This video might help explain 3D Gaussian splatting.
https://www.youtube.com/watch?v=wKgMxrWcW1s
Essentially, an entirely new graphics pipeline with different fundamental techniques which allow for high performance and fidelity compared to... what we did before(?)
Cool.
It’s amazing tech, it’s just a solution looking for a problem.
It feels a bit like the original Segway’s over-engineered solution versus cheap Chinese hoverboards, then the scooters and e-bikes that took over afterwards.
Why would I be paying all this money for this realistic telepresence when my shitbox HP laptop from Walmart has a perfectly serviceable webcam?
I'm curious about practical application in everyday life of these avatars - but in the real life, not examples provided by marketing department. With that price Vision Pro still feels like a toy for wealthy people, or perhaps for CEOs of companies who can afford conferences in virtual environment. But then exactly, why? Majority of world tested during pandemic video calls, conferences and all sorts of other activities like virtual crowds for tv programs (pretty sure British panel shows shown grids of people as a substitute for studio audience). News services were inviting their guests by video calls when Skype was still around.
I always viewed the current generations of 'cheap Chinese hovebords' etc being a direct descendant of the Segway, and that Kamen and his believers weren't quite as ridiculous as we thought them at the time - they were just ahead of their time and expecting too much from too low a point in the technology curve.
Many use cases come to mind. If (retinal?) identities were private, encrypted, and “anonymized” in handshake:
web browsing without captchas, anubis, bot tests, etc. (“human only” internet, maybe like Berners-Lee’s “semantic web” idea [1][2])
Non “anonymized”:
non-jury court and arbitration appearances (with expansion of judges to clear backlogs [3])
medical checkups and social care (eg neurocognitive checkups for elderly, social services checkins esp. children, checkins for depressed or isolated needing offwork social interactions, etc.)
bureaucratic appointments (customer service by humans, DMV, building permits, licenses, etc.)
web browsing for routine tasks without logins (banks, email, etc)
CorridorDigital recently used the tech to assist in remaking the rooftop bullet-time scene from The Matrix. It's used for making the environment instead of modeling it from scratch.
They also had an earlier video that more heavily featured gaussian splats. Using them to recreate the inside of the universal studios theme park without permission. I was very impressed with how it handles reflections on glass.
1. The scanning is fast, it takes longer to set up a fingerprint on a macbook air. Just turning the head from side to side, then up and down, smiling and raising one's eyebrows.
2. I used the M5, and the processing time to generate the persona was quick. I didn't time it, but it felt like less than 10 seconds.
3. My cheeks tend to restrict smiling while wearing the headset, it works but people that know me understood what I meant when I said my smile was hindered.
4. Despite the limited actions used for set up, it reproduces a far greater range of facial movements. For example if I do the invisible string trick, it captures my lips correctly (when you move the top lip in one direction and the lower lip in the opposite direction, as if pulled by a string.)
5. I wasn't expecting this big of a jump in quality from the v1.
How's the latency? Latency is what makes Zoom et al painful for me now - it ruins the ability to politely interject, give confirmatiom, etc. Does Apple do a better job of this than Google/Zoom? In theory you could get 20-30ms (just spitballing numbers I used to get playing shooters!) but i've never got anywhere near that with vid conferencing.
Even so, latency-in-zoom kind of becomes an attribute of the medium and you learn to adapt. How does it feel with the Vision Pro though? The article talks about a really convincing sense of being in the same place with someone - how does latency affect that? (And does it differ based on if you're all physically in Silicon Valley or not?)
I would assume any added latency is negligible -- the sensors + interpretation + rendering should be very fast.
But you've still got all the network latency including Wi-Fi latency on both ends. And you always need a small audio buffer so discrete network packets can be assembled into continuous audio without gaps.
So I wouldn't expect this latency to be any different from regular videoconferencing.
The laws of physics means that the longer the path for your network packet, the higher the latency.
One way latency on the Internet across fiber is about 4μs to 5μs per kilometer in my experience.
For example, SF to Paris is ~40ms one way (it used to be 60ms 15y ago, latency and jitter have really improved).
Double those values for the round trip allowing you to interject in a conversation.
Add wifi, which has terrible latency with a lot of jitter (1ms to 400ms jitter is not uncommon). Wi-Fi 7 should reduce the jitter and latency in theory. We shall see improvements in the coming decade. Cellphone 5G did improve latency for me, so I don't doubt WiFi will eventually deliver.
In other words you need to be within 3Mm (3000km) away to get a chance at a 30ms roundtrip. And that's assuming peer to peer without wifi nor slow devices.
For a conference call, everybody connects to a central server acting as the relay. So now the latency budget is halved already.
What is missing from the article is that creating a model from a few pictures is not that hard (well it is to do well, but hear me out)
The difficult part is animating it realistically with the sensors you have, in real time.
Extracting signal from eye-gaze cameras with a sighlty wider field of view, that allows realistic not not uncanny valley animation is quite hard to do on the general public Peoples faces are all different sizes and shapes, to the point that even getting accraute gaze vectors is hard, let alone smile and check position (those are done with different cameras, not just eye gaze. )
Came this close to buying an AVP, before learning that they only mirror a single screen with no virtual monitors.
Like, guize, c'mon. Virtual desktop can do three. For 3.5k you gotta do better. I don't particularly need a virtual me in space as much as I need more screens that can do, like, actual work.
