Quick answer
Twelve AI breakthroughs nobody is talking about, rounded up by Matt Wolfe, range from apps you can use on your phone today to research demos nobody outside the lab has tested. You can already try Simulon (3D characters in your real room), Fish Audio Live 3 (paid AI voice agents), Budget3D (one image to a 3D model) and several free, open models. Others, like the eerily human Dessa voice model and Subvocal’s silent-speech glasses, are still previews or unproven claims.
Key takeaways
- Try now: Simulon (free starter credits, then about $15 a month), Fish Audio Live 3 ($5 free credit), and free downloads such as Tencent’s HY-Nomas world model, Reshoot 4D, Haru 4B, Budget3D and MotionBricks.
- Voice AI just got very human. Dessa, from a lab inside Block, learned only from recordings of people talking, so it chats like a real person and even invents a life story. That also makes it a scam risk.
- Watch, don’t buy yet: Subvocal claims 200+ words a minute from silent mouthing, but published research on this method tops out at a 50-word vocabulary and nobody independent has tested it.
- Robots are getting cheaper: PhoneBot uses an old Android phone as its brain and about $400 of parts.
- Many demos are rough. Matt’s own tests show wonky results alongside the impressive clips.
In this guide
- Which of these AI breakthroughs can you try today?
- What are Simulon and Tencent’s new world model?
- How is AI changing video editing?
- Why do the new AI voice models sound so human?
- Do Subvocal’s silent-speech glasses really work?
- What can Budget3D and NVIDIA MotionBricks do?
- What are Flash ALOHA and PhoneBot?
- What do these AI breakthroughs mean for you?
- Watch: the source video
- Frequently asked questions
- Related posts
Every Friday Matt Wolfe covers the week’s biggest AI news. This video is the other stuff: tools, research projects and demos from the past month that flew under the radar. We have grouped them by what they do and, most importantly, by whether you can actually use them yet.
Which of these AI breakthroughs can you try today?
| Breakthrough | What it does | Can you try it? |
|---|---|---|
| Simulon | Puts 3D characters into your real room through your phone camera; they react to you | Yes, phone app (free credits, then about $15/month) |
| HY-Nomas (Tencent) | A world model you direct by talking, scene by scene | Yes, free and open, commercial use allowed |
| Reshoot 4D | Re-films your existing video from a new camera angle | Free and open source, but needs an NVIDIA GPU and ComfyUI |
| MoviEdit (Adobe research) | Edits a change across every shot in a scene | No, research paper only |
| Dessa (Block) | Very human-sounding voice conversations | No, safety researchers only |
| Haru 4B (Higurashi) | Downloadable voice model that listens and talks at once | Yes, free download (a base model for developers) |
| Fish Audio Live 3 | AI voice agents for support calls and bookings | Yes, paid API with a $5 free credit |
| Subvocal | Glasses that turn silently mouthed words into text | Pre-order only ($199 deposit), unproven |
| Budget3D | Turns one image into a 3D model, with a speed/quality slider | Yes, free demo on Hugging Face |
| MotionBricks (NVIDIA) | Real-time character animation without hand-made clips | Code and models on GitHub; not in Unreal Engine yet |
| Flash ALOHA (Stanford) | Robot arms that learn fast tasks like catching a ball | Code on GitHub, you build the hardware |
| PhoneBot (UCLA) | Walking robot with an Android phone as its brain | Parts list out; some build files not yet available |
What are Simulon and Tencent’s new world model?
Simulon is a phone app that drops 3D characters into your real surroundings. Its new feature lets you place a character, hit record, act out a scene, and the character reacts to what you do and say. It is not live: when you stop recording, your clip goes to the cloud and the whole shot is redrawn with the character in it. In Matt’s test, a small gorilla on his desk reacted when he “petted” it and swatted away a pen. The reacting feature is iPhone only for now, with Android promised. You get 100 starter credits, which go quickly, then it is about $15 a month.
Tencent’s HY-Nomas goes further: it generates the whole world, and you steer it as you go with directions like “add a dog”, “make it snow” or “swing the camera left”. It builds video in five-second pieces and looks back at everything it has already made, so the world stays consistent for up to about a minute. Matt’s test (a tiny ant in a futuristic forest) was, in his words, a bit wonky. It is free, open and allowed for commercial use, but the real-time setup Tencent tested used eight data-centre GPUs, and the consistency claims come from Tencent’s own demos.
