Modern Creator
Matt Wolfe · YouTube

Insane AI Breakthroughs Nobody Is Talking About

A monthly sweep of twelve under-the-radar AI and robotics releases, each sorted by whether you can actually use it today.

Posted
4 days ago
Duration
Format
Listicle
educational
Views
210.9K
3.1K likes
Big Idea

The argument in one line.

The most useful AI news is often what never hits the headlines, and the real test of each release is whether you can run it today or only watch a demo.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • A creator or video editor who wants to know which new AI tools for AR, world models and reshooting footage can be tried today.
  • A developer or hobbyist who wants open-weight models and code on GitHub or Hugging Face rather than closed demos.
  • Someone who follows the weekly AI news cycle and wants the interesting releases that did not make it into those roundups.
SKIP IF…
  • You want one deep tutorial on a single tool; this is a fast tour of twelve projects, most shown only briefly.
  • You only care about the biggest model launches from OpenAI, Google or Anthropic, which this video deliberately skips.
TL;DR

The full version, fast.

Matt Wolfe covers twelve AI projects that missed the weekly news cycle. Simulon drops reactive 3D characters into phone video, a Tencent world model keeps scenes consistent for about a minute, Reshoot 4D re-films existing footage from new angles, and Adobe's Movie Edit changes whole scenes from one prompt. Block's Dessa and the downloadable Haru 4B talk without a text step, Fish Audio's Live 3 targets business calls, and SubVocal claims silent speech through glasses. Budget 3D, Motion Bricks, Flash ALOHA and Phonebot round out the 3D and robotics picks. For each, he flags what you can download versus what is still a paper or demo.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →
Chapters

Where the time goes.

00:00 – 00:56

01 · Cold open

Sets up the video as the under-the-radar companion to the weekly news show and previews the lineup.

00:56 – 04:37

02 · Simulon AR characters

A phone app that drops 3D characters into your room and redraws the recording so they react to you.

04:37 – 08:18

03 · Tencent world model

An open model that builds a consistent world in five-second pieces, steered by text and camera moves.

08:18 – 10:30

04 · Reshoot 4D

Rebuilds existing video in 3D so a virtual camera can reshoot it from a new angle.

10:30 – 11:23

05 · Movie Edit

Adobe-linked research that edits every shot in a scene from one instruction, with no public model yet.

11:23 – 15:00

06 · Dessa voice model

Block's research voice model, trained on raw audio, that sounds like a real person on a phone call.

15:00 – 16:59

07 · Haru 4B

A downloadable base model that listens and talks at once with no text step in between.

16:59 – 18:46

08 · Fish Audio Live 3

A paid voice agent for business calls, demoed as a drive-thru order taker.

18:46 – 20:50

09 · SubVocal glasses

Glasses that claim to turn silent mouthed words into text using radio sensing, unverified so far.

20:50 – 23:12

10 · Budget 3D

Single-image 3D generation with a slider that trades quality for speed and memory.

23:12 – 24:40

11 · Motion Bricks

NVIDIA's real-time character animation model that generates transitions and object interactions itself.

24:40 – 26:41

12 · Flash ALOHA

A Stanford two-armed robot that learns fast tasks like catching a ball from hand-guided demonstrations.

26:41 – 28:49

13 · Phonebot

A UCLA walking robot with an old Android phone for a brain and about $400 in parts.

28:49 – 29:25

14 · Outro

Asks for comments on which to try first, plugs the Friday news video and the newsletter.

Atomic Insights

Lines worth screenshotting.

