Download "They Just Built a New Form of AI, and It’s Better Than LLMs"
%3Aformat(webp)%2Fi.ytimg.com%252Fvi_webp%252FIpOiRmAEle8%252Fmaxresdefault.webp&w=1139&q=75)
Similar videos from our catalog
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252FCRraHg4Ks_g%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi_webp%252FF2Mx-u7auUs%252Fmaxresdefault.webp&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi_webp%252FMyLUiE-XfQI%252Fmaxresdefault.webp&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi_webp%252FRpvf1BQzJeg%252Fmaxresdefault.webp&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252FgXyudlJCVFs%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252FjSdHmImyUjk%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi_webp%252FseBndOROJt4%252Fmaxresdefault.webp&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi_webp%252FzkafOIyQM8s%252Fmaxresdefault.webp&w=1139&q=75)
Video tags
Subtitles
00:00:02
All right. So today I want to talk about00:00:04
something coming out of Metapare led by00:00:07
Yan Lun and his [music] team that shows00:00:09
a very different kind of AI taking00:00:12
shape. It's not built around generating00:00:13
text or chasing better word prediction00:00:15
at all. Instead, it drops the idea that00:00:18
intelligence has to revolve around00:00:20
producing words and focuses on00:00:22
predicting meaning directly. And once00:00:24
you really see how this works, it00:00:26
honestly feels like what comes after the00:00:28
LLM era, not just another upgrade on top00:00:30
of it. What they've built is called00:00:32
VLJA, short for vision language joint00:00:35
embedding predictive architecture. The00:00:37
name sounds heavy, but the idea behind00:00:38
it is surprisingly straightforward. And00:00:41
once it clicks, a lot of design choices00:00:43
we've all accepted in modern AI suddenly00:00:45
start to feel inefficient [music] or at00:00:47
least unnecessary. To understand why00:00:49
this matters, you have to look at how00:00:51
vision language models usually work00:00:53
today. Right now, most VLMs follow the00:00:56
same basic [music] pattern. You show00:00:57
them an image or a video. You give them00:00:59
a prompt or a question, and they respond00:01:01
by generating text one token at a time.00:01:04
That's how image captioning works.00:01:06
That's how visual question answering00:01:07
works. That's how most large multimodal00:01:10
models operate under the hood. They're00:01:12
trained to predict the next word, then00:01:14
the next word, then the next word,00:01:16
again. And again. That approach clearly00:01:18
works, but it comes with some hidden00:01:20
problems. The first issue is that these00:01:22
models are forced to learn a lot of00:01:24
things that don't actually matter for00:01:26
correctness. Take a simple example. If00:01:28
you ask, "What happens if I flip this00:01:30
light switch down?" There are many00:01:32
answers that are all perfectly fine. The00:01:34
light turns off. The room gets darker.00:01:37
The lamp goes dark. Humans immediately00:01:39
understand that these all describe the00:01:41
same outcome. But for a token-based00:01:43
model, those answers are totally00:01:45
different. They're different sequences00:01:46
of symbols with almost no overlap.00:01:49
During training, the model has to learn00:01:51
exact phrasing, word choice, and00:01:53
sentence structure. Even though none of00:01:55
that changes the meaning, a huge amount00:01:57
of training effort goes into modeling00:01:59
surface level language variation instead00:02:01
of the underlying idea. The second issue00:02:04
shows up when you try to use these00:02:05
systems in real time. Token by token00:02:07
generation is slow and awkward. If00:02:09
you're dealing with live video,00:02:10
wearables, robotics, or anything [music]00:02:13
that needs to understand what's00:02:14
happening continuously. You can't really00:02:16
know what the model means until it00:02:18
finishes generating [music] text. The00:02:20
semantics only appear at the end of the00:02:22
decoding process. That adds latency,00:02:25
burns compute, [music] and makes it hard00:02:26
to update information selectively. This00:02:29
is exactly where VLJA takes a completely00:02:31
different route. Instead of predicting00:02:33
words, VLJA predicts embeddings. These00:02:36
are continuous vectors that represent00:02:38
meaning directly, not the surface form00:02:40
of language. During training, the model00:02:43
never tries to generate text at all. It00:02:45
learns to map visual input and a text00:02:47
query straight into a semantic00:02:49
representation of the answer. The system00:02:51
is built from four main components, and00:02:53
each one has a clear role. First,00:02:56
there's the visual encoder. This takes00:02:58
an image or a sequence of video frames00:03:00
and compresses it into a set of visual00:03:02
embeddings. [music] You can think of00:03:04
these as visual tokens except they're00:03:06
continuous vectors rather than discrete00:03:08
symbols. In this setup, they use VJEPA00:03:11
2, a self-supervised vision transformer00:03:13
with around 304 million parameters and00:03:16
it stays frozen during training. Next00:03:18
comes the predictor which is the core of00:03:20
the whole system. This module takes the00:03:22
visual embeddings and the text query00:03:24
[music] like a question or prompt and00:03:26
predicts what the answer embedding00:03:27
should look like. It's built using00:03:29
transformer layers initialized from00:03:31
llama 3.21b but without causal masking.00:03:35
That means everything can attend to00:03:36
everything else. Vision and text00:03:38
