Riemann solved, unforced Navier-Stokes, realtime waifus, new GPT, Gemini Neon: AI NEWS
AI Search · 2026-11-22
💡 Quick Take
1. LongCat Video Avatar 1.6 turns a photo and audio into a realistic talking video over 10 minutes long without quality loss, and the code is available on GitHub.
2. Fab reconstructs a person and environment in 3D from video, transferring human movement to UniTree G1 robots, and is available in sizes under 4 GB.
3. Apple's Depth Pro 2 estimates scene depth from images or video with sharp edges and consistent frames, but weights are currently restricted to non-commercial research.
4. Levent from Anthropic used Claude to write 17 million lines of lean code claiming a proof for the Riemann hypothesis, though the math community is still verifying it.
5. Duo is a real-time conversational model that listens and speaks simultaneously, requiring high-end hardware due to a 47 GB quantized download size.
📊 Detailed Explanation
Several major video and 3D generation tools launched this week, giving creators new ways to handle avatars and spatial computing. LongCat Video Avatar 1.6 takes a static photograph and an audio file to generate speaking videos with expressive faces, hand movements, and consistent quality exceeding ten minutes. Alongside it, Fab reconstructs full 3D bodies and surrounding environments from regular video feeds, even translating human motions like dancing or walking directly into animations for humanoid robots like the UniTree G1. Fab offers three versions scaling up to 3.8 gigabytes, making local execution feasible on consumer hardware. Meanwhile, Apple released Depth Pro 2 for estimating scene depth across video frames with high temporal consistency, and Mave converts ordinary videos into editable 3D Blender scenes, though Mave's dataset and code are slated for future release.
In the realm of foundational mathematics, researchers announced potential solutions to two major Millennium Prize problems using AI assistance. A mathematician at Anthropic named Levent utilized Claude to build a 17-million-line codebase in lean, a proof-checking framework, to tackle the Riemann hypothesis relating prime numbers to the zeta function. While independent mathematicians like Andrew Sutherland verified that the code compiles successfully, the broader academic community continues scrutinizing the dense proof. In a parallel breakthrough, Levent and Tristan Buckmaster tackled the unforced Navier-Stokes equations, presenting an 80-page paper and 600,000 lines of lean code proving that fluid turbulence or finite-time blowup can occur naturally without an external force. However, the critical verification data certificate has not been released yet, meaning independent confirmation remains pending.
Real-time interaction models and audio processors also saw broad updates aimed at low latency and multimodal understanding. Duo provides continuous two-way audio and video conversations where users can interrupt the model mid-sentence, though running it locally requires heavy computing resources due to its 47 gigabyte quantized footprint. For audio cleanup, a new Voice Isolator model removes background noise and competing voices while maintaining the primary speaker, offering a real-time streaming version with a processing latency of just 63 milliseconds under a permissive Apache 2 license. Additionally, new image tools like Cling Image 4 debuted online with support for 8K resolution, text rendering up to 2,000 characters, and multi-image storytelling features.
Frontier labs and open-source communities pushed out several notable model architectures and reasoning frameworks. OpenAI released GPT 6.2 Astra and Soul, which the host found somewhat underwhelming after noticing performance regressions compared to earlier iterations. Google introduced Gemini Neon as a mid-tier efficiency model sitting between Carbon and Hydrogen, which punches above its weight class in select benchmarks. In open weights, Tencent launched High-4 Flash, a massive 224 billion parameter mixture-of-experts model with only 12 billion active parameters designed for cost-efficient enterprise scale. Smaller open options like Sense Nova Loop 9B use a shared loop architecture to let data pass through internal layers multiple times, punching well above its 9 billion parameter class while fitting into standard consumer hardware under 5 gigabytes.
🎯 Tech Expert Opinion
The week's developments highlight a fascinating shift where artificial intelligence is moving past simple pattern matching and entering rigorous scientific verification and complex spatial computing. The claims surrounding the Riemann hypothesis and unforced Navier-Stokes proofs using Claude and lean codebases represent a massive leap forward. Even though formal academic peer review takes time, the ability of AI-assisted systems to write millions of lines of verifiable proof code signals a permanent change in how mathematical research operates.
At the same time, the release of practical tools like Fab, Depth Pro 2, and LongCat Video Avatar shows that heavy computational graphics tasks are rapidly decentralizing to local consumer hardware. The host points out that models under 4 gigabytes can now handle complex 3D body tracking and depth estimation. This lowers the barrier for independent developers who want to experiment with spatial video editing, robotic imitation learning, and synthetic data generation without relying exclusively on expensive cloud APIs.
However, users must approach several of these releases with measured caution. Commercial restrictions apply to high-end weights like Apple's Depth Pro 2 and RMBBG 3.0, and massive architectures like Duo and Tencent's High-4 Flash demand enterprise-grade infrastructure to run locally. Furthermore, the speaker's lukewarm take on OpenAI's GPT 6.2 update reminds us that larger version numbers do not always translate to linear performance gains, making it wise to benchmark new closed models thoroughly before migrating production workflows.
Kanal: AI Search