A video of a robot catching a ball tells you almost nothing about whether you can build one. A voice assistant that takes an interruption in stride tells you more than a polished audio sample. Matt Wolfe's roundup is most interesting where he moves beyond the showreels and tries things himself: a generated deer looks wrong, a 3D render gets the front but guesses the back, and a voice agent holds up when he goes off-menu.
Here are all 12 projects, with the original links and a distinction that matters: a demo you can watch, a model you can download, and a tool you can use today are three different things.
Try Fish Audio Live 3 for a live conversation or the Budget3D browser demo for one-image 3D. Simulon has an iPhone app; Haru 4B has downloadable weights. Treat Subvocal's glasses and the two physical robot demos as projects to follow, not products to count on.
Watch Matt Wolfe's Roundup
Source Note: The hands-on observations below are Wolfe's, from his video and its transcript. Access and technical details were checked against the linked project pages, papers and repositories on 19 November 2026. A project team's demo is not an independent test.
Characters, Worlds and Video
1. Simulon: film a character that reacts to your room
Access: iPhone app. 00:51 in the video · Official app and examples.
Wolfe positions a digital character in the camera preview, then records himself interacting with it using his hand and a pen. The distinction is easy to miss: you place the character live, but the reactive footage is rendered in the cloud after recording. Simulon's staged launch clips show the creative goal; Wolfe's clip gives a better feel for the capture-and-wait workflow.
2. Hy-Nomas: give a generated world another instruction
Access: browser demo, code and weights. 04:27 in the video · Tencent project · GitHub.
Tencent's model extends a scene in segments. You can direct the next segment with text or a camera move and the model draws on earlier footage. Wolfe starts with an ant in a forest, adds a deer and changes the viewpoint. The clip picks up where it left off, but the deer looks misshapen and the camera controls feel awkward. That's enough for experimenting with a shot, not evidence that the scene obeys game-engine rules or responds instantly on your hardware.
3. Reshoot 4D: move the camera after filming
Access: public editor and a local GPU workflow. 08:13 in the video · Examples and editor · Code and requirements.
Set a new virtual camera path for existing footage, then render a view from that path. The project offers a browser editor, while its documented rendering workflow uses ComfyUI and a compatible GPU. The model has to invent what the original camera never saw. That's useful for a creative alternate shot, but inspect faces in motion, backgrounds and newly exposed surfaces before trusting a take.
4. MoviEdit: make one change across several shots
Access: research examples, paper and benchmark. 10:26 in the video · Adobe and KAIST project.
Reshoot changes where the camera appears to be. MoviEdit changes what's in the shot, then carries the change across cuts. The team shows edits driven by a text instruction or by one or more edited reference frames, on clips up to 25 seconds. That's useful when an object needs to look the same from several angles, but the research page is not a release announcement for a Premiere feature or a full-length editing tool.
Voice and Silent Speech
5. Dessa: converse without taking strict turns
Access: public audio samples and a waitlist. 11:26 in the video · Block's research page.
Block says it trained Dessa on conversational audio without starting from a text model. What stands out is the timing: the system can listen while speaking, rather than waiting for a clean handoff. Wolfe also plays a clip in which the voice invents a backstory. Natural pauses and a convincing personality don't make that story true.
6. Haru 4B: downloadable conversational voice
Access: public weights and inference code. 14:58 in the video · Higurashi project and samples · Model card.
Haru is the more hands-on option if you're interested in running a conversational voice model yourself. It's a pretrained base model, not a finished voice assistant. Experiment with the cadence and interaction, but don't assume it will follow policy, look up facts reliably, or handle a support call without oversight. Check the model card for setup and terms.
7. Fish Audio Live 3: test a voice agent in the browser
Access: hosted demo and API. 16:56 in the video · Try the voice demos.
Wolfe speaks to a restaurant demo, interrupts it and requests things that aren't on its menu. The voice recovers and sticks to its listed options. That's a more useful test than listening to a perfect script. It shows how this configured demo handles a detour, not that an integration can take real orders or deal with every accent and background noise.
8. Subvocal: glasses for words you mouth silently
Access: prototype and preorder information. 18:48 in the video · Company site · Launch update.
The company says its glasses use radio-frequency sensing to pick up small facial movements as you deliberately mouth words. That's silent articulation, not thought reading. The appealing use case is dictating in a quiet shared space; the open questions are accuracy across different wearers, correction effort and how the device handles sensor data. Wolfe says he'd wait for independent testing before putting money down.
3D, Animation and Robots
9. Budget3D: spend less compute where the detail is sparse
Access: community browser demo, code and weights. 20:54 in the video · Project · GitHub · Try it.
One photo goes in and a 3D Gaussian splat comes out. A smaller token budget can save time and memory at the expense of detail. Wolfe compares budgets and rotates his generated object: the front looks plausible, but the back of the head is flat. Use it as a visual draft, not as an accurate scan or an animation-ready mesh.
10. NVIDIA MotionBricks: animate the transitions, too
Access: research examples and public code. 23:15 in the video · NVIDIA demos · Whole-body control repository.
A character doesn't just have to walk and grab an object. It has to slow down, put its feet in the right place and reach without an abrupt jump. MotionBricks shows how generated motion can connect those actions and respond to changes in speed, style or target pose. Convincing motion in a rendered character doesn't establish that the same action is safe on physical hardware.
11. Flash ALOHA: learn quick movements from demonstrations
Access: research project, paper and code. 24:59 in the video · Project, videos and resources.
The Stanford team demonstrates a two-arm robot catching a ball and clearing a table after learning from physically guided demonstrations. The team reports over 80% success on its tested tasks using 50 demonstrations per task. That's a result from a particular lab setup, not an estimate of how often the system would succeed in a kitchen with different objects and lighting.
12. PhoneBot: give an old smartphone legs
Access: research prototype with partial build resources. 26:29 in the video · Hardware and demos · Paper.
The phone supplies a camera, motion sensors and computing for perception and walking; a separate board talks to the motors. The project estimates about $400 for hardware, excluding the phone. The phone is not doing everything: the paper describes external compute for the demonstrated conversation pipeline. The site configures links for the parts list and CAD, but the printable-parts bundle and assembly tutorial were not linked at review. Check for a complete software release before buying parts, and note that the published mechanical materials carry a noncommercial license.
Two Demos to Try First
Test recovery in a voice call. In Fish Audio's restaurant demo, change your mind halfway through an order, ask for an unavailable item, and see whether the agent remembers the rest. Use fictional details rather than real customer data.
Look at the back of a generated asset. Give Budget3D one photo you own, run it at two budgets, and rotate both outputs. If all you need is a front-facing prop for a storyboard, the faster result might be fine. If the back needs to match a real product, capture more views instead of paying for a more convincing guess.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | AI breakthroughs you probably missed |
| 00:51 | Simulon: AI characters in your room |
| 04:27 | Tencent Hy-Nomas: direct a generated world |
| 08:13 | Reshoot 4D: move the camera after filming |
| 10:26 | MoviEdit: carry an edit across cuts |
| 11:26 | Dessa: a voice model that can listen while talking |
| 14:58 | Haru 4B: downloadable conversational voice |
| 16:56 | Fish Audio Live 3: hosted voice agents |
| 18:48 | Subvocal: glasses for silent articulation |
| 20:54 | Budget3D: turn one image into a 3D asset |
| 23:15 | NVIDIA MotionBricks: character animation |
| 24:59 | Flash ALOHA: robots that catch and clear |
| 26:29 | PhoneBot: a walking robot with a phone brain |