed against the problem, $15k stops sounding expensive.

Framed as "an AI tool," it never will.

[plutos @plutos_eth]

Most people think AI agents fail because the model isn't smart enough

That's almost never the reason

The real failure: no independent judge

The model writes the answer, grades the answer, and tells you it did a good job

Of course it does, it knows what it meant to do

A verifier that reads the output cold, with no memory of the reasoning

No emotional attachment to the draft catches things the worker never will

Separate the worker from the verifier

That one design rule fixes more broken agent workflows than any prompt rewrite ever will

[plutos @plutos_eth]

Andrej Karpathy:

"The agent works in an autonomous loop, accumulating git commits to the training script as it finds better settings — the architecture, the optimizer, all the hyperparameters."

He stripped his nanochat trainer down to one file, ~630 lines, and handed it to an agent.

The human writes a markdown file saying what to improve.

The agent runs the change → train → score loop on its own, keeping only what beats the last best.

~100 experiments a night.

He left it running and woke up to a model 11% faster to GPT-2 level.

[plutos @plutos_eth]

Satya Nadella, CEO of Microsoft:

"You really are going to have to perfect that loop."

He's not talking about prompts

He's talking about the fact that right now, you are the loop

You start it, you move the output, you decide what's next, and the moment you step away, everything stops

The people pulling ahead aren't prompting better

They built a system that runs without them

[plutos @plutos_eth]

Pedro Franceschi, CEO of Brex:

"Every good AI product you've ever used is an agent loop with tools. That's it."

At Brex, when a customer's chat with their expense agent goes badly, it files a bug automatically.

That bug triggers another agent that rewrites the code and the prompts until the case passes — and only if it can't does a human step in.

The goal, he says, is a system that watches everything overnight and re-learns by morning.

A company that debugs itself.

[plutos @plutos_eth]

Brian Grinstead, Mozilla Engineer:

"Our goal isn't a bunch of bugs that are hard to find. Our goal is zero bugs."

Firefox is tens of millions of lines of code. And human focus fades in a way an agent's doesn't.

So Mozilla pointed agents at it:

Score the riskiest files → run a main loop → verify in a sub-loop → hand a human an actionable fix.

~500 security bugs fixed in a single month.

The bugs were always there. what changed is the cost of finding them.

[plutos @plutos_eth]

Dario Amodei, CEO of Anthropic:

"In a world where AI can generate anything, having basic critical thinking skills may be the most important thing to success"

That's the hidden risk in every loop you build.

Hand the whole task to the agent and stop understanding it, and the muscle quietly atrophies.

His fix isn't to avoid the tools. it's to use them deliberately.

The people who design loops and still understand the work stay in control.

The ones who delegate their thinking become passengers in their own process