  • Simulon redraws your recorded video in the cloud with a 3D character that reacts to your hands, your props and your words, so nothing happens live.
  • Tencent's world model builds video in five-second pieces and re-reads everything it already made, which is why the scene stays consistent without a game engine.
  • A world model that runs on one RTX 4090 is a compressed, slower version; the real-time demo used eight data center GPUs.
  • Reshoot 4D rebuilds a video as a 3D scene over time, so a virtual camera can orbit, push in or do bullet time on footage you already shot.
  • Reshoot 4D cannot invent what was never on camera, so anything behind the subject can change between angles.
  • Adobe's Movie Edit applies one text instruction across every cut in a scene, solving the old problem where an edited jacket reverts at the next shot.
  • Block's Dessa was trained on raw audio of people talking with no text and no language model, so it behaves like a guy on a phone call instead of an assistant.
  • Because Dessa has no assistant personality, it invents a whole life, including a newborn named Darius Johnson, and that is why it sounds so human.
  • Haru 4B lines up your audio, its own audio and its words about seven times per second, which lets it say mm-hmm without stopping its sentence.
  • An open voice model arrived just three weeks after the closed one it mimics, and a paid product followed, all doing roughly the same thing.
  • Fish Audio's Live 3 handles interruptions in over 30 languages for about 3.5 cents per minute, a launch price that doubles at the end of the year.
  • SubVocal claims 200 words per minute, but published radio-sensing research tops out around a 50-word vocabulary, so the claim needs independent testing.
  • Budget 3D lets you slide how much compute a 3D generation spends, trading about twice the speed for visibly worse quality.
  • Most 3D models spend equal effort on empty space; Budget 3D groups the boring regions and spends its compute where the detail is.
  • Motion Bricks learned from about 350,000 motion clips and generates every transition live, replacing the hand-wired animation graphs studios build today.
  • Flash ALOHA teaches a robot by letting a human physically move its arms at full speed, which beats remote-rig teaching for fast tasks like catching.
  • Phonebot uses an old Android phone as the robot's brain, using its camera, gyroscope and GPS, with about $400 in other parts.
  • Nearly every project here ships with a catch: no download yet, flat ground only, or claims that come from the maker's own demos.
Takeaway

Check whether you can run it today

WHAT TO LEARN

The useful skill with new AI releases is sorting them by access: open and downloadable, paid, closed preview, or just a paper with claims no one has tested.

02Simulon AR characters
  • Simulon's cloud redraw lets a character react to your hands and props, but the free tier is small: 100 starter credits, then about $15 per month.
  • Reactive effects like poking the character are iPhone only for now, so check platform support before building a video workflow on it.
03Tencent world model
  • Tencent's world model keeps a scene consistent by re-reading its own earlier five-second pieces, but its consistency claims come from the company's own demos.
  • Running it on a single RTX 4090 means a slower compressed version; the real-time setup used eight data center GPUs.
04Reshoot 4D
  • Reshoot 4D can orbit or push in on footage you already shot, but it cannot add what was never on camera, so backgrounds shift between angles.
  • It needs an NVIDIA card and ComfyUI knowledge, and you cannot upload your own footage to the online demo yet.
05Movie Edit
  • Movie Edit applies one instruction across every cut in a scene, which fixes the old problem of edits reverting between shots, but it is only a paper for now.
  • A paper plus sample videos with no downloadable model is a watch-this-space item, not something to plan a workflow around.
06Dessa voice model
  • Dessa sounds human because it was trained only on audio of people talking, so it plays a person on a call instead of a helpful assistant.
  • That realism makes scam calls easy, which is the reason access is limited to a small group of safety researchers.
07Haru 4B
  • Haru 4B shows the same listen-and-talk-at-once idea in an open model you can download, though it is a base model and not ChatGPT-level smart.
  • It tracks your audio, its own audio and its words about seven times a second, which is how it knows when to say mm-hmm instead of cutting in.
08Fish Audio Live 3
  • Fish Audio's Live 3 handles interruptions in over 30 languages at about 3.5 cents per minute, but it is closed and the launch price doubles at year end.
  • Three similar voice products in three weeks suggests the capability is commoditizing, so compare price and openness before committing.
09SubVocal glasses
  • SubVocal's radio-sensing glasses claim 200 words per minute, while published research of this type reaches only a 50-word vocabulary, so wait for independent tests.
  • A refundable $199 deposit toward a $499 price is still money down on an unverified prototype with a Q1 2027 ship estimate.
10Budget 3D
  • Budget 3D lets you trade quality for speed with a token budget slider, giving a rough preview quickly and a clean version when you need it.
  • It runs in under 8GB of VRAM, so most recent gaming cards can handle it, and a free Hugging Face space lets you test your own images.
11Motion Bricks
  • Motion Bricks generates character movement and object interactions in real time from about 350,000 motion clips, removing hand-wired transitions.
  • The code and models are public, but it is not in Unreal Engine yet, so it suits solo developers willing to integrate it themselves.
12Flash ALOHA
  • Flash ALOHA teaches fast skills by letting a person move the robot's arms at full speed, which beats slow remote-rig demonstrations.
  • Each task needed only 50 demonstrations and regular cameras, but an 80 percent success rate still means missing about one catch in five.
13Phonebot
  • Phonebot shows a cheap phone can supply the camera, motion sensors and GPS for a robot, with about $400 in other parts.
  • Its build files and tutorial were not yet available, and it only walks on flat ground and has no arms, so it is a project to watch.
Glossary

Terms worth knowing.

