Grok 4.6 is here, and it builds games, 3D worlds, and trailers
A two-word post from Elon Musk — "Grok 4.6" — officially teases xAI's next-generation model. The hands-on clips that followed point to a steep jump in multimodal generation.
Musk's announcement was characteristically terse. The entire reveal was a single phrase — "Grok 4.6" — posted with no chart, no benchmark table, and no launch blog. In the hours that followed, the community filled in the blanks. Creators began sharing what the model can generate: playable game prototypes, explorable 3D worlds, cinematic game trailers, and even instructions for real-world 3D prints. The through-line is a shift away from text and still images and toward interactive, spatial output — assets a user can walk around inside rather than merely view. If the demos hold up, Grok 4.6 marks the point where xAI's flagship moves from answering questions to producing worlds.
Qwen3.8-27B flies on a laptop
Alibaba's Qwen team says the 27-billion-parameter model runs comfortably on a laptop and is already blending into everyday work. Apps like Atomic Chat have adopted it, pushing an open model out of the datacenter and onto local hardware where developers control their own inference.
GLM 5.3 aces security evals, deepsec cuts review costs 3×
rauchg reports that evals of GLM 5.3's cybersecurity capabilities place it at the new open frontier. Because its costs run lower, the full-repository security scanner deepsec can be run at least three times more often — a direct win for defensive teams that previously had to ration deep scans.
intelligence wants to be free
— hardmaru
Qwen3.8-27B tops Hugging Face's trending chart
Within hours of release, Qwen3.8-27B became the most popular model on Hugging Face. Alibaba's team thanked the community — the announcement drew thousands of likes — and invited developers to try the model and report back. It is a sign of how quickly a laptop-scale open model can become the default choice for tinkerers.
Intelligence per joule jumps 18× in 16 months
Replit CEO Amjad Masad shared a chart showing that the intelligence delivered per joule of energy has improved eighteen-fold in sixteen months. It is a reminder that progress in AI is measured not only in benchmark scores, but in how efficiently a model converts electricity into useful reasoning — and that this curve, more than raw scale, is what quietly rewrites the economics of the field.
125 years of super-exponential compute
Masad argues that the claim AI "structurally centralizes power" ignores more than a century of super-exponential gains in compute price-performance. Algorithms and hardware efficiency keep improving, he says, so there is no reason to assume concentration is permanent.
Qwen downloads pass 3 billion
The Qwen series has crossed three billion cumulative community downloads, Alibaba said while thanking developers.
Games, 3D worlds, and trailers
Musk amplified a demo calling Grok 4.6 "scary good" at building games, 3D worlds, trailers, and real 3D prints.
Qwen 3.8 27B is excellent — but it overthinks
Simon Willison tested the 27B vision model on a MacBook Pro and DGX Spark, praising it while flagging a default "xhigh" reasoning mode that can burn tens of thousands of tokens.
Grok Harvey + Cursor Ultra for $99 a month
A community tip maps out how to bundle Grok Harvey, Cursor Ultra, and Grok Bot quotas under a single $99 subscription.
CodePilot adds Grok 4.6, DeepSeek V4 Pro, GLM 5.3
The coding assistant now supports the latest models and lets users spend Grok image and video quotas inside the editor.
Enable Codex's 1M-token context
A config.toml tweak unlocks a million-token window, though dotey argues shorter, well-compressed context is often faster and cheaper.
An open tool strips AI watermarks
watermarks-remover removes C2PA metadata and multi-vendor provenance marks from images, documents, and code.
ChatGPT can now open PRs directly
Pointing ChatGPT at a GitHub repo lets it clone, edit code, and submit a pull request without manual uploads.
Auto-resume past the 5-hour limit
Claude Desktop now continues main and sub-agents when the session limit hits, reusing existing sub-agents automatically.
GPT-5.6 Sol exported 5,302 bookmarks
It took over Chrome and rescued years of X bookmarks back to 2014.
Benchmarks need human judgment
For non-verifiable domains, human opinion is the benchmark — AI folks should learn qualitative methods.
AI Overview is reshaping the web
Google's summaries are already changing what information people consume and act on.
Old models still write exam questions
o3-mini in an agentic loop produced questions on par with standardized tests.
OLMo 1-3 remains underrated
Its contributions to open science deserve more attention, the author says.
Spec trailers are now effortless
AI makes producing polished game trailers dramatically easier.
AI adds flavor to every project
levelsio says AI lets people add personality to work that would have been impossible to hand-build five years ago.
A blind robot that still performs
New footage shows a visionless robot completing tasks, drawing "end game" reactions.
A UI plugin for DeepSeek Harness
A community project is beautifying the interface and bundling service-provider plugins into a client.
Hyderabad AI meetup
soumithchintala is collecting names of local deep-tech, AI-product, and GPU experts for a gathering next week.
A visual guide to the $99 bundle
A step-by-step graphic explains how to claim Grok Harvey, Cursor Ultra, and Twitter Premium+ together.
"Code is truth"
The two authors of Pi argue code needs no memory system or RAG, and that Bash plus scripts beat MCP for most tasks.
Fable 5 struggles to break its frame
Given only an optimization goal, even strong models optimize within the existing framework rather than finding a new path.
How Claude's text watermark works
Anthropic's note traces Claude's text watermarking to Google DeepMind's SynthID-Text, published in Nature in 2024.