Cool figure of #Gemini3 with #DeepThink on ARC-AGI-2. The big jump looks quite like the results on #IMOBench that we obtained previously with Gemini 2.5
MACHINE LEARNING
-
Astonishing LLM progress: opportunity landscape changes in 6 months
By
–
LLM progress keeps its astonishing pace! I'm reminded of Adobe founder Charles Gerschke saying "shoot where the duck is going to be"
In 6 months, thanks to AI, the opportunity landscape will be vastly different from today. -
Google launches free Antigravity coding IDE
By
–
BREAKING 🚨: Google is launching Antigravity, a free vibe coding IDE.
— 🚨 AI News | TestingCatalog (@testingcatalog) 18 novembre 2025
– Agent model: access to Gemini 3 Pro, Claude Sonnet 4.5, GPT-OSS
– Unlimited Tab completions
– Unlimited Command requests
– Generous rate limits *
Free testing time 👀 https://t.co/sq3xsRhu7o pic.twitter.com/u7cDZ5lyh5BREAKING : Google is launching Antigravity, a free vibe coding IDE. – Agent model: access to Gemini 3 Pro, Claude Sonnet 4.5, GPT-OSS
– Unlimited Tab completions
– Unlimited Command requests
– Generous rate limits * Free testing time -
Gemini 3.0 just one-shotted Apple website clone
By
–
Gemini 3.0 just one-shotted this Apple website clone. pic.twitter.com/tkGkryMh9Y
— God of Prompt (@godofprompt) 18 novembre 2025Gemini 3.0 just one-shotted this Apple website clone.
-
DA3 as feed-forward 3D Gaussian splatting outperforms specialized models
By
–
And then the wild part DA3 becomes a feed-forward 3D Gaussian splatting engine. Add one head.
No optimization loops.
Just straight image → 3D Gaussians → novel views. DA3-based 3DGS outperforms every specialized model. -
Teacher-student training with RANSAC turns bad data into high-quality supervision
By
–
The teacher–student training strategy is doing heavy lifting. They train a synthetic-only teacher, align it with real-world noisy depth using RANSAC, and suddenly all these “bad datasets” become high-quality supervision signals. That’s how DA3 learns sharp, dense geometry
-
DA3 beats VGGT and Pi3 across all geometry benchmarks
By
–
The numbers are insane. DA3 beats VGGT and Pi3 across every geometry benchmark: → HiRoom
→ ETH3D
→ DTU
→ 7Scenes
→ ScanNet++ Pose accuracy. Reconstruction accuracy. Depth consistency.
Across all 5, DA3 sits at the top. -
DA3’s universality: one model handles depth, geometry, pose, and splatting
By
–
DA3’s biggest flex isn’t accuracy, it’s universality. One model handles monocular depth, multi-view geometry, pose estimation, and even 3D Gaussian splatting without changing the backbone. Just adaptive cross-view attention + a dual-DPT head. This is what “simple architecture,
-
Single Vanilla Transformer Outperforms All SOTA 3D Models
By
–
Holy shit… a single vanilla transformer is outperforming every SOTA 3D model. Depth Anything 3 just rewired 3D perception. A single plain transformer no fancy architecture now reconstructs full 3D geometry from any set of images. One photo, 18 photos, posed or unposed, it