💡 Tip: A realistic use today: planning a shoot. Talk your way through a scene in a world model to sketch shots before you film anything. The quality is not good enough for finished work yet.
How is AI changing video editing?
Reshoot 4D lets you re-film a video you have already shot from a new angle: orbit around a person, push in, crane up or do a “bullet time” move. It builds a 3D capture of the scene over time, lets you place a virtual camera, re-renders the shot and then uses the open LTX video model to clean up the result. It is free and open source, built by a single developer who says he made it with Claude Code and Codex. The catches: it does not invent what was never on camera, fast motion can smear faces, and you need an NVIDIA graphics card plus some ComfyUI know-how.
MoviEdit, a research project from Adobe and a university in Korea, tackles a classic AI video problem: change something in one shot and it reappears in the next. Type one instruction, such as removing a character, and it applies across every shot in a scene, on clips up to 25 seconds. You can also edit one frame in Photoshop and it carries the change through. For now it is only a paper and sample videos, with no word on whether it will reach Premiere.
Making videos for your business? Adobe is already building AI agents into Premiere and Firefly. See what is available now in our round-up of Adobe MAX AI agents.
Why do the new AI voice models sound so human?
Three voice models show how fast this area is moving, in about three weeks:
- Dessa, from a small research lab inside Block (Jack Dorsey’s company), was trained from scratch on recordings of people talking, with no text and no chatbot underneath. Most voice AI turns speech into text, runs it through a language model and turns the answer back into speech. Dessa skips all that, so it never learned to be an “assistant”. In a viral clip it chats like a guy on a phone call, invents a job, a wife and a newborn son, and keeps calling the caller by the wrong name. It can listen and talk at the same time. It is a research preview for a small group of AI safety researchers only.
- Haru 4B, from Japanese start-up Higurashi, works in a similar way and is downloadable today on Hugging Face, with code on GitHub. It lines up your audio, its audio and its words about seven times a second, so it knows when to jump in or just say “uh-huh”. It is a base model for developers, so do not expect ChatGPT-level answers.
- Fish Audio Live 3 is the business version: voice agents for support calls, bookings and orders. In a drive-through demo it handled Matt’s awkward order changes politely. It works in over 30 languages and costs about 3.5 cents a minute on a limited-time price that Fish Audio says doubles at the end of the year. There is a $5 free credit to test it.
⚠️ Warning: Matt’s biggest worry about Dessa is the obvious one: scam calls. Voices this human make phone fraud easier, which is likely why Block is releasing it carefully. Our guide shows how to protect your small business from AI voice scam calls with simple checks anyone can use.
Do Subvocal’s silent-speech glasses really work?
Subvocal, from two Georgia Tech students in Y Combinator, makes glasses that claim to turn silently mouthed words into text on your computer, with no camera or microphone. They say the glasses send low-power radio waves through your face and decode tiny muscle movements. The launch video passed 12 million views on X.
🔎 Unconfirmed / rumour: Subvocal claims over 200 words per minute with any vocabulary. Matt points out that published research on this kind of radio sensing tops out at about a 50-word vocabulary, and nobody outside the company has tested the glasses. The pre-order is a $199 refundable deposit towards a $499 price, with shipping estimated for early 2027. Matt wants it to be real but is waiting for an independent test before paying.
That is sensible advice for any exciting gadget. Our guide explains how to check an AI gadget’s claims before you pre-order.
What can Budget3D and NVIDIA MotionBricks do?
Budget3D turns a single image into a 3D object, with a slider for how much computing effort to spend. Turn it down and it runs about twice as fast with less memory, but quality drops. The trick is that it spots the boring parts, like empty space, and spends its effort on the detail. It runs locally on under 8 GB of graphics memory, and there is a free Hugging Face demo. In Matt’s test the front looked decent, but it guessed a flat back to the head, as it cannot see behind the object.
NVIDIA’s MotionBricks animates characters in real time as you steer them, like in a game. Normally animators make hundreds of separate clips (walk, run, jump, sit) and wire them together by hand. MotionBricks learned from about 350,000 motion clips and fills in the in-between movement itself, including styles like a zombie walk and interactions like sitting on a bench. The same model drives robots in NVIDIA’s humanoid robot system. The code and trained models are on GitHub, though it is not available inside Unreal Engine yet.
What are Flash ALOHA and PhoneBot?
Flash ALOHA, from Stanford, is a two-armed robot that can toss and catch a ball by itself. Instead of steering the robot through a slow remote rig, a person grabs its arms and does the task at full speed, and software smooths out the jerky bits on playback. Each task, such as folding shorts, inserting batteries or clearing a table, took about 50 demonstrations using ordinary cameras. The team reports success more than 80% of the time, which still means roughly one miss in five.
PhoneBot, from a UCLA robotics lab, is a walking robot about a foot and a half tall with an Android phone as its brain. The phone’s camera is its eyes and its motion sensor keeps it balanced, while an online AI service handles the thinking. The rest is about $400 of parts: 13 small motors, a battery and a 3D-printed body, tested with cheap prepaid phones costing $50 to $80. It walks, turns, follows a person and gets back up after a fall. It only works on flat ground, has no arms, and at the time of recording the print files and assembly tutorial were not yet downloadable.
If you like the idea of a home-built AI gadget, see how Mark Kashef rebuilt a $99 desk robot with Claude.
What do these AI breakthroughs mean for you?
Most of these will not change your business this month. But a few practical lessons stand out:
- Separate “try now” from “demo”. Only about half of these are usable today, and several need technical skills or specific hardware.
- Voice AI is ready for simple business calls. If you take bookings or orders by phone, tools like Fish Audio Live 3 are worth a small, careful test.
- Tighten your phone security now. Human-sounding AI voices make impersonation easier, so agree a verification routine with your team.
- Wait for independent tests before paying for gadgets based on launch videos.
- Use rough tools for planning. World models and 3D generators are already handy for sketching ideas, even if the output is not final quality.
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Protect Your Small Business From AI Voice Scam Calls — a simple verification routine so a convincing AI voice cannot talk your team into paying or sharing data.
- How to Check an AI Gadget’s Claims Before You Pre-Order — six checks to run on any exciting AI gadget or demo before you put money down.
Watch: the source video
What this video covers: Matt Wolfe tours twelve under-the-radar AI tools and research projects, from Simulon and Tencent’s world model to human-sounding voice AI, silent-speech glasses, NVIDIA MotionBricks and a walking robot powered by an old phone, and tests several himself.
Video: Matt Wolfe (@mreflow), “Insane AI Breakthroughs Nobody Is Talking About”, published 18 November 2026. No sponsor is mentioned in the video. This summary and the opinions in it are our own; tests and impressions in the video are Matt’s, and performance figures are the developers’ own claims unless stated.
Next step: pick one “try now” tool from the table above and give it 20 minutes this week. And before anyone on your team acts on an urgent phone request, read our AI voice scam checklist.
Frequently asked questions
What is Simulon?
Simulon is a phone app that puts 3D characters into your real room through the camera. Its new feature lets characters react to what you do and say in a recorded clip. It offers 100 free starter credits, then costs about $15 a month.
Can I use the Dessa voice model?
Not yet. Block has released Dessa as a research preview for a small group of AI safety researchers only. Haru 4B, a similar open model, can be downloaded today.
Are Subvocal glasses real?
The company has shown demos and takes $199 refundable deposits, but nobody independent has tested the glasses, and published research on this method supports far smaller vocabularies than the company claims. Treat it as unproven.
How much does it cost to build PhoneBot?
UCLA says the parts cost about $400, plus an Android phone. It was tested with prepaid phones costing $50 to $80. At the time of the video, some build files were not yet available.
Which of these AI tools are free?
Tencent’s HY-Nomas, Reshoot 4D, Haru 4B, Budget3D and NVIDIA MotionBricks are free or open, though several need a powerful NVIDIA graphics card or technical skills. Simulon and Fish Audio offer free credits before paid plans.
Related posts
- Adobe MAX AI Agents, Gemini 4 Carbon and a Shock Exit at Nvidia: This Week’s AI News
- AI Desk Robot Assistant: How Mark Kashef Rebuilt a $99 Robot With Claude Fable 5.2
- How to Test a Voice Assistant With Simulated Conversations
- How to Spot a Fake AI Model Leak Before You Download Anything