Norris's dodge on iPhone scanning is telling—processing on-device keeps it secure and magical, but imagine Personas popping up in FaceTime cameos or ARKit bridges to iPads. How soon until we see cross-device ecosystems like Microsoft's Mesh, but with Apple's polish? Eager for that affordability leap; until then, thanks for the vivid demo, Scott—now I need a Vision Pro buddy just to test this out.
28 comments
[ 2.4 ms ] story [ 112 ms ] threadJust in time for Vision Pro to go big. Right?
It feels a bit like the original Segway’s over-engineered solution versus cheap Chinese hoverboards, then the scooters and e-bikes that took over afterwards.
Why would I be paying all this money for this realistic telepresence when my shitbox HP laptop from Walmart has a perfectly serviceable webcam?
To me it would be a shortcoming of the device if I couldn't show me and the thing I'm working on at the same time.
web browsing without captchas, anubis, bot tests, etc. (“human only” internet, maybe like Berners-Lee’s “semantic web” idea [1][2])
Non “anonymized”:
non-jury court and arbitration appearances (with expansion of judges to clear backlogs [3])
medical checkups and social care (eg neurocognitive checkups for elderly, social services checkins esp. children, checkins for depressed or isolated needing offwork social interactions, etc.)
bureaucratic appointments (customer service by humans, DMV, building permits, licenses, etc.)
web browsing for routine tasks without logins (banks, email, etc)
[1] <https://www.newyorker.com/magazine/2025/10/06/tim-berners-le...> [2] <https://newtfire.org/courses/introDH/BrnrsLeeIntrnt-Lucas-Nw...> [3] <https://nysfocus.com/2025/05/30/uncap-justice-act-new-york-c...>
https://www.youtube.com/watch?v=iq5JaG53dho&t=2s
https://www.youtube.com/watch?v=cetf0qTZ04Y
The previous beta ones were terrifying frankenstein monsters. The new ones fooled my boss for 30 minutes.
There's a bit of uncanny valley left, nevertheless. My persona's smile reminds of the horrible expressions people like to make in Source Filmmaker.
1. The scanning is fast, it takes longer to set up a fingerprint on a macbook air. Just turning the head from side to side, then up and down, smiling and raising one's eyebrows.
2. I used the M5, and the processing time to generate the persona was quick. I didn't time it, but it felt like less than 10 seconds.
3. My cheeks tend to restrict smiling while wearing the headset, it works but people that know me understood what I meant when I said my smile was hindered.
4. Despite the limited actions used for set up, it reproduces a far greater range of facial movements. For example if I do the invisible string trick, it captures my lips correctly (when you move the top lip in one direction and the lower lip in the opposite direction, as if pulled by a string.)
5. I wasn't expecting this big of a jump in quality from the v1.
Perhaps how their heads, eyes move with this weird "fluid" effect and way too much blurred faces?
There's regular latency due to distance, just like on a phone call if you're chatting with someone halfway across the world.
But on a normal connection, audio and the persona should always be in sync, the same way audio and video are over Zoom or FaceTime.
There shouldn't be any extra latency for the audio only.
Even so, latency-in-zoom kind of becomes an attribute of the medium and you learn to adapt. How does it feel with the Vision Pro though? The article talks about a really convincing sense of being in the same place with someone - how does latency affect that? (And does it differ based on if you're all physically in Silicon Valley or not?)
But you've still got all the network latency including Wi-Fi latency on both ends. And you always need a small audio buffer so discrete network packets can be assembled into continuous audio without gaps.
So I wouldn't expect this latency to be any different from regular videoconferencing.
One way latency on the Internet across fiber is about 4μs to 5μs per kilometer in my experience.
For example, SF to Paris is ~40ms one way (it used to be 60ms 15y ago, latency and jitter have really improved).
Double those values for the round trip allowing you to interject in a conversation.
Add wifi, which has terrible latency with a lot of jitter (1ms to 400ms jitter is not uncommon). Wi-Fi 7 should reduce the jitter and latency in theory. We shall see improvements in the coming decade. Cellphone 5G did improve latency for me, so I don't doubt WiFi will eventually deliver.
In other words you need to be within 3Mm (3000km) away to get a chance at a 30ms roundtrip. And that's assuming peer to peer without wifi nor slow devices.
For a conference call, everybody connects to a central server acting as the relay. So now the latency budget is halved already.
What is missing from the article is that creating a model from a few pictures is not that hard (well it is to do well, but hear me out)
The difficult part is animating it realistically with the sensors you have, in real time.
Extracting signal from eye-gaze cameras with a sighlty wider field of view, that allows realistic not not uncanny valley animation is quite hard to do on the general public Peoples faces are all different sizes and shapes, to the point that even getting accraute gaze vectors is hard, let alone smile and check position (those are done with different cameras, not just eye gaze. )
https://www.youtube.com/live/ucRukZM0d1s?t=1h1m50s
https://zju3dv.github.io/freetimegs/
https://www.4dv.ai/
The videos can be played back in real-time, though they require multiple cameras to capture.
Like, guize, c'mon. Virtual desktop can do three. For 3.5k you gotta do better. I don't particularly need a virtual me in space as much as I need more screens that can do, like, actual work.