What stays scarce14 minWhen thinking becomes too cheap to meter, the bottleneck moves to the physical world: robots, energy, matter, and maybe even human demand. A debate between an AI futurist and a lunar robotics founder, from helium-3 and moon dust to the 'last AI' and the heat death of the universe.Read now →

An AI Voice Invented a Newborn Son, Glasses Read Silent Words and a Robot Runs on an Old Phone: 12 Breakthroughs You Missed
Not every big AI story makes the headlines. From Block's eerily human Dessa voice model and silent-speech glasses to Tencent's talk-to-it world model, Nvidia's MotionBricks and a $400 walking robot with an Android phone for a brain, here are 12 under-the-radar breakthroughs, what you can try today and what's still hype.
The biggest AI news of the week gets all the oxygen. New frontier models, billion-dollar deals, lawsuits.
Meanwhile, a lot of the genuinely strange and exciting stuff slips by almost unnoticed.
An AI voice that, mid-phone call, casually mentioned its newborn son, Darius, a child that doesn't exist. Glasses that claim to turn silent mouth movements into text with no camera and no microphone. A walking robot whose entire brain is an old Android phone. A world you can direct just by talking to it.
Here are 12 under-the-radar breakthroughs from the past month or so, grouped by what they do. For each one: what it is, how it works, whether you can actually use it, and the catch.
Part 1: Putting AI into the real world
1. Simulon: AI characters that react to you
What it is: a phone app that places 3D characters into your real room through the camera. Its newest feature lets you place a character, hit record and act out a scene. The character then reacts to what you do and say.
The company's showcase films are surprisingly convincing: a caveman walking among real people and climbing into a real car, and a tiny Dracula riding a real RC car across a table, cape flowing, then jumping off before the car goes over the edge. According to the founder, that jump wasn't prompted. He just put the character in the scene and drove the car.
How it works: none of this happens live. You record your take, the app sends it to the cloud, and the service redraws the whole video with the character in it, analyzing the table, the car, your hands and what you said to build a new version of the shot.
Hands-on: in a test on the free plan, a white gorilla was shrunk down, spun around and placed on a desk. After a couple of minutes of rendering, the video showed it reacting to a hand petting it and swatting away a poking pen.
The catch: the reaction feature is iPhone-only for now, with Android promised. You get 100 starter credits, which go fast, then it's about $15 a month. Adding your own characters requires an upgrade. The long-term goal, according to the founder, is to run this live in AR glasses.
2. Tencent Hy-Nomas: a world you direct by talking
What it is: an open interactive world model from Tencent's Hunyuan team. Instead of one prompt producing one clip, the video keeps going while you keep giving directions: add a dog, make it snow, swing the camera left. And it stays the same world.
How it works: it builds video in roughly five-second chunks. Each new chunk takes your latest direction, text, a camera move or both, and before drawing anything it looks back at everything it already made. That memory keeps the world consistent: look away and look back, and the scene is still there. Tencent says it can continue for up to about a minute. Unlike some earlier consistent world models, there's no game engine tracking the scene; the video model remembers on its own.
Hands-on: you can start with or without an image and choose first-, second- or third-person views. A test of a tiny ant in a futuristic green forest worked, with the model enriching the scene. Adding a deer produced a "wonky" deer, and camera control via the keyboard was confusing, with keys not mapping intuitively. Generation is iterative, less real-time than Google's Genie.
The catch: the code and model are open, including for commercial use. Tencent says it can run on a single RTX 4090, but that's a compressed, slower version; its real-time setup used eight data-center GPUs. The consistency claims come from Tencent's own demos. Today it's best for planning shots, talking your way through a scene before filming, not final production.
3. Reshoot 4D: change your video's camera angle after filming
What it is: a free, open-source tool that re-shoots a video you've already filmed from a new angle: orbit around a person, push in, crane up, even do bullet-time.
How it works, in four steps:
- It builds a 3D capture of your scene over time, working out where the original camera was and how the subject moved.
- You place a virtual camera: path, lens, focus, or presets like orbit and crane.
