New AI enzymes, 700+ math breakthroughs, AI for DNA mutations, room temp semiconductors: AI NEWS

Summary This video is a weekly AI news roundup presented by the YouTube channel AI Search. The host reviews major model releases, research preprints, and open-source tools across scientific AI, open models, robotics, and mathematics.

What is shown

• WorldPlay2 [01:08]: Interactive demonstration of real-time world generation based on Wan 2.2, showing keyboard-controlled navigation (walking, running, driving) and interaction events (petting a dog, turning on a stove).
• MotionSpaceFlow (MSFlow) [02:25]: Text-prompted 3D skeleton animation generator with spatial constraints on limbs.
• Kandinsky 6.0 Video [03:30]: Video clips with synchronized 44 kHz audio, including animal performances and cinematic clips, plus a demonstration of its super-resolution pipeline.
• ChatGPT Intelligent UI [04:55]: Interactive UI generation in ChatGPT for interactive explanations (interactive bicycle anatomy diagram, Monty Hall interactive widget, dynamic Sunday roast planner with adjustable servings checklist).
• ALoDLM (Adaptively Looped Diffusion Language Models) [06:23]: Benchmark tables and code implementation showing token generation refinement over recurrent loops.
• AlphaProtein Novo [08:04]: Protein structure generation pipeline combining diffusion with AlphaFold 3 and LigandMPNN to design functional enzymes from scratch for piperidine synthesis and DEHP breakdown.
• BOTANIC-1 & Gemma 4 [10:22]: Bioinformatics workflow mapping melon plant DNA mutations causing flower color alterations.
• OpenAI Mathematics Manuscript Collection [12:21, 18:50]: Walkthrough of OpenAI’s repository of 719 manuscripts across 372 families, detailing proofs for the Quasi-Riemann Hypothesis, Hilbert’s tenth problem over $\mathbb{Q}$, the matrix multiplication exponent ($\omega \le 2.25$), and Birch and Swinnerton-Dyer formula extensions.
• Higgsfield AI Influencer 2.0 & Genjutsu [17:18]: Custom character generation interface and motion-transfer demo across video clips.
• EmbeddingGemma 2 [26:37]: Multimodal embedding benchmarks across text, image, audio, and video on mobile devices.
• Mistral Large 4 (“Le chonk”) [28:40]: Model architecture overview and Artificial Analysis intelligence/cost benchmark charts.
• GenIA & CoDimRecon & DistScene [30:08, 39:35, 40:36]: 3D scene and object reconstruction pipelines from single/multi-view images.
• Room-Temperature Antiferromagnetic Semiconductors [30:52]: Solid-state candidate analysis designed and screened by Claude Opus 5.5 agents.
• TERRA & FlashDexRetarget [35:34, 38:10]: Musculoskeletal whole-body locomotion simulation over obstacles and dexterous robotic hand retargeting.
• Iris-3B & Nano Banana 2.1 [36:35, 41:27]: Diffusion without VAE (Iris-3B) and Google’s Nano Banana 2.1 generation examples.
• Claude Haiku 5.5 [42:40]: Benchmark comparisons against Haiku 4.5, GPT-6 Luna, and Sonnet 5.5 on coding and knowledge tasks.

Claims & numbers

• WorldPlay2: Operates in real time at 16 FPS; includes a 4-step turbo model checkpoint weighing 69.8 GB based on Wan 2.2 [01:16, 02:11].
• Kandinsky 6.0 Video: Generates 5-second video clips with synchronized 44 kHz audio; released in Lite (3.8B parameters, 6.38 GB) and Pro (29B parameters, 60.3 GB), with up to 4× latent upscaling [03:47, 04:22, 04:41].
• ALoDLM: Amazon’s model delivers +8.5% throughput and requires 13.6% less compute compared to baseline diffusion language models; available in 1.7B (4.08 GB) and 8B (16.4 GB) sizes [07:27, 07:46].
• Google DeepMind AlphaProtein Novo: Generates de novo catalytic motifs confirmed in wet-lab experiments, achieving a 100% hit rate in round 2 for piperidine transferase activity [09:18].
• BOTANIC-1: Foundation model pretrained across 320 plant species; achieved a 90% top-ranked success rate pinpointing the causal variant out of 2,494 single-nucleotide mutation candidates in melon plants [11:17, 11:52].
• OpenAI Mathematics Release: 719 manuscripts covering 372 result families; 162 are Lean-verified (approx. 22.5%) [13:24]; claims bounding zeta zeros to $\text{Re}(s) \le 7/8$, disproving a universal algorithm for Hilbert’s 10th problem over $\mathbb{Q}$, and lowering the matrix multiplication exponent bound to 2.25 [14:24, 20:13, 22:33].
• EmbeddingGemma 2: Requires 270M parameters for text, with optional 170M vision and 300M audio modules (total size ~1.53 GB, runs in ~191 MB active RAM on Google Pixel 11 Pro) [27:22, 28:28].
• Mistral Large 4: 1 trillion total parameters with 52 billion active parameters in an MoE architecture with a 1M token context window [28:55].
• Claude Haiku 5.5: Priced at $0.10 per million input tokens and $0.50 per million output tokens, scoring 72.4% on OSWorld 2.1 and 46.4% on Agentic Coding (FrontierCode 11) [29:53, 42:40].

Notable quotes

• [00:27] “This is like the biggest day in the history of mathematics.”
• [08:17] “This is an AI that can basically design entirely new enzymes from scratch.”
• [26:13] “The bottleneck right now has flipped to where we are no longer limited by compute or the intelligence of AI, but rather by the capacity of us humans to actually process all of these discoveries.”

Assessment This is an independent weekly roundup and analysis video summarizing published open-source releases, research preprints, and company announcements. While code repositories, project pages, and paper benchmarks are accurately showcased on screen, the presenter notes that many of the highlighted scientific and mathematical claims (such as OpenAI’s unverified proof preprints and the semiconductor crystal candidates) remain theoretical or preprints pending independent experimental and peer verification.

Described by gemini-3.8-flash on 2026-10-19 from the video’s audio and frames.