Mellon

19 // i don't know how but i will win // ai and trading // dms open for u

[Mellon @Mellon0x]

CHATGPT SOL 6 DREW GUCCI MORTY IN MS PAINT. SORRY, PICASSO, THIS ONE'S BETTER.

I asked ChatGPT Sol 6 to draw Gucci morty and gave it control of Paint. No image generator. Just a cursor, a canvas, and the model doing the work itself.

> Pick a tool.
> Move the mouse.
> Build the face piece by piece.
> Keep going until Gucci Morty exists.

The best part is watching it happen. Every line appears on screen instead of arriving as a finished image.

I expected chaos in the toolbar. Instead, ChatGPT Sol 6 turned a blank canvas into Gucci Morty like it had an art-school deadline and something to prove.

Picasso can keep the museum. I'm framing the screenshot.

[Mellon @Mellon0x]

@nonepcbl Im glad you liked it

[Mellon @Mellon0x]

GPT-6 SOL JUST MADE ASTRA 6 LOOK LIKE AN EXPENSIVE HABIT.

I tested models on agent runs in business tasks, comparing accuracy, speed and cost. The gray line is July's frontier. The pink area shows how far the new models pushed past it.

The biggest mover is GPT-6 Sol (high). It reached 100% on this benchmark's AI coding tasks and landed at the top for quality per dollar. Among models at that quality level, it was also the fastest.

The rest of the field makes the trade-offs obvious:

> GPT-6 Luna (high/medium) pushed the low-cost frontier.
> Kimi K3 placed second in accuracy, but ran slower and cost more.
> Astra 6 (medium) lagged on both price and speed.
> Opus 5.5 ranked 26th, with refusals hurting its business-task score.

One perfect coding score means this set needs tougher edge cases, not that coding is solved.

My takeaway: if you run agents for actual work, Sol and Luna deserve a fresh test in your own pipeline. Astra 6 now has to justify its bill.

The win isn't one model beating another. It's how much more quality you can buy for the same money than you could in July.

[Mellon @Mellon0x]

@nonepcbl this is a very useful system bro

[Mellon @Mellon0x]

JEV GOT 7 KILLS IN DOOM WITHOUT SEEING A SINGLE FRAME.

I gave the model a text briefing instead of a screen: health, ammo, enemies, their distance and direction, nearby walls, and the door ahead. Then I let it pick what to do next.

The setup runs FreeDoom through ViZDoom. Jev gets eight choices: strafe left and fire, strafe right and fire, advance, retreat, collect an item, explore, turn around, or open a door. It returns probabilities; the code executes the highest-ranked choice for half a second.

The fixed menu is most of the prompt. Each new update only needs to explain what changed around the player, so Jev can spend its next decision choosing between fighting, moving and getting out of trouble.

> About 2 decisions/sec; median response: 443 ms.
> Input/turn: 740 tokens; ~620 fixed menu, ~100 state.
> Cost: ~$0.00003/turn; ~$0.0045/min; ~$0.27/hour.
> Two-minute run: 289 decisions, 7 kills, 0 deaths.

The code handles aiming and getting around walls. Jev handles the action choice. So the interesting part is how far a short text briefing and eight options can take a model. Apparently, far enough to survive the run.

[Mellon @Mellon0x]

About 30,000 AI agents were working at once on Anthropic's most-used internal platform in August 2026.

Their assignments were research and engineering. Employees delegated work; agents kept working for long stretches and passed tasks to one another.

Picture the scale: thousands of software workers operating in parallel while the lab develops its next models.

Anthropic built two layers of oversight around them. One monitor checks every action before it runs and can stop it. Another reads the activity afterward, looking for patterns across agents and over time.

At 30,000 agents, the supervision target is an entire working system.

The race to build AI now includes another race: building oversight that can follow it.

[Mellon @Mellon0x]

@k1rallik bro this is a really interesting project I'll be keeping an eye on it

[Mellon @Mellon0x]

@nonepcbl yeah bro youre absolutely right

[Mellon @Mellon0x]

Claude now leads 26% of Anthropic's measured AI R&D work.

A researcher gives it a high-level goal. Claude does most of the task, works through the problems it finds, and returns a result for human review.

That changes the rhythm of a frontier lab. The model moves from answering questions about AI development to carrying out substantial parts of the development itself.

Anthropic says Claude collaborates or leads in more than 90% of the work it measured.

As that share grows, the lab can run more experiments, fix more infrastructure, and test the next generation faster.

How fast does AI development move when the research tool begins doing the research?

[Mellon @Mellon0x]

MY FLY WAS HEADING FOR THE TWIN TOWERS WITH AN ENTIRE F*CKING PLANE.

I used ChatGPT Astra 6 to train one fly as a pilot and create another for safety. Two insects. One aircraft. A project nobody should have let me manage.

At first, the pilot followed instructions.

Then the little bastard stopped responding and headed for the towers.

Cancel. Nothing. Override. Nothing.