- It re-renders the scene from the new angle, which introduces fuzziness and artifacts.
- A final pass with the open LTX video model cleans it up so it looks like real footage.
Hands-on: a web demo of the editor lets you keyframe camera moves on sample footage, though you can't upload your own there yet.
The catch: it doesn't invent what's behind your subject, so off-camera areas can change between angles (in one demo, the view out a window shifts). Fast motion can smear a face for a few frames. You need an Nvidia GPU and some comfort with ComfyUI. Its solo developer says he built it with Claude Code and Codex.
4. MovieEdit: one instruction, every shot
What it is: a research model from Adobe and a Korean university that edits across multiple shots at once.
Why it matters: most AI video editing works one shot at a time. Change a character's jacket, cut to the next shot, and the old jacket might reappear. MovieEdit takes one instruction, such as removing a man in a black tactical suit, and applies it to every shot in the scene. You can also edit a single frame in Photoshop or anywhere else and it carries the change through all the cuts.
Specs: clips up to about 25 seconds, with quick rough drafts so you can check an edit before a full render.
The catch: it's a paper and example videos only. No model to download, and no word on whether it'll reach Premiere.
Part 2: Voices that sound human
5. Block's Dessa: the voice that made up a life
This is the one to hear.
What it is: a voice model from a small research lab inside Block, the company led by Twitter co-founder Jack Dorsey. You can have a natural back-and-forth with it.
A clip posted by someone on the team went viral for being hilarious. The model keeps calling the caller "Ronaldo" even after being corrected, talks about being tired from its job picking produce, tells a long story about its wife and a crying baby, and announces a new baby boy, born October 20, named Darius Johnson. None of it exists. It made up a whole life.
Why it acts that way: most voice AI turns your speech into text, runs that through a language model like GPT or Claude, then turns the answer back into speech. Dessa doesn't. Block says it was trained from scratch on recordings of people talking, with no text and no existing language model underneath. Everything it knows came from listening.
That's why it never developed an "assistant" personality. It learned what a phone call between two people sounds like, so it acts like a guy on a phone call, with a job, a family and stories.
It's also full-duplex: it listens and talks at the same time, like people do. In the clip, you can hear the human saying "yeah" and "uh-huh" while Dessa keeps going.
The catch: you can't use it. It's a research preview, with early access limited to a small group of AI safety researchers. Block says it can run on a MacBook Pro with an M4 Max chip, but outsiders haven't been able to verify that. The obvious risk is scam calls: a voice this human, with no assistant persona, is a gift to fraudsters, which likely explains the cautious release.
6. Haru-4B: the open one you can download
What it is: a similar full-duplex speech model from Higurashi, a Tokyo speech-AI startup. It listens and talks at the same time, with no text step in the middle. And it's open-weight, available on Hugging Face under an Apache 2.0 license, with code on GitHub.
How it works: at every step it lines up three streams, your audio, its own audio and the words it's saying, about seven times a second (each audio frame is about 133 milliseconds). That's how it knows when to jump in and when to just say "mhm." On a MacBook, the company says, the whole loop takes about a fifth of a second.
In a demo, asked for its favorite romcom, it picked How to Lose a Guy in 10 Days, and let the human's little "okay"s pass without stopping.
The catch: it's a base model, a starting point for developers, so don't expect ChatGPT- or Claude-level answers. The Hugging Face demo space works by recording or uploading audio rather than live dialogue, so it's not a true test of its real-time skills. It supports Japanese plus several other languages.
The timing is the story: this open model appeared roughly three weeks after Block's closed Dessa.
7. Fish Audio Live 3: the paid product
What it is: a real-time voice agent model from Fish Audio, the company behind the Drama 3 voice model, built for businesses: support lines, bookings and the like.
Hands-on: a public drive-thru demo for a fictional "Harbor Burger" handled a realistic order. Asked for a triple cheeseburger, it said the biggest was a double at $4.99. Asked to add a patty, it politely refused. Told to bring a double and a single so the customer could combine them, it confirmed the order. Asked for an Oreo shake, it offered vanilla, chocolate or strawberry. You can interrupt it and talk over it.
