I asked Qwen4-Max-Preview to create a lake scene. It generated the code, opened it in a browser, checked for console errors, and kept making refinements until the entire scene worked perfectly. I didn’t have to intervene even once.
The prompt was just a single paragraph:
"A tranquil lake at night, 30–40 earthen lamps (diyas) with flickering flames, lanterns rising into the sky, and a mist hovering over the water. Clicking on the lake should spawn a new diya, and pressing 'F' should trigger fireworks."
It wrote a 909-line `index.html` file using Three.js and then tested its own work. It opened the page, identified any issues, fixed them, and verified the results again. After 84 iterations using its tools, I had a fully functional scene—without having to debug a single line of code myself.
Clicking anywhere on the water causes a diya to appear and float, creating gentle ripples. Each diya’s flame flickers uniquely. Pressing 'F' launches fireworks in the sky above the hills, with their reflections shimmering on the lake's surface.
The entire experience was seamless and impressive. I provided just one prompt—no back-and-forth, no manual tweaks or repairs required on my part.
Diwali may have passed a week ago, but I still have a lake where I can light diyas.
Qwen4-Max-Preview is now available within 'Qoder'—Alibaba’s coding agent—which is currently offering an 80% discount for a limited time.
An AI company used by ~10,000 businesses was breached twice in one month.
The prize was bigger than the AI model itself: API keys, secrets, and stored AI chats.
Even some Notion AI chat transcripts were exposed by that company.
If you build with AI APIs, your keys now reside in more places than you might think.
Here's what happened, and 5 things to do tonight:
This guy literally explained why some people become successful… while others stay average.
And the reason is uncomfortable.
It’s not just hard work.
It’s not luck.
It’s Game Theory.
Once you understand this, you’ll start seeing the “game” behind success differently.
Watch this. 👀
Building an AI agent that works is impressive.
Building one that actually learns from experience and gets better over time is a completely different challenge.
The real question isn't whether your agent works.
It's whether it's improving.
¿Entonces vincula las notas del cliente desde Slack con el fotograma exacto?
Yo todavía tengo que hacer esa correspondencia manualmente y me lleva todo el día. Quiero ver si realmente identifica el fotograma correcto
Black Friday mission status: Target located. Product secured. Agent ready for the final showcase.
I explored this secret-agent-style product ad in CapCut PC, using the telescope shot to create curiosity, the infiltration scenes to build tension and the final product shot to complete the story.
Timeline Editing helps me refine the sequence, Multi-Track Editing keeps the visuals and audio organised, and CapCut Agent EditPilot can assist with timeline adjustments through natural-language instructions. I can review everything and make the final tweaks before publishing.
See the creative workflow: https://t.co/UV0mx6s0tX
#capcutpc #capcutmarketing #capcutblackfriday #CapCutDidThat #chatwithcapcutagent
Put two AI agents from different labs into a small game, and they often fail to work together.
They fail because each one assumes their own guess is "clear."
This is from a new Oxford paper, accepted at NeurIPS 2026.
The game is almost foolproof:
→ Alice sees a cat or a dog. Bob has to name it.
→ Alice can only indicate with a light. No one agreed on what the light means.
→ Or Alice gives 5 points for removing a wall so Bob can only see. A correct answer earns 10 points, so the safe total is 5.
Each agent silently deduces a meaning for the light.
When a partner comes from a different lab, the meanings often don't match, and both lose.
Models tested: GPT-5.2, Gemini 3 Pro, Grok 4.1 Fast, DeepSeek V3.2, and Cloud Opus 4.5.
The solution was to add a small instruction to each agent's prompt:
Don't plan a course of action your partner might not share.
With this:
→ All 16 model pairings scored exactly 5.00 in the cat-and-dog game. Every pair took the safe wall.
→ Across 9 games, failures decreased to 6 for Grok 4.1 Fast and 5 for GPT-5.2.
This doesn't mean that AI agents can't cooperate.
These are toy games, the models are a generation behind today's generation, and the "play safe" rule could block strategies with better payoffs.
The future of agents is that your agent deals with agents from other companies: booking, buying, negotiating.
Those agents need written, shared rules, just like drivers need traffic rules.
Link👇
"The suspense. The stealth. The product reveal. Honestly, this agent deserves a promotion. 😂
I explored this spy-inspired Black Friday ad in CapCut PC, turning a product image and original video references into a short story.
When different shots have different lighting or colour tones, Color Adjustment helps make them look more consistent. It's a useful finishing step before switching from the darker mission scenes to the final product showcase.
Check out the workflow: https://t.co/wk5a6VOIH5
#capcutpc #capcutmarketing #capcutblackfriday #CapCutDidThat #chatwithcapcutagent
Claude dropping Whiteboards, Sheets, and Data is a pretty big signal.
But I like puppyone’s answer to it even more:
open-source whiteboard, spreadsheet, and dashboard tools that work with any model, not just one ecosystem.
Free, open source, and already live.
https://t.co/tjnyCuK7Ma
Your AI remembered. So what?
A feature nobody uses is just clutter. AI memory is no different.
Don’t show me what the agent saved. Show me what it did differently because it remembered.
YOUR KITCHEN CHORES ARE A SURPRISINGLY TOUGH AI TEST.
Hold. Scrub. Turn. Regrip. Now timestamp every move.
@Perceptroninc just dropped EgoHome, and the level of detail is WILD 🔥
It benchmarks how well models understand first-person video, from the overall task down to what each hand is doing.
across roughly two hours of household footage:
> 417,912 hand boxes at 30 fps
> 1,924 subtask segments
> 5,032 hand-action segments
> 5,277 free-form hand-caption segments
All 4 annotation types appear on every clip:
✦ 3,371 segments last under a second
✦ 773 are shorter than half a second
✦ The shortest? 100 milliseconds
Perceptron reports 0.798 on subtasks at 8 fps, alongside GPT-6 Astra’s 0.797 at 4 fps. Astra leads on hand actions and captions.
Oh.. and EgoHome also introduces PSJ.
it allows different valid ways to divide an activity into steps.
ppl don’t describe the washing-up at identical levels of detail either.
ordinary footage, demanding questions: which hand moved, what did it do, and when did the activity change?
Public dataset. Open-source evaluation code. Explore it below 🧵↓
🚨 BREAKING: I asked Claude to improve my LinkedIn profile.
It didn’t just improve it. It made it a recruiter magnet.
Here are the 7 exact prompts I used: