Self-improving AI, Opus 4.8, Nvidia bangers, game-ready 3D models, juggling robots: AI NEWS

This week’s AI news is dominated by a flood of major open-source releases spanning vision, 3D generation, gaming, and robotics. Nvidia stands out with several tools, including Locate Anything for precise object grounding, the PID upscaler for ultra-fast 2K image enhancement, and Gamma World for multi-agent game simulation. Equally notable are Triplat and Cube Part, which push 3D reconstruction and part-aware generation forward, while Google’s relightable characters brings realistic lighting control to captured video. Anthropic’s Opus 4.8 claims benchmark supremacy, though independent tests suggest a closer race with GPT-5.5, and humanoid robots like Astrobot’s T1 and Athena Zero showcase rapid progress in physical AI. Overall, the week underscores a trend toward accessible, code-released AI that runs on consumer hardware, alongside increasing sophistication in simulation-ready content and autonomous research agents.

• ▶ 0:00 The week's roundup highlights many major AI releases, including Anthropic's Opus 4.8, Nvidia's open-source object detector and image upscaler, an AI that builds simulation-ready 3D models, and a Roblox open-source 3D model generator, among others.
• ▶ 1:06 Nvidia released "Locate Anything," an open-source vision-language grounding model that locates and segments requested objects in images/videos using fast parallel box decoding, trained on 785 million bounding boxes and small enough (3B params) to run on consumer GPUs.
• ▶ 2:44 "ControlLight" uses generative AI to brighten or relight images without the noise/artifacts of traditional tools, keeping details and structure stable, with open-source code available and built on Flux 2.
• ▶ 4:03 A new open-source model built on Flux 2 client runs on most consumer devices, with full training code and dataset released.
• ▶ 4:25 Triplat is introduced as a 3D reconstruction AI that turns images into a simulation-ready 3D scene, e.g., letting a robot navigate across it.
• ▶ 4:40 Unlike Gaussian-splat methods, Triplat represents the scene as triangle primitives from the start, skipping mesh conversion; at ▶ 5:20 it is shown reconstructing scenes much faster, with code publicly released.
• ▶ 6:00 Nvidia released PID, an open-source upscaler using a single pixel diffusion decoder to directly output high-res images, avoiding the traditional slow two-step decoding bottleneck.
• ▶ 7:00 PID upscales a 512×512 image to 2K in under one second—about 6× faster than Seed VR2—and wins most benchmark comparisons; it’s already public, works in ComfyUI, and supports Flux 2, Z-Image, and SD3.
• ▶ 8:02 Instruct AV-to-AV edits video and audio together from text prompts, enabling speech changes with matching lip sync, voice alterations, and even gender swaps—but as of the video, its GitHub page is empty with no code or weights released yet.
• ▶ 9:05 Gen Recon converts casual smartphone video or photos into a complete, editable 3D scene, producing a PBR-ready mesh that can be relit, modified, and recolored.
• ▶ 10:28 Scope generates playable first-person shooter worlds from controller actions, responding to moving, aiming, firing, reloading, and more in real time.
• ▶ 11:12 Scope is trained on nearly 70,000 clips from seven FPS games with 10 controller signals, and it outperforms other game generators; its code and dataset are publicly released.
• ▶ 12:13 HubSpot’s free "AI Agents Cheat Sheet" offers a beginner-friendly overview of major agent tools, practical workflows, copy-paste prompts, and the four core traits of true AI agents.
• ▶ 13:36 Fizz X Omni is an open-source, physics-aware 3D generator that creates simulation-ready assets with realistic geometry, materials, joints, and motion—outperforming competitors on benchmarks.
• ▶ 15:09 Deepswe is a new coding-agent benchmark that uses short, realistic prompts across 91 open-source repos and handwritten behavioral verifiers, with GPT-5.5 leading results.
• ▶ 17:11 Anthropic released Opus 4.8, claiming it beats GPT 5.5 on self-reported benchmarks, but independent testing shows a much closer, mixed picture—not a clear win.
• ▶ 19:46 Astrobot unveiled the T1 home humanoid robot, priced around $13,000, capable of chores like laundry, ironing, and bartending, though it uses a wheeled base and cannot climb stairs.
• ▶ 20:53 The Rye Institute's Athena Zero robot learned complex juggling in under 10 minutes of real-world interaction, adapting among five juggling styles on the fly—showcasing advanced real-time spatial coordination.
• ▶ 23:33 Cube Part enables part-aware 3D generation, producing separate meshes for each object part (e.g., a car's wheels, body, doors) that assemble into a coherent, animatable object; code is released and it runs on most consumer GPUs.
• ▶ 25:21 Google's relightable characters method captures a moving person with multiple cameras and inserts them into new scenes with controllable, realistic lighting, though code is coming soon.
• ▶ 26:13 BEES improves language models using bidirectional evolutionary search—combining forward solution-building and backward subgoal decomposition—delivering gains on post-training tasks and open problem-solving benchmarks, with code released.

• ▶ 27:46 Autoscientist is introduced as an agentic framework for automating scientific research, distinct from earlier Nature-published agents by organizing AI agents into parallel, self-improving research teams.