The specs: more than 30 languages, about 3.5 cents per minute as a limited-time price that doubles at the end of the year, and a $5 free credit to try it. It's closed and runs through Fish Audio's site and API.
The pattern: in about three weeks, conversational voice went from an impressive closed research demo (Dessa), to an open model that works almost as well (Haru), to a paid product (Live 3). That's how fast this corner of AI is moving.
Part 3: New ways to talk to machines
8. Subvocal: glasses that read silent speech
What it is: glasses that claim to let you talk to AI without making a sound. In the demo, the wearer moves his lips silently and his words appear on screen.
Who: The Subvocal Company, founded by Georgia Tech computer science graduates in Y Combinator's Fall 2026 batch. The idea began at a hackathon in February. Its launch video drew more than 12 million views on X.
How it works: no camera, no microphone. The glasses send low-power radio waves through the face and measure how they change as your muscles move while you silently articulate words. Models decode those signals directly into text.
The catch, and it's a big one:
- The company claims over 200 words per minute with any vocabulary. Published research on this kind of radio sensing has topped out around a 50-word vocabulary. That's a huge gap.
- No independent testing yet.
- The prototype looks 3D-printed; the founders say the shipping product will look far better.
- Preorders take a $199 refundable deposit toward a $499 price, with shipping estimated in Q1 2027.
Silent dictation in a coffee shop or on a plane would be wonderful. Wait for an independent test before putting money down.
Part 4: 3D, animation and robots
9. Budget3D: you choose how hard it thinks
What it is: an open image-to-3D model. Give it one image, get a 3D object back.
What's different: a slider that sets how much computing it spends. Turn it down and it runs about twice as fast with less memory, at a quality cost. Most such models spend equal effort everywhere, including the empty space around the object. Budget3D identifies boring regions, starting with empty space, lumps them together and spends its effort where the detail is.
Hands-on: a cowboy toy and a toy train looked good at full budget and visibly degraded at 40%. A custom upload came out decent from the front, with a flat back since the model can't see behind the object.
The catch: there are speed figures but no published quality metrics per budget, so you judge by eye. It needs under 8 GB of VRAM, so most recent gaming cards work. Code is on GitHub, weights on Hugging Face, and there's a free browser demo.
10. Nvidia MotionBricks: animation without animators
What it is: a real-time generative motion model from Nvidia research. You steer a character like a game character, and every movement is generated on the spot, with no handmade transitions.
Why it matters: in a normal game, animators build hundreds of clips (walk, run, jump, sit) and then experts wire them together so motion flows. MotionBricks learned from about 350,000 motion clips and handles all the in-between itself.
It has two parts:
- Movement: give it a speed, direction and style, and switch styles on the fly: zombie walk, injured leg, skipping.
- Objects: mark a few key poses, such as a hand on a sword or sitting on a bench, and it fills in the approach, contact and follow-through.
The same model drives robots: it's the motion layer inside Nvidia's GR00T whole-body control for humanoids.
The catch: no Unreal Engine plug-in yet, but code and trained models are on GitHub. For solo game developers, it's a glimpse of character animation without an animator budget.
11. Flash ALOHA: robots that catch
What it is: a Stanford two-armed robot system that tosses and catches a ball with nobody controlling it. Catching is hard because there's no time to think.
The insight: robots usually learn by watching a human steer them through a remote-control rig, which gets clumsy at speed, so robots end up slower than people. Flash ALOHA uses arms you can grab and move by hand. A person does the task at full speed, the robot records it, and small software smooths jerky spots on playback.
What it can do: catch, fold shorts, insert batteries into a device and clean a table, each learned from just 50 demonstrations using regular cameras.
The catch: reported success is over 80%, so it can still miss one in five. "Close to human speed" isn't precisely defined. But the code is on GitHub with a full hardware list for builders.