World model
An AI model that generates a continuous, explorable environment you can steer, rather than a single fixed video clip.
Open weights
A model whose trained parameters are published so anyone can download and run it, often with the code on GitHub or Hugging Face.
Bullet time
A camera move that appears to orbit around a subject frozen or slowed in mid-action, made famous by The Matrix.
ComfyUI
A node-based interface for running image and video AI models locally, which tools like Reshoot 4D require you to know.
Full-duplex voice model
A voice AI that listens and talks at the same time, so it can add small acknowledgments without being interrupted or stopping.
Base model
A model trained on raw data that has not been tuned to follow instructions, meant as a starting point for developers rather than a finished assistant.
RF sensing
Using low-power radio waves to measure small changes in the body, here the tiny facial muscle movements made when mouthing words silently.
Token budget
A setting that caps how much computation a model spends on an input, trading quality for speed and memory.
VRAM
The memory on a graphics card, which limits how large a model you can run locally.
Resources

Things they pointed at.

00:56productSimulon
04:15productTencent world model (Hynomus)
08:18toolReshoot 4D
08:30toolLTX
10:30productMovie Edit (Adobe research)
11:23productDessa (Block)
15:00productHaru 4B (Higurashi)
16:59productFish Audio Live 3
18:46productSubVocal
20:50toolBudget 3D
23:12toolMotion Bricks (NVIDIA)
24:40toolFlash ALOHA (Stanford)
26:41toolPhonebot (UCLA)
29:19linkFuture Tools newsletter
Quotables

Lines you could clip.

00:00
“This is the really cool stuff that kind of flew under the radar.”
clear premise statement that frames the whole video→ IG reel cold open↗ Tweet quote
12:15
“It's the most human sounding AI conversation I've heard so far, full stop.”
strong claim tied to a funny, listenable clip→ TikTok hook↗ Tweet quote
20:48
“I really want these to be real and work the way they claim.”
honest skeptical optimism about unverified hardware→ newsletter pull-quote↗ Tweet quote
The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

I'm going to show you some of the coolest AI and tech stuff I've come across recently. So every Friday, I cover the biggest AI news that everyone needs to know about. This is not that video.
Today, we're talking about all the other stuff. The things that I've found in the past month or so that didn't make it into those videos. This is the really cool stuff that kind of flew under the radar.
We've got new world models that are better than Google's. Crazy AI voice technology that you swear was human. I'm gay.
How you doing, Renato? I'm good, I'm good. Glasses that claim to read words you never actually said out loud, a robot that can catch a ball out of the air, a walking robot that uses an old Android phone as its brain, and you know, a bunch of other crazy tech that'll show you that we're actually currently living in the future.
But let's get us started with some really cool augmented reality tech. So this Dracula clip, came from a phone app.
It's called Simulon and it puts 3D characters into your real room through the camera. And they just added a feature where you place a character, hit record, act out a scene, and then the character actually reacts to what you do and what you say. So here's a short film that Simulon created to sort of show off what it's capable of.
You can see this like caveman here walking around and he's walking through a real environment. These are real people. This is a real video, but the caveman is actually being added in later.
Like for example, here's the caveman jumping into an Uber interacting with a real car that's really sitting there in a real video. But if we fast forward to the end of the video, we can actually see that this is a clip of real footage and nothing was ever actually there. And here's that full Dracula clip from the founder.
And this started just as a single image. So check this video out. So you can see the RC car is real and he's riding the real car across the table.
You can see his cape is even like moving around, flowing behind the car. And then when the car goes over the edge, the animated character knows to sort of jump off and not fall with it. And the founder says none of this was actually prompted.
He just put that. character into the world and then drove the car around now none of this stuff is happening live you record your take and then when you stop the app sends it to a cloud and then redraws the whole video with the character in it it actually looks at what's in the scene like the table the car your hands what you said and then it builds a new version of the shot from all that information okay so i'm in the simulon app on my phone right now and we have all these characters we can choose from or we can actually add our own character if we want inside the library but That one requires you upgrade.
And I'm just playing with a free plan right now. Grab this like white gorilla here. And now you can actually see it switched over to my camera on my phone.
So here's our gorilla sort of floating in the air in my office. So I need to like shrink it down. So let me just sort of grab it and shrink it.
And now that he's shrunk down, I could put him on my desk right here. And let's see if I could spin him around. Come on.
There we go. Okay. So now he's sort of facing me.
He's on my desk. So I can place him in that spot. And now you can see he's just like living in that spot in augmented reality.
There's my camera. Now, if I record a video, I theoretically can interact with him. So I can like, you know, pet him like that.
I could grab my pen over here. poke him with a pen and theoretically my video he will interact with whatever i'm doing now if i click create video it says your render is on the way so it will take you know a couple minutes to render this all right so here's the video we got but you can see there's the gorilla on my desk I put my hand down and when I go to pet it, you could actually see he's interacting with the hand that I just put over him.
When I poke him with the pen, you can actually see him interacting with me, poking with the pen and like swatting the pen away. It looks pretty cool. And that reacting feature where I poked it with a pen and petted it, that's actually iPhone only right now.
It'll be on Android soon. And you can actually try this for free. They give you 100 starter credits.
And then after that, it costs 15 bucks a month to use. those starter credits go really fast. But if you have like a video workflow where you're trying to put something in your real environment, this could be pretty cool.
Also picture this live running in a pair of like AR glasses because the founder says that's what they're building toward. They want you to be able to wear glasses and see this stuff and interact with it through glasses, which will be pretty cool. All right, so we just saw characters dropped into the real world.
Well, Tencent just released a model that generates the whole world and you direct it by talking to it. Now this one's called a Hynomus. So most AI video works like this.
You write one prompt, you get one clip, and if you want something different, you start from scratch. With this one, the video keeps going and you keep giving it direction. You can say things like add a dog, then, you know, make it snow, then swing the camera to the left.
and it stays the same world the whole time. So if we watch this clip here, we could see this astronaut looking around, and then when he comes back, you can actually see it's the same exact world he originally saw when he was originally looking in that direction. Some of the other world models we've looked at in the past, they lose that consistency.
They look one direction, they look away, and when they look back, it looks different. This one maintains that consistency. So the way this one works is it builds the video in five -second pieces.
Each new piece takes your latest direction, which can be text, a camera move, or both. And before it draws anything, it actually looks back at everything it already made. And that memory is what keeps the world from changing on you.
And it can keep going for about a minute. If you remember last month's video, we looked at world models that stayed consistent by letting a game engine keep track of everything. This one has no game engine.
It's the video model remembering everything on its own. I'll link up in the description if you want to play with this. one, but you can start with an image reference or without an image reference.
You could select first person, third person, second person, third person observer. Let's just go ahead and do third person for our environment. I'll just do like a lush green forest in the future.
For our subject, let's do a tiny ant. We can actually see when I jump into this as it's simulating it, that it actually improved my subject and my scene. It added a lot more detail here.
All right, so here's my little ant walking around in our lush green forest here. Let's say a deer runs by. And every time we do this, it kind of does this simulating what happens next.
So it sort of builds it out iteratively, but it's not happening as much in real time as something like Genie was. Now we see our video picks up right where it left off and.
I mean, it's a kind of wonky looking deer and we can change our camera control and I can actually do this by adjusting the perspective here using my arrow keys and my WASD keys. And so let's rotate it around. So maybe it's looking at like the front view of it.
A little hard to. control because if I hit W, it's going to the right. If I hit A, now it feels like it's moving forward.
But OK, let's go ahead and generate that camera position there. And now we get this, which is looking a little wonky. Not going to lie about that, but we can see the camera angles changing quite a bit based on the movements that I was making with my arrow keys earlier.
Yeah, not the most impressive. Result, very interesting. All right, so a few quick notes on that.
This is fully open. So the code and model are free to use and. You can even use it commercially if you want.
Tencent says it can run on a single RTX 4090, but that's like a compressed slower version. The real time setup that they tested actually used eight data center GPUs and Tencent hasn't published numbers comparing it to other world models. So the consistency claims come from like their own demos.
Where I can see something like this being used is if you're trying to plan out a shot, you can talk your way through a scene, prompt different things you want to see in the scene, and then sort of map it out before going and actually filming anything. Clearly the quality on this isn't good enough to use in any final like production work, but it is cool to see new world models, especially ones that are becoming open weight and available for anybody to use.
Now that last one was moving a camera around a world that AI made up. This next one moves the camera around a video that you've already shot So last month I showed ByteDance doing this to an old movie scene But the catch was that nobody could use it yet. This one you can actually download today It's called Reshoot 4D.
You give it a video and you reshoot it from a new angle. You can orbit around the person, push in, crane up, even do like a bullet time move. So there's actually four steps for this.
First, it builds a 3D capture of your scene as it moves through time. It works out where your original camera was and how your subject moved. Second, you have a virtual camera.
You set where it goes, the lens, the focus, or you pick a preset like orbit or crane. Third, it reshoots that scene from the new camera angle, which does create some weird like fuzziness and artifacts in the background. So step four, it uses an open model called LTX and it does a final pass.
So it actually looks like real footage and remove some of the weirdness from the edges. And they actually let you demo the editor. So you can't actually upload your own footage yet, but you can see how they works.
So here's the video that we can edit. And if we switch to the editor view, we can actually see the various like orbits and things that we can do with it. So I can actually grab this camera here and have it move over this way, move this here.
I don't actually know what I'm doing, but let's just see what happens. And now we can see our shot in the bottom right corner where it moves to these various camera spots that were key framed. Again, we can actually get like the full view of this.
It's not fully available yet. And a few other quick notes it doesn't build anything new behind your subject so things that were off camera can change from one angle to the next we can actually see in this example here what's outside the window slightly changes with each new shot and they do say that fast motion can smear a face for like a few frames so it's not gonna look great if there's a lot of fast motion right now and you do need an nvidia graphics card specifically as well as some knowledge of how to use comfy ui which is not the you know smoothest learning curve, but it is free and open source.
And it was built by one developer who says he made the whole thing from Claude code and codex. And if changing the camera isn't enough for you, this next one, I'm going to show you. changes what's actually in the shot.
So this is movie edit, which is a research model from Adobe and a university in Korea. So you type in one instruction, like remove the man in the black tactical suit, and it removes him from every shot in that scene. Now, the old problem is that AI video editing just works with one shot at a time.
You change a character's jacket, the video cuts to the next shot, and the old jacket might be back again. You can also show it what you want. Edit a single frame in Photoshop or anywhere else, and it carries that change through the whole clip.
across all the cuts it can handle clips up to 25 seconds long and it makes quick rough drafts first so you can check the edit before a full render now there is a catch to this one and that's that it's just a research paper at the moment as well as like a set of example videos there's no model to download and no word on whether it's going to end up in premiere or not so This is just one we're going to have to keep an eye on because we can't play with it ourselves yet.
All right, I'm going to give your eyes a break for a second because this next one just needs for you to listen. This next one's called Dessa and it's a voice model from a small research lab that's inside of Block, which is Jack Dorsey, one of the original Twitter founders companies. And with this one, you can have a real back and forth conversation with it.
In fact, this is one of my favorite clips of recently. Somebody on the team posted this and it's pretty hilarious. Listen to a small clip of this.
Hey, how you doing? How you doing, Rinaldo? I'm good.
How you doing, Rinaldo? I'm good. I'm good.
I'm not Rinaldo. You're mixing me up with another guy. Where's that?
No, I'm Devere, man. I'm not Rinaldo. No, you're not Ronaldo.
And it's funny because if you listen to this whole conversation, he asked him how the family is and he says, my kid's good, man. I got a new baby boy from October 20th. We're calling him Darius Johnson.
He's got a daughter who likes that name. And like this character from Dessa just makes up all sorts of like stories about his life. And then it ends.
That's awesome, man. All right. Good talking to you, man.
And Dessa responds with. Yeah, you too, Renato. See you later.
He calls. and Rinaldo again at the end. I don't know why that's so funny to me.
But I mean, one of the more fascinating things besides how realistic it sounds is that, again, the AI said it had a newborn son named Darius Johnson. Earlier in the call, it complains about being tired from its job picking produce. Later, it tells this long story about his wife and the baby crying, but none of it actually exists.
It made up a whole life. And the reason that it acts like that is, well, A lot of the voice AI turns your speech into text, runs that text through a language model, like a GPT or Claude model, and then it turns that answer back into speech.
DESA doesn't. Block says they trained it from scratch on recordings of people talking with no text and no existing language model underneath it all. Everything it knows, it's purely picked up by listening to audio.
That also means that it never got an assistant personality similar to ChatGPT or Claude. It learned what a phone call between two people sounds like, so it just acts like a guy on a phone call with a job and a family and a story about his kids. It can also listen and talk at the same time, just the same way people do.
In the clip, which I think is a must listen if you haven't listened to the whole thing, you can actually hear Devere say yeah and uh -huh underneath it while the Dessa model keeps talking. You're alive. You know what I'm saying?
We'll do a whole nother situation, but I didn't want to give him up. I know. Take him and let me have it.
Yeah. So now a couple of quick caveats on this one. You can't use it yet.
It's a research preview and Block is only giving access to a small group of AI safety researchers right now. And they do say it can run on a MacBook Pro using an M4 Max chip, but nobody outside of Block has actually been able to test it yet. And again, I think these clips are hilarious.
They sound really good. It's the most human sounding AI conversation I've heard so far, full stop. And I think that's because it's not.
trying to help you with anything like an assistant would. But the obvious downside to something like this is, well, the scam potential. Like this can really easily be used for scam calls, which is why I think they're being so careful with releasing it to everybody.
I do. It's my life and I'm sticking with it. And I'm going to make you watch me die.
No, no, I don't want that. Like, why would I want to see that? Okay, so you can't use DESA yet.
But here's one that works the same way and you can actually download it today. It's called Haru 4B from a Japanese startup called Higurashi. It's the same idea.
It listens and talks at the same time with no text step in the middle. Are you there? Yeah, I'm ready.
I'm excited. Oh, John. Yeah.
How are you? How are you? Yeah, I'm good.
I'm good. Actually, I wanted to ask you about what's your favorite? rom -com of all time uh well i think i am very familiar with american pop culture so i probably would say that my favorite rom -com would be How to Lose a Guy in 10 Days.
Notice how there was like a little okay and yes and mm -hmms in there while the AI was talking and it didn't actually interrupt and cause the AI to stop talking. So the way it works is that at every step, it lines up three things. It lines up your audio, its own audio here, and then the words that it's saying.
And it does that about seven times per second. And that's how it knows when to jump in and when to just say mm -hmm. And on a MacBook, they say this whole loop takes about a fifth of a second.
Now there's a slight catch to this one. It is a base model, which means it's a starting point for developers to build on. It's going to chat with you, but don't expect like ChatGPT level or Claude level intelligent answers.
But the model is available over on Hugging Face right now, and they put the code up on GitHub. So you can use this and download it right now. And if you do want to test it yourself, they have a Hugging Face space where you can play around with it, but I don't.
quite understand how to use it in this hugging face space to have like an actual dialogue. It looks like you either record or upload audio and then it. adds its own audio back afterwards so it doesn't feel like a real test of its capabilities i think the most interesting thing about this one is the timing of it this open version showed up just three weeks after that desa closed one which nobody can use right now and there's a third audio model i want to talk about from fish audio this is the same company behind the drama 3 voice model from the last month video that i did like this one this model is called live 3.
and This one is built for businesses for things like, you know, support calls and booking things and things like that. Like here's an example of a drive -thru restaurant.
You can actually demo this yourself. So let's start a call here. Welcome to Harbor Burger.
What can I get for you today? Can I get a large fries, a triple cheeseburger and a chocolate milkshake?
We don't have a triple cheeseburger. Sorry, the biggest is the double cheeseburger, $4 .99. Can you just add another meat patty to it for me?
Sorry, I can't add an extra patty. The double comes as it is. Okay, then give me a double cheeseburger and give me a single cheeseburger and I'll combine them myself.
Okay, so it's a double cheeseburger, a cheeseburger, a large fries, and a chocolate milk shake. Make it an Oreo shake instead.
We don't have Oreo shakes, sorry. Just vanilla, chocolate, or strawberry. Which would you like?
Okay, keep it a chocolate. So, I mean, it sounds pretty good. You can interrupt it.
You can talk through it. And they have some impressive demos. It works in over 30 languages.
And it costs about three and a half cents per minute right now. But that's actually a limited time offer. That price doubles at the end of the year.
However, if you do want to try it, you can get a $5 free credit to just go and test it out. There isn't a model or anything like that to download. This one runs through their site.
in their API. It's not open. This is a subscription payment model.
So over the course of about three weeks, we went from one research demo that was really impressive to an open model that works almost as good and now a paid product. And they all kind of do the same thing. about as good as each other.
Okay, this next one's a little weird and a little bit different. So everything we've just heard needs you to talk out loud to it. Well, these are glasses that claim to let you talk to AI without actually making a sound.
This is from a company called SubVocal and it's two Georgia Tech students in the current Y Combinator batch. And they started this at a hackathon back in February. So check this out.
He's actually talking, but no audio is coming out. And as he's talking, it is putting up what he said. on the screen now i do have this muted but you see their lips moving but no noise is coming out of their mouth while they're talking when they're wearing these glasses There's no camera and no microphone doing this.
SubVocal uses RF sensing from a pair of glasses to pick up the tiny facial movements you make while silently articulating words. We send low power radio waves through the face and measure how they change as your muscles move. Our models then decode those signals directly into text.
Their launch video got over 12 million views on X. Now, a few quick notes on this. The company claims over 200 words per minute with any vocal.
vocabulary. However, the published research on this kind of radio sensing tops out at a 50 word vocabulary. So that's a pretty big gap.
And no one outside the company has actually tested these glasses yet. The founders also say that this current prototype looks a lot worse than the actual glasses on the website that you'll end up getting. If you look at these, they just look 3D.
printed. Now a pre -order for this is 199 bucks and that's a refundable deposit towards a $499 price tag. They estimate they'll ship in Q1 of 2027.
So just within the next few months. Now I personally really want this to be real. I think dictating to my computer, like in a coffee shop or like on a plane without anybody hearing me would be pretty awesome, but I'm still kind of waiting for like an independent test before actually putting money down on these.
I don't totally know what to think yet, but I really want these to be real and work the way they claim. All right, so back to things you can actually run yourself.
At the top of the video, Simulon turned a photo into a 3D. character well here's a tool for that step it's called budget 3d you give it one image and it can give you a 3d object back now what's different about this one is this slider here you choose how much computing power it spends you can turn it down and it runs about twice as fast and uses less memory but you're also going to get a bit of a quality hit so we can see in this example here this is the input image of this little cowboy toy when using the full token budget you can see it comes out pretty good quality but when we change this slider, you can see the quality degrade, but also the speed increase.
And if I bring it all the way up to 40 % here, it used a lot less tokens, a lot faster speed, but I mean, you can see the quality of the 3D that it generated is quite degraded. Here's another example of a toy train at max speed and lowest budget. Not great.
Crank it all the way up. It's going to work a lot slower, use a lot more tokens, but you're going to get a pretty good quality output. Normally, these models will spend the same effort on every part of the object, including the empty space around the object.
