Matt Wolfe’s 29-minute roundup covers twelve projects. His hands-on results are mixed, and several entries remain research previews.

1. Simulon — 00:51

  • Simulon places characters in camera footage and generates their reactions after recording.
  • Wolfe’s desk gorilla reacts to his hand and pen; processing happens in the cloud.
  • The current site offers 100 starter credits, Plus at $14.99/month and Pro at $39.99/month.

2. Hy-Nomas — 04:27

  • Tencent’s world model takes successive text instructions and camera moves while retaining video history.
  • Wolfe’s forest, ant and deer trial produces awkward movement and camera control.
  • Its project page advertises 60-second sequences, but says public resource links are still pending, despite Wolfe’s release claim.

3. Reshoot 4D — 08:13

  • Reshoot 4D reconstructs moving footage, adds a virtual camera and renders a new shot through LTX 2.3.
  • The local ComfyUI workflow offers camera paths, lens settings and a browser editor demonstration.
  • Unseen backgrounds can change between angles, and fast motion can smear faces.

4. MoviEdit — 10:26

  • Adobe and KAIST’s research propagates a text instruction or edited reference frame across multiple shots.
  • Its project page shows edits on clips up to 25 seconds, with low-resolution drafts before final renders.
  • The linked paper and benchmark support a research preview; they do not establish a shipping Premiere feature.

5. Dessa — 11:26

  • Block’s speech model learns from audio-only conversation and listens while speaking.
  • The showcased persona invents a job, wife and newborn son while maintaining natural conversational timing.
  • Block offers early access to selected safety researchers; realistic speech increases impersonation and scam risks.

6. Haru 4B — 14:58

  • Higurashi releases weights and inference code for a full-duplex speech base model.
  • User audio, model audio and text advance together at 7.5 steps per second.
  • Wolfe finds the browser demo awkward to test; conversational timing and assistant reliability need separate evaluation.

7. Fish Audio Live 3 — 16:56

  • Live 3 handles interruptions and menu restrictions in Wolfe’s restaurant-order demonstration.
  • Its advertised $0.035/minute rate ends December 31, 2026; premium models and phone calls add charges.
  • This is a hosted voice-agent product, distinct from Haru’s downloadable base model.

8. Subvocal — 18:48

  • Subvocal claims glasses can convert intentional silent mouthing into text at over 200 words per minute.
  • Its site describes personal calibration, but provides no independently verified accuracy figure in the retrieved material.
  • Wolfe withholds a preorder pending outside testing; silent articulation should not be confused with thought reading.

9. Budget3D — 20:54

  • Budget3D trades token budget for generation speed and memory use when producing 3D Gaussian splats.
  • Wolfe’s low-budget examples lose detail; his head example also shows the ambiguity of unseen geometry.
  • The project’s memory tables vary by model and budget, contradicting a blanket claim that every configuration needs under 8GB.

10. MotionBricks — 23:15

  • NVIDIA’s motion framework combines direction, speed and style controls with object-interaction poses.
  • Its authors describe a backbone trained on over 350,000 motion clips, with public models, runtime and authoring tools.
  • An Unreal showcase demonstrates integration work; it does not establish a ready-made plugin for every game project.

11. Flash ALOHA — 24:59

  • Flash ALOHA records demonstrations by having people physically guide backdrivable robot arms.
  • Its learning pipeline smooths actions and masks the operator from training camera views.
  • Researchers report over 80% task success from 50 demonstrations per task; that is a bounded experimental result.

12. PhoneBot — 26:29

  • PhoneBot combines smartphone sensing and policy inference with a separate actuator-control computer and 13 joints.
  • The project estimates $400 in hardware excluding the phone, and demonstrates walking and recovery.
  • Printable-parts and assembly-tutorial links remain inactive; the hardware materials carry a noncommercial licence.

“It’s not happening as much in real time as something like Genie was.” — Wolfe on his Hy-Nomas trial, 06:27

Caveat: these are creator demonstrations and project claims, not a controlled comparison. The source’s headline promises “breakthroughs”; the evidence spans usable tools, research previews and unverified prototypes.