12. PhoneBot: a $400 walking robot with an old phone for a brain
What it is: an open, low-cost humanoid from a UCLA robotics lab, about a foot and a half tall, whose brain is an Android phone.
How it works:
- The phone's camera is the robot's eyes.
- The phone's motion sensor, the one that rotates your screen, keeps it balanced.
- The phone runs the walking policy and sends commands over Wi-Fi to 13 small servo motors.
- It connects to an online AI service for higher-level intelligence.
- Everything else is about $400 of parts: motors, a battery and a 3D-printed body. Tested phones were cheap prepaid Androids costing $50 to $80.
It walks forward and backward, turns, follows a person using the camera and gets itself back up after falling on its face. It learned to walk in simulation by trial and error: points for staying upright and keeping pace, penalties for dragging its feet.
The catch: at the time of testing, the parts list and CAD files were posted but the printable parts and assembly tutorial links weren't live yet. It walks only on flat ground and has no arms.
Why this matters
Put these together and a picture forms.
- The real world is becoming AI's canvas. Simulon, Reshoot 4D and MovieEdit all work on footage of reality rather than generating everything from nothing.
- Voice crossed the uncanny valley. Dessa, Haru-4B and Live 3 show natural, interruptible conversation going from lab to open model to product in weeks.
- Interfaces are getting weirder. Silent-speech glasses may or may not work as claimed, but the race to replace typing and talking is on.
- Robotics is getting cheap and open. MotionBricks' motion layer, a natural voice model and a $400 phone-brained robot are all public building blocks.
Practical takeaways
- Try today: Simulon (iPhone for reactions), Hy-Nomas, Haru-4B, Fish Audio Live 3's demo, Budget3D's browser demo, and code for MotionBricks, Flash ALOHA and Reshoot 4D.
- Watch, don't buy yet: MovieEdit and Dessa aren't available; Subvocal is unverified.
- Guard against voice scams. Agree on a family code word and verify unexpected calls asking for money, whatever the voice sounds like.
- Check hardware needs. Several open tools need an Nvidia GPU; Budget3D needs under 8 GB of VRAM.
Counterpoints and risks
- Demo reels flatter. Most claims come from creators' own demos and project pages.
- Hype gaps. Subvocal's 200 words per minute versus a 50-word research ceiling is the clearest example.
- Human-sounding AI cuts both ways. Dessa's realism is exactly what makes it dangerous for fraud.
- "Open" often means "for experts." ComfyUI, GPUs and GitHub setup still gate many tools.
FAQ
What is Block's Dessa?
A research-preview conversational voice model trained only on recordings of human speech, with no text or pretrained language model. It listens and talks simultaneously. Access is limited to selected AI safety researchers.
Can I download a full-duplex voice model now?
Yes. Higurashi's Haru-4B is an open-weight base model on Hugging Face, meant for developers to build on.
Do Subvocal's silent-speech glasses work?
Unknown. The company's claims exceed published research, and no independent tests exist yet. Preorders are a refundable $199 deposit toward $499.
What is PhoneBot?
A UCLA open-source humanoid robot about 18 inches tall that uses an Android phone as its camera, balance sensor and computer, with roughly $400 of other parts.
Is Tencent's world model free?
Yes. Hy-Nomas code and weights are open and allowed for commercial use, though real-time performance needs serious GPUs.
The bottom line
The headlines track the giants. The future often shows up first in the small stuff: a voice that accidentally invents a family, glasses that may read your lips by radio, a robot built from a drawer phone.
Not all of these will pan out. But together they show AI leaving the chat window and moving into cameras, voices, faces and bodies. That's the part worth watching.
Sources
- Dessa: the conversation model trained only on speech (Block)
- Block's new voice model sounds eerily human (TechCrunch)
- Haru-4B model card (Hugging Face)
- The Subvocal Company (Y Combinator)
- Hy-Nomas: Towards Free-form Interactive World Generation (arXiv)
- MotionBricks project page (Nvidia Research)
- PhoneBot: A Low-Cost Open Humanoid Robot Platform Reusing Smartphones (arXiv)