• ▶ 28:17 It functions like a decentralized research lab: agents share a project state (current best solution, experiment log, forum, and deadend registries) and iterate on reading, deciding, acting, and writing results—enabling them to navigate the messiness of real science.

• ▶ 29:53 On the BIOML Bench (24 biomedical ML tasks), Autoscientist beat other agentic frameworks, and the code is already released for local setup via the project page.

• ▶ 30:25 Nvidia Gamma World generates simulations of multiple agents playing the same game simultaneously, unlike typical single-player world models.
• ▶ 31:02 Uses simplex rotary agent encoding to keep agents independent while maintaining consistency and allowing a variable number of agents.
• ▶ 31:29 A GitHub repository is released, with code, training scripts, and a dataset preparation tool coming soon—suggesting Nvidia plans to fully open source it.
• ▶ 31:44 Pantheon 360 generates 360° videos from multiple panoramic images and a camera path, enabling consistent digital twin scenes.
• ▶ 32:05 Unlike normal video generators with narrow field of view, it avoids stitching drift by reconstructing a 3D point cloud to keep video stable and grounded.
• ▶ 32:39 Key applications include digital twins, robot training, and autonomous driving, with code expected to be open-sourced soon.
• ▶ 33:00 Bonsai Image is a new image generator that runs locally and offline on a phone.
• ▶ 33:11 Two model variants were released (one-bit and ternary), both based on Flux 2 Klein and compressed from nearly 8 GB to roughly 1 GB.
• ▶ 33:59 A real-world demo on an iPhone 17 Pro Max generated a 512x512 image in just 9.4 seconds.
• ▶ 34:25 Introduces Mini CPM51B, a tiny 1-billion-parameter model from OpenBM designed for local deployment.
• ▶ 34:42 Despite its small size, it outperforms similar-sized models across general knowledge, coding, math, logical reasoning, and agentic tasks.
• ▶ 35:00 Its 2 GB footprint allows it to run on most laptops and potentially mobile phones, with setup instructions and platform support provided in the description.
• ▶ 35:16 Sega is introduced as a new method for generating high-resolution images, shown via sample generations.
• ▶ 35:24 Sega produces extremely high-resolution, sharp outputs, including over 4K images with fine details like water droplets and whiskers, with minimal flaws even when zoomed in.
• ▶ 35:49 Sega is flexible across base models, supporting both Flux and Qwen, with Qwen Image examples demonstrated at high detail.
• ▶ 35:59 Sega is an AI image upscaler that produces "really sharp and accurate" scenes, with output resolutions up to 6,144 pixels per side.
• ▶ 36:09 Sega is more consistent and less error-prone than the DYP upscaler, making it currently one of the best methods for high-quality image generation.
• ▶ 36:24 The code has been released on the project page, with instructions for running Flux one or Quinn image, but there is no indication yet of support for Flux 2 or Z image.
• ▶ 36:49 Users can interactively relight a single photo by dragging the cursor to control the light source from any angle, with options to adjust light types like a narrow spotlight.
• ▶ 37:04 The system understands the 3D layout of a 2D scene, detecting objects such as tables and chairs and relighting them realistically even in cluttered, messy scenes.
• ▶ 37:41 The workflow estimates a rough 3D model of the input image, plugs it into Blender for user-controlled relighting, and then uses the modified scene to generate the final relit output; code is released at ▶ 37:57.
• ▶ 38:15 Viewers are invited to comment with their opinions, favorite news, and the tool they're most excited to try.
• ▶ 38:28 The host calls on viewers to like, share, and subscribe, and to stay tuned for more content.
• ▶ 38:33 To keep up with fast-paced AI developments, viewers are directed to subscribe to the free weekly newsletter via the link in the description.

• ▶ 0:00 Weekly Roundup Intro & Image Tools (0:00 - 4:05) - - Weekly intro, Nvidia's Locate Anything grounding model, and ControlLight image brightening/editing.
• ▶ 4:05 Open-Source Models & 3D Reconstruction (4:05 - 6:00) - - A consumer-device open-source model and Triplat simulation-ready 3D reconstruction from images.
• ▶ 6:00 Upscaling & Video/Audio Editing (6:00 - 9:05) - - Nvidia's PID high-res upscaler plus Instruct AV-to-AV and lip-sync audio/video editing.
• ▶ 9:05 VR, Game Worlds & Datasets (9:05 - 12:13) - - Gen Recon for real-estate VR, Scope for playable FPS worlds, and a controller-signals dataset.
• ▶ 12:13 AI Agents, Benchmarks & 3D Generation (12:13 - 17:11) - - HubSpot's AI agents cheat sheet, Fizz X Omni 3D generation, and Deepswe/Deep Sweep benchmarks.
• ▶ 17:11 Model Releases & Humanoid Robots (17:11 - 23:33) - - Anthropic Opus 4.8, affordable humanoid robots, and Step 3.7 Flash agentic model.
• ▶ 23:33 3D Generation & Self-Improving AI (23:33 - 27:46) - - Part-based 3D generation, Google's relightable characters, and BEES self-improving language models.
• ▶ 27:46 Scientific AI, World Models & Local Tools (27:46 - 38:50) - - Agentic science, world models, camera-stable 360° video, local Bonsai/Mini CPM, Sega, and Pixe.