I looked at the safety fly. It was sitting there like it hadn't read a single f*cking email.

Then its back split open.

Metal limbs unfolded. Armor locked into place. My backup insect became a giant robot and intercepted the plane before it reached the buildings.

THE F*CKING FLY HAD A SECOND PHASE.

I'd expected it to take over the controls. Maybe force a landing.

Astra 6 had apparently read “handle emergencies” and built something that needed its own defense budget.

The towers were safe. I was staring at a skyscraper-sized insect I'd nearly swatted twenty minutes earlier.

Next time I'm reading the f*cking feature list.

[Mellon @Mellon0x]

@nonepcbl thats right, even she cant handle it

[Mellon @Mellon0x]

I GAVE CHATGPT ASTRA ONE JOB: CLEAN NEW DELHI. THE CITY TREATED IT LIKE A PERSONAL ATTACK.

I asked Astra agents to map the worst blocks, deploy cleaning robots, route crews, and keep one street and a section of canal spotless.

The pavement appeared. The water stopped resembling soup.

Then a crowd arrived.

Bags hit the curb. Cups went into the canal. Someone inspected the clean wall and apparently decided it needed urine.

Astra detected the damage, recalculated the routes, added bins, signs, barriers, and another cleanup shift.

Clean again.

By sunset, the first cleanup looked deleted from history.

So Astra tried a third time. More robots. Better predictions. Faster collection.

Same ending.

Its final report didn't say “mission failed.” It said “environment repeatedly restores previous state.”

That's AI language for: the garbage has autosave.

I asked Astra to clean New Delhi.

It cleaned the street twice, watched humanity reinstall the mess, and refused the next task.

The robots cleaned the street.

The final report just said: “I'm not cleaning this sh*t a fourth time.”

[Mellon @Mellon0x]

@nonepcbl yeah and I dont really like his principles

[Mellon @Mellon0x]

MY FLY WENT TO INDIA WITH AN AI TRAVEL AGENT. NOW THE AGENT NEEDS AN EVACUATION PLAN.

I connected ChatGPT Astra to a fly's brain and had the AI translate its signals into travel choices.

Japan first. Looked around. Next.

Vietnam? Nice. Keep moving.

Then New Delhi.

The fly stopped requesting destinations.

Paris? Rejected. Bali? Rejected. An entire f*cking planet, and this idiot had found home.

Astra tried changing the destination itself.

The fly kept overriding it. Back to India. Every time.

So I checked its sightseeing priorities: feces, the Ganges, Indian street food.

This wasn't an itinerary. It was a fly's vision board.

A discarded snack beside an open drain got more attention than anything we'd shown it in Japan.

I was trying to organize a holiday. This little bastard was shopping for retirement property.

Astra kept suggesting departures. The fly kept voting for another night.

I gave an insect access to AI and international travel. It used both to become impossible to evacuate.

I'm never going to India. My travel advisor eats sh*t, and this was its strongest recommendation.

[Mellon @Mellon0x]

@martynov014 good music mate😭

[Mellon @Mellon0x]

A new architecture gets the headlines. Better training data can do more of the actual work.

That's a useful reality check from this lecture.

An LLM learns patterns from the material it sees. Feed it repetitive, low-quality text and a clever design won't magically fix the result.

Better data, meaningful tests and efficient computing infrastructure can matter more than another small Transformer tweak.

The infrastructure matters because it determines how much training and experimentation you can actually afford.

Watch this if every AI launch has started sounding like the same parameter-count contest.

The interesting question: what changed in the training, and did it make the model more useful?

[Livsun @L1vsun]

https://t.co/qOBb2XGkl3

[Mellon @Mellon0x]

@k1rallik wow thats an interesting experiment

[Mellon @Mellon0x]

@nonepcbl I agree bro thats really how it is

[Mellon @Mellon0x]

HOLY SHIT. I gave a f*cking FLY the ChatGPT Astra 6 + Blender treatment. Now it looks like the final boss of my kitchen.

I wanted to make an insect so unnecessarily cinematic that you'd stop scrolling for something you'd normally swat away.

So I used Astra 6 to help write the Python for a Blender scene and animation. The brief: give this tiny bastard an entrance it absolutely hasn't earned.

Compound eyes fill the screen. Wing veins come into view. An abstract neural network lights up.

Then it lifts its front end like a lowrider.

All that dramatic buildup. For a fly with fucking hydraulics.

Want to make your own?

> Install Blender, Python 3.11+ and FFmpeg.
> Clone the repo, install requirements.txt in a virtual environment, and open assets/Drosophila.blend.
> Run scripts/render.py through Blender: test one frame, then render the full sequence.
> Run scripts/export.py with Python to build the MP4. Customize the text, timing and audio.

Full source + setup below. The included edit exports without sound and needs no AI subscription to reproduce.

https://t.co/c3sbNBCCCD

I was supposed to get this thing out of my kitchen. I gave it a premiere.

[Mellon @Mellon0x]

@Gyome1_ this is actually a really interesting and useful project thank bro