This one actually finds the boring parts, starting with that empty space and then lumps them together. And then it spends all of its effort where the details are. And this one, they do have a demo available over on Hugging Face as a Hugging Face space where you can upload your own image here and set your token budget.
So let's give it this image here, do the max token budget to start. And here's what it generated. I mean, obviously it doesn't what the back of it looks like.
So we get a flat head, but from the front, not too bad. Crank down the token budget all the way, generate again. It generated a lot faster, but yeah, that's what we got out of it.
Now this page does show how much faster it gets by changing the token budget, but there's no like numbers or metrics towards the quality difference. So you kind of have to go by your own eyes on, you know, which level works for your needs. Now, To run this model locally, it only needs under 8GB of VRAM, so pretty much any recent gaming card should be able to handle it.
The code is available on GitHub, the model weights are available over on Hugging Face, and you can demo this free inside of this hugging face space in your browser right now. So once you have a 3D character, you may now want that character to actually move.
Nvidia's research team has a new answer to that. It's called Motion Bricks, and it animates the character in real time. You steer the character like you would in a game, and every movement is generated on the spot.
So everything in this clip comes from the model. Nvidia says there are no handmade transitions in this at all. In a normal game, animators build hundreds of separate clips, like a walk, a run, a jump, a sit.
Then somebody has to wire them together by hand so that the character moves smoothly from one point to the next. That wiring is a lot of expert work. So here, one model learned from about 350 ,000 motion clips, and it handles all the in -between by itself.
So there's two parts. The first is the movement. You give it a speed, a direction, and a style, and you can switch styles on the fly.
So you've got like design, Bombi walk here. You've got an injured leg style here.
You've got a skipping style here. And then the second is objects. You mask a few key poses like hand on the sword and sitting on the bench.
And the model fills in the approach, the contact and the follow through. So here we can see it goes and picks up the object. And here we can see it goes and sits on the bench.
And the same models driving robots. It's the motion layer inside of Groot, which is NVIDIA system for humanoid robots. Now these demos, look really really good but it's not available to use inside of unreal engine just yet so game developers can't just drop it into any project yet but the code and the trained models are on github right now if you're a solo game developer this is character animation without needing an animator budget and speaking of robots this next one is about speed this is flash aloha out of stanford what we're seeing is a two -armed robot tossing a ball and catching it with nobody controlling it.
And catching is actually pretty hard for a robot because there's really no time to think. And most robots learn stuff like this by watching a person and then end up a lot slower than the person because the way we usually teach them is slow. Somebody steers the robot through a remote rig and that gets clumsy at high speed.
But here's the fix that they found. They built arms you can grab and move by hand. So a person takes hold of the robot's arms and does the task at full speed and then the robot robot records that, then a small piece of software smooths out the jerky spots when the robot plays it back.
It could also do things like fold your shorts, put batteries into a device for you, and actually cleans a table for you after eating. Each one of these tasks took just 50 demonstrations to learn, and it only uses regular cameras. Now, they do report that it succeeds more than 80 % of the time, which still means that it could miss one in five catches.
And they also say it's close to human. speed in their description. The page doesn't really specify what that means.
But the code for this one is on GitHub and their documentation tells you all the hardware you would need to purchase if you want to build one of these yourself. For me, the catching was the most impressive thing. I haven't seen a robot that learned from a demonstration react as fast as what we're seeing here.
That part was really cool. All right, so last one. If those robots feel out of reach for you, here's one you can actually build yourself with an old phone that's in your drawer right now.
I actually made a short all about this and it turned into one of of the most viral shorts I've posted in a long time. So I wanted to give it a little more time in this video.
This is Phonebot from a robotics lab at UCLA. It's a walking robot about a foot and a half tall and its brain is an Android phone. Now the phone does all of the hard parts of making this robot work.
Its camera is the robot's eyes. The same sensor that rotates your screens is the sensor that they use to keep it balanced. And it does connect to the internet to use the GPS.
for the actual like AI that runs the brain and how it moves. Everything else is about $400 worth of parts. They even give you the list of parts you need to buy, the various 3D print files, and an assembly tutorial to build this thing.
It's really just 13 small motors, a battery, and a 3d printed body and they tested it with some old prepaid android phones that cost anywhere from 50 to 80 bucks it could walk forwards and backwards it turns and it can follow a person around using the camera and when it falls on its face it can get itself back up without needing any extra help from anybody else which is pretty wild to watch it actually learned how to walk in a simulation by trial and error and it got points for staying upright and keeping pace and it lost points when it dragged its feet now there is one thing that i did notice it's a bit of an issue they do tell you what files you need to purchase and they do have like the cad files but i just noticed that the printable parts and the actual tutorial, these buttons aren't clickable.
So they aren't quite ready for you to grab yet. You can't get the 3D print files or the assembly tutorial as of the recording of this video. It also only walks on flat ground right now and also has no arms.
So, you know, it's not going to be doing your dishes or folding your laundry anytime soon. But just think about everything we've looked at so far today. A motion model from NVIDIA that already drives humanoid robots.
A voice model you can download that talks like a person. Put those into a form. $400 robot with a phone for a brain.
And that's something that would be kind of fun to build. In fact, I might. Try that.
That's a good idea for a future video. But that's what I got for you today. I'm curious which of these you want to try first.
Let me know down in the comments. And if you see something cool that nobody else is talking about, that's kind of flying under the radar, drop that in the comments too, because that might end up in next month's video. Also, I put out an AI news video every Friday.
So if you want to keep up with the news with just one video a week, make sure you like this video and subscribe to this channel. That'll make sure this video shows up in your YouTube feed. Also, I do have a free newsletter over at future tools.
Thank you so much for nerding out with me today and hopefully I'll see you in the next one. Bye -bye.
The Hook

