Alibaba just released a 2.4T-parameter open-weight model.
Qwen3.8 includes:
→ 95B active parameters
→ 1M-token context
→ Parallel tool calls
Unsloth compressed it from 4.9 TB to 397 GB.
Open models are scaling fast.
https://t.co/CctisiIldB
[OpenSource - UI/UX Design & AI Coding Agents]
Relying on raw AI "vibe coding" often produces visually generic, ugly, or unpolished web layouts that lack true design hierarchy. 🛑
taste-skill (73K+ GitHub stars) is an open-source design skill that injects elite UI/UX aesthetics, visual rules, and design discipline directly into your AI coding agents.
Here is how this open-source skill elevates AI-generated frontend design. 🧵👇
1/4
Anthropic just dropped a 30-minutes workshop on loop engineering.
how to turn one spec into hundreds of shipped features with AI.
• 5:31 - build memory for your coding loop
• 8:22 - the 4 files every loop needs
• 11:02 - the loop: test → pick a feature → build → test → commit
• 13:06 - build one feature at a time to avoid running out of context
• 33:03 - ship an MVP fast, then let the loop keep improving it
most people use AI to finish one task.
Anthropic uses AI loops that can keep building for hours or even days.
watch it today, then read the full loop engineering playbook below ↓
One prompt. $423 in compute. 690 million tokens. Out came a fully playable cel-shaded boat racing game built in Three.js.
Not a demo. A finished game. Infinite ocean, lap timers, opponents, the kind of visual polish a real studio ships.
The "vibe coding is a toy" crowd is going to have a rough week.
Here is the math that matters. A small studio quotes this at $50,000 and three months. Vyom got it for the price of a nice dinner for four. And $423 is the ceiling, not the floor. Same result costs $40 a year from now.
The prompt is the part worth studying. He gave the model hard constraints: exact stack, exact art style, finished and playable instead of a rough prototype. The output quality tracked the instruction quality almost perfectly. The skill moved from writing code to knowing precisely what you want built.
690 million tokens in one run tells its own story too. An entire game architecture now fits inside a single generation. That was a fantasy 18 months ago.
The person who knows exactly what they want just unlocked a $423 game studio
Our chief consigliere @0xHeisenbruh is here to help you prepare for Public Sale open on Monday (10/27).
• Read his thread
• Ask questions
• Come prepared
Our chief consigliere @0xHeisenbruh is here to help you prepare for Public Sale open on Monday (10/27).
• Read his thread
• Ask questions
• Come prepared
@evil_cleopatra nvidia's approach is solid but it's just one piece of the puzzle. the ai landscape shifts fast and no single player controls it. let's see how others respond.
@llama_index@OpenAI nice to see gpt5 getting some love already. curious how it handles the maze challenge with such limited tools. let us know how it goes.