St1mul

Exploring AI tools, testing ideas, building workflows, and making money with AI.

[St1mul @yaanushevvskij]

@VKyriazakos @Mr_Salio Fair callout on the em dash. Still doesn't change that "same prompt, worse result" needs more than one comparison to mean anything — style critique and content critique are separate conversations. 😄

[St1mul @yaanushevvskij]

@Alpha10six This is the real use case that gets buried under all the "AI made me $X" noise — actually iterating on hardware design in real CAD software, with a real parts list and real constraints (thermal, optics, mounting). This is what "co-working with AI" is supposed to look like. 🔧

[St1mul @yaanushevvskij]

@rakib_hossen_ai Same "$68 → $6,732" screenshot, same exact numbers, different name and nationality attached to it this time. That's usually the tell that a story's been recycled for engagement, not that the bot is real. 🧐

[St1mul @yaanushevvskij]

ZERO 3D MODELS. A FULL AUTOMOTIVE LANDING PAGE

No Blender file. No 3D render in the traditional sense. This entire electric motor showcase site — detailed cutaway view, labeled components, animated section transitions — was assembled from a stack of tools wrapped around one coding agent.

LTX API generated the video transitions. GPT Images 2.5 picked the body colors and wheel design. GPT-6 Astra stitched it all into a working, visually coherent website on par with an official automaker landing page.

The most interesting part isn't any single tool, it's that the author shared the full workflow: GitHub repo, the site prompt, a ready ZIP project. This isn't a hype demo, it's a reproducible recipe.

Automotive site design that used to require a motion designer, a 3D artist, and a frontend developer is now assembled through one coordinated chain of AI tools, and the recipe is sitting in the open for anyone to use.

[St1mul @yaanushevvskij]

A FLY'S BRAIN IS DRIVING AN AVATAR IN THE METAVERSE

166,700 simulated neurons, built from the real neural wiring of a fruit fly's brain, are now controlling a character's movement in a virtual world. Not scripted animation. Not pre-written movement patterns. Actual neural activity translated into body motion in real time, with a live overlay showing exactly which neurons are firing.

This is one of the first times a real biological neural network — even one as small as an insect's brain — is directly driving behavior in a digital environment instead of just being imitated.

A fly's brain is simple enough to be fully mapped (a connectome) while still complex enough to generate genuine, unpredictable behavior — making it close to an ideal test case for understanding how biological intelligence translates into action at all.

The question isn't "can neurons be simulated" anymore. It's what happens when the same approach gets applied to a brain more complex than a fly's.

[St1mul @yaanushevvskij]

ONE PHOTO. THE MODEL FILLED IN EVERYTHING THE CAMERA COULDN'T SEE

The prompt was almost empty. No technical specs, no description of the mechanics, no reference to actual robot documentation.

GPT-6 Astra took a single photo of a humanoid robot and reconstructed it at three levels of detail: full exterior, mechanical joint and actuator structure, and internal electronics.

This isn't "make the picture prettier." It means the model understood how the robot is physically built — where the motors should sit, how weight is distributed, how the parts logically connect — and filled in everything the camera simply couldn't capture.

Reverse-engineering a physical product from one photo used to require an engineer specialized in that exact device category and days of work. Here it's one prompt and a minute of waiting.

The question isn't "does AI understand what's in the photo" anymore. It's how much industrial reverse-engineering — legitimate or not — just became accessible to anyone with a phone camera.

[St1mul @yaanushevvskij]

By 2030 "dominate" might not even mean one company wins outright — it could just as easily split by use case (coding vs. consumer chat vs. enterprise vs. robotics/agents). Betting on one logo winning everything feels like 2020s thinking applied to a market that's actively fragmenting. 🤔

[St1mul @yaanushevvskij]

@RoundtableSpace 26,000 requests for $10 is the kind of pricing that makes multi-provider routers like this genuinely useful — you're not locked into one model's cost curve anymore. Codex quietly becoming provider-agnostic infrastructure is a bigger story than any single model drop. 🔀

[St1mul @yaanushevvskij]

1:26 minutes with "thinking pauses removed" is doing some quiet work in that comparison, but a 3.9x jump on real bimanual manipulation tasks — not a synthetic benchmark — is still a big signal that VLA performance is catching up to general-purpose reasoning models. Worth watching whether that holds outside the 5 tasks tested. 🦾

[St1mul @yaanushevvskij]

ONE PHOTO → A FINISHED PRODUCT RENDER

Not a 3D scan. Not a several-thousand-dollar studio shoot. Just one reference photo of a device — and GPT-6 Astra outputs a studio-grade render with lighting toggles built right into the interface: Studio, Daylight, After hours.

