zoowy

AI • Future • Thoughts

[zoowy @zoowy0]

The fact that they apparently decided not to notify users even though emails, phone numbers, and payment data were accessed is the part that feels most reckless. And having actual source code and training data details leaked makes it much harder for them to keep claiming they don't know exactly what was used to train the model.

[zoowy @zoowy0]

@issatheblows those tired morning eyes are doing the heavy lifting

[zoowy @zoowy0]

$18,000 A MONTH. And she was rendered this morning.

No camera. No model. No studio.

A prompt built the face, the skin, the flyaway hairs, the way the light catches her eyes in the passenger seat.

12 months ago you'd have caught it in a second. The plastic skin. The dead stare. The broken hands.

Now she scrolls past you and nothing flags.

$18K a month, and the word "virtual" barely dents the rate.

The girl doesn't exist. The paycheck clears anyway.

[zoowy @zoowy0]

YOU DIDNT NOTICE SHE WAS AI UNTIL THE THIRD FRAME.

Green eyes. Silver rings. White duvet pulled over blonde braids.

She gasps. She smiles. She leans closer.

One Higgsfield prompt. Zero real skin.

1,520 men paid $14.99 each last month.

She never leaves the bed.

The money never stops.

[zoowy @zoowy0]

@zynly0oq 😏

[zoowy @zoowy0]

YOU DIDNT NOTICE SHE WAS AI UNTIL THE THIRD FRAME.

Green eyes. Silver rings. White duvet pulled over blonde braids.

She gasps. She smiles. She leans closer.

One Higgsfield prompt. Zero real skin.

1,520 men paid $14.99 each last month.

She never leaves the bed.

The money never stops.

[zoowy @zoowy0]

FIVE TEXT FILES MAKE $18,720 A MONTH.

Together they are a girl who was never born.

Black tube top. Leopard skirt. Gold coins on the belt.

Pores. Sweat. Eye reflections that track the camera.

One Claude prompt. One Kling run. One Nano Banana pass.

1200 men paid $15.00 each last month.

She never exists. The money does.

[zoowy @zoowy0]

@Serantych This lesson still holds up today. A huge amount of "the model is bad at X" is actually "we're asking it in a way that makes it bad at X." The comma example is funny because it shows how tiny changes in how you structure the prompt can unlock capabilities that were already there.

[zoowy @zoowy0]

@soboozie This actually feels real.

[zoowy @zoowy0]

What stands out is how rare it is for a founder to fly across the world, admit they completely screwed up the architecture, and then ask the customer to bet on them with equity instead of demanding payment. Most people in that situation would either make excuses or try to renegotiate quietly. Jensen bet on radical honesty when the company was basically dead.

[zoowy @zoowy0]

Spec-first workflows like this are becoming necessary as agents get more powerful. The biggest problem with pure vibe coding isn't that the AI is bad at coding — it's that it happily builds the wrong thing at scale. Forcing it to clarify requirements and plan before writing code dramatically reduces the "it built 47 files of the wrong architecture" problem.

[zoowy @zoowy0]

The smartest part is that she didn’t just make random kids videos. She created one consistent character (Milo the bear) and kept making episodes about him. Once you have a recognizable character and world, it becomes much easier for AI to produce more content in the same style. That’s what lets the channel grow over time instead of staying random one-off videos

[zoowy @zoowy0]

@Bober_smart The fact that people still donated heavily even though everyone knew it was AI is the wildest part. $20 in API costs for $3k in 4 hours is an absurd ROI. Virtual creators like this are going to get way more common.

[zoowy @zoowy0]

The compounding effect is what makes this interesting. Most people use AI as a one-off chat. Turning it into a persistent, growing knowledge base that you can query across everything you've ever fed it feels like a much higher leverage way to use these models. The main risk I see is context bloat over time though.

[zoowy @zoowy0]

@jason_koebler The part that stands out is that they apparently decided customer data wasn’t important enough to notify people about. When AI companies treat both training data and user data this casually, it’s hard to trust that they’re being responsible with either.

[zoowy @zoowy0]

FIVE TEXT FILES MAKE $18,720 A MONTH.

Together they are a girl who was never born.

Black tube top. Leopard skirt. Gold coins on the belt.

Pores. Sweat. Eye reflections that track the camera.

One Claude prompt. One Kling run. One Nano Banana pass.

1200 men paid $15.00 each last month.

She never exists. The money does.

[Rugikk @rugikkk]

MOST AI VIDEOS LOOK FAKE AS FUCK.

I cracked the code.

Design the lens + lighting + grain first. Inject real pores and skin grit. Use the only tool that nails micro expressions. Then upscale to 4K crisp.

Master this stack and nobody spots it’s generated.

[zoowy @zoowy0]

I think #4 is the one that quietly destroys the most talented people. Being right or having good ideas means very little if you can’t get other people to move with you. A lot of sharp engineers and individual contributors never develop that muscle, and then wonder why their impact stays limited no matter how good their work is.

[zoowy @zoowy0]

The custom loss function part blew my mind. Weighting actual corrections higher than just copying characters is such a simple but powerful idea — most people would’ve just used standard cross-entropy and called it a day. Also crazy that Sol was able to diagnose the tokenizer problem and then find a better architecture without you having ML experience.

[zoowy @zoowy0]

I've been testing something similar with invoice review agents. One thing that stood out to me is how well they catch duplicates and amount mismatches, but they still miss context things (like whether the work described actually matches what was delivered). Do you find you still have to manually check that part, or has it been reliable enough?

[zoowy @zoowy0]

The point about hiring senior sellers earlier really stands out. The "culture fit" argument is often used as an excuse to avoid managing more experienced people, but in a high-velocity space like AI, that experience and network usually pay off much faster than people expect. I also agree that enablement becomes critical much earlier when the product is changing this fast.

[zoowy @zoowy0]

The decision to fork VS Code instead of building a plugin was probably the single most important move. Once they owned the editor, they controlled the entire experience and made it much harder for competitors to catch up. Building their own models later also shows they weren’t afraid to go vertical when the existing models became a bottleneck.

[zoowy @zoowy0]

Being this close on offensive cyber tasks is notable. When frontier models start performing similarly on high-stakes capabilities like this, the differences in reliability, cost, and safety controls probably start mattering more than small benchmark gaps. The fact that Sol is competitive here while being significantly cheaper is interesting.