The bait, then the rug-pull.

Every Friday Matt Wolfe covers the big AI news. This is the other video: a month of releases that never made the cut, with a straight answer on whether you can use each one.

Frameworks

Named ideas worth stealing.

03:40list

Can I use it today? checklist

  1. Is there a downloadable model or code?
  2. Does it need special hardware?
  3. Does a free demo exist?
  4. Are the claims independently tested?

The implicit test Matt applies to each project: separate released tools from research papers, closed previews and unverified claims.

Steal forvetting any new AI tool before spending time on it
08:31list

Reshoot 4D pipeline

  1. Build a 3D capture of the scene over time
  2. Set a virtual camera path and lens
  3. Re-render from the new camera
  4. Clean up with an open model called LTX

Four steps that turn existing footage into a re-cameraed shot.

Steal forplanning camera moves on footage already shot
CTA Breakdown

How they asked for the click.

VERBAL ASK
28:49newsletter
“Let me know down in the comments... I put out an AI news video every Friday... I do have a free newsletter over at future tools.”

Asks for comments first, then subscribe, then the newsletter, with a reason tied to next month's video.

Storyboard

Visual structure at a glance.

cold open
hookcold open00:01
Simulon
valueSimulon01:11
world model
valueworld model04:30
Reshoot 4D
valueReshoot 4D08:29
Dessa
valueDessa11:53
SubVocal
valueSubVocal19:08
Motion Bricks
valueMotion Bricks23:35
Phonebot
valuePhonebot27:30
CTA
ctaCTA29:06
Frame Gallery

Visual moments.

One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Watch next

More from this channel + related breakdowns.