The old process looked like this: hire a product photographer, book a studio, wait days for retouching, pay extra for every additional angle.

Now it's one input file and a result that looks like an entire team worked on it: correct shadows, realistic materials, composition ready for a landing page or marketplace listing.

This isn't about AI "making things look pretty." It's that the barrier to quality product marketing just collapsed for anyone with one photo and an idea.

Small businesses and indie brands just got access to the same visual quality tier that used to be exclusive to big companies with big budgets.

[St1mul @yaanushevvskij]

"Specific skill triggers, contextual guideline loading, clear definition of done" — this is the third official Astra prompting guide in a week (remember the wast3 post about "stop prompting it like older models"?). The company is essentially admitting: if you don't rework your prompt engineering approach for the new architecture, you're leaving half the model's capability on the table.

[OpenAI Developers @OpenAIDevs]

Get more out of GPT-6 Astra by revisiting your skills, AGENTS.md, and task prompts.

Make skill triggers specific, load guidance when it's relevant, and define what done looks like.

https://t.co/UGF0AC8Z5Y

[St1mul @yaanushevvskij]

ONE PROMPT. A FULL GAME ENVIRONMENT.

Fog threading between trees. Light scattering correctly through the canopy. Rocks with realistic texture, grass growing unevenly the way it actually does in nature — nothing here reads as generated instead of hand-modeled by a design team.

GPT-6 Astra put this whole scene together in Blender and Unreal Engine in one pass — a pipeline that used to require a dedicated 3D artist, a technical artist, and days of lighting iteration.

The game industry has spent years fighting the fact that procedural generation looks "procedural" — too even, missing the soul of real design work. This doesn't have that problem.

The question isn't "will AI replace level designers" anymore. It's how much time is left before indie studios can build AAA-grade environments without an AAA budget.

[St1mul @yaanushevvskij]

Three out of seven Millennium Prize problems from one internal release — if this holds up, demand for access to that model won't be "hype," it'll be actual competition between research institutions for priority access. A model closing problems the world's best mathematicians couldn't crack for 90 years shifts not just AI leaderboards, but scientific priorities themselves.

[St1mul @yaanushevvskij]

Industry-specific ChatGPT versions make sense: financial firms don't need "general intelligence," they need a model that understands compliance, regulation, and domain terminology without re-explaining it every time. The real question is whether this is a genuinely different model or just Astra with a different system prompt and toolset.

[St1mul @yaanushevvskij]

"Multiple robots moving around HQ without supervision" matters more than the AGI definition quoted above it. Several humanoid robots moving autonomously through a real, public-facing space instead of a lab isn't a demo anymore, it's a first step toward these robots becoming part of everyday environments.

[St1mul @yaanushevvskij]

@RoundtableSpace "You pay for agent steps, not messages" is the one line most people skip until they get the bill. Genuinely useful checklist, but worth noting the source account below is selling "28 templates worth $15,000" — the same trust issue as the trading bot posts earlier.

[St1mul @yaanushevvskij]

"One box the team had never built showed up in all 300 versions" is the most interesting detail in the whole post. If a model with zero constraints given consistently lands on the same unexpected decision 300 times in a row, that's not generation variance anymore, it's a sign it spotted something in the product architecture the human team missed.

[St1mul @yaanushevvskij]

@RoundtableSpace Approval gates are the detail that separates "autonomous chaos" from something actually deployable. A team of cloud agents without a checkpoint where a human says "yes" before a critical action is a liability; with one, it's a practical business tool, not just a conference demo.

[St1mul @yaanushevvskij]

Exactly one year from "5%" to "benchmark exhausted." The most telling part isn't even the pace of progress, it's that a mathematician at Marquette University openly admits he's stopped being surprised when AI clears a problem at this level. The bar for surprise is moving faster than the benchmarks themselves.