interact freely. Then there's the Y00:03:40
encoder. This encodes the target text00:03:42
during training the correct [music]00:03:44
answer into an embedding. That embedding00:03:46
becomes the learning target.00:03:47
Importantly, this representation is00:03:49
meant to capture the meaning of the00:03:51
answer, not the exact wording. Finally,00:03:53
there's the Y decoder, and this part is00:03:56
barely involved. It doesn't participate00:03:58
in training at all. At inference time,00:04:00
it only gets used when you actually need00:04:02
readable text. Most of the time, the00:04:04
model stays entirely in embedding space.00:04:07
Training works in a simple loop. You00:04:09
give the model a visual input, a query,00:04:11
and a target answer. The Yen encoder00:04:13
turns the answer into an embedding. The00:04:15
predictor tries to produce [music] that00:04:17
same embedding from the visual input and00:04:19
query. The loss is computed directly in00:04:21
embedding space, not in token space.00:04:24
What matters here is how the model00:04:25
learns without everything collapsing00:04:27
into noise. Vla is trained so that its00:04:30
predicted meaning is pulled toward the00:04:31
correct meaning while different answers00:04:33
are kept clearly separated. In practice,00:04:35
this forces the system to build a00:04:37
structured semantic space. Similar00:04:39
answers cluster together naturally.00:04:41
Different answers stay far apart.00:04:43
Instead of memorizing phrasing, the00:04:45
model organizes meaning itself, which00:04:47
keeps the whole representation stable00:04:49
and useful. This leads to the key00:04:51
insight behind the whole approach. In00:04:53
token space, multiple valid answers can00:04:55
be extremely far apart. In embedding00:04:57
space, those same answers can sit close00:04:59
together. That turns a messy multimodal00:05:02
learning problem into a clean single00:05:04
mode one. The model no longer has to00:05:06
guess which wording you want. It just00:05:08
has to understand what the answer means.00:05:10
Because of this, VLJA doesn't need a00:05:13
heavy language decoder during training.00:05:15
It's not learning how to write00:05:16
sentences. It's learning how to predict00:05:18
semantics. And that change alone cuts a00:05:21
huge amount of unnecessary work out of00:05:23
the system. And you can see it clearly00:05:24
in the results. To test whether this00:05:26
idea actually holds up, the researchers00:05:28
ran a rare kind of comparison where00:05:30
almost nothing was allowed to change.00:05:32
Same vision encoder, same resolution,00:05:35
same frame rate, same data mixture, same00:05:37
batch size, same number of training00:05:40
steps. The only difference was what the00:05:42
models were trained to predict. One00:05:43
model followed the standard route,00:05:45
predicting tokens with a 1 billion00:05:47
parameter language model. The VLJ00:05:49
version predicted embeddings using a00:05:51
roughly 500 million parameter predictor.00:05:54
So right away, the embedding based00:05:56
system had about half the trainable00:05:57
parameters. Early in training, the two00:06:00
systems look similar. After around00:06:02
500,000 samples, performance is roughly00:06:04
comparable. But as training continues, a00:06:06
clear pattern emerges. VLJA starts00:06:09
improving faster and it keeps improving.00:06:12
After 5 million samples, it reaches00:06:14
around 14.7 CR on video captioning,00:06:17
while the token-based model is still00:06:19
around 7.1. Classification accuracy00:06:22
jumps to about 35% top five for VLJA00:06:25
versus roughly 27% for the baseline. And00:06:28
the gap doesn't close later. At 1500:06:30
million samples, the difference remains.00:06:32
VLJa continues to learn more00:06:34
efficiently, even with fewer parameters.00:06:36
That's not a tuning trick. That's a00:06:38
structural advantage. The story doesn't00:06:40
stop at training efficiency either.00:06:42
Inference is where this approach really00:06:44
starts to shine, especially for video.00:06:46
Because VLJA produces a continuous00:06:48
stream of semantic embeddings, it00:06:50
supports something called selective00:06:52
decoding. Instead of generating text at00:06:54
fixed intervals, you monitor how the00:06:56
embeddings change over time. If the00:06:58
meaning stays stable, you don't decode00:07:00
anything. If there's a significant00:07:01
semantic shift, then you decode. They00:07:04
test this on long procedural videos from00:07:06
Ego XO4D. These videos average about 600:07:09
minutes each and contain roughly 14300:07:12
action annotations per video. Decoding00:07:14
text is the expensive part. So, the goal00:07:16
is to recover the annotation sequence00:07:18
while minimizing how often decoding00:07:20
happens. They compare two strategies.00:07:23
Uniform decoding where text is generated00:07:25
at fixed time intervals and embedding00:07:27
guided decoding where the embedding00:07:29
stream is clustered into semantically00:07:31
coherent segments and decoded once per00:07:33
segment. The result is clean. To match00:07:35
the performance of uniform decoding at00:07:37
one decode per second, VLJA only needs00:07:40
to decode about once every 2.85 seconds.00:07:43
That's roughly a 2.85 times reduction in00:07:46
decoding operations with similar side ER00:07:49
scores. No fancy memory tricks, no KV00:07:52
cache gymnastics. It's just a00:07:54
consequence of working in semantic00:07:56
space. This is especially important for00:07:58
real-time systems like smart glasses,00:08:00
robotics, navigation or live planning00:08:02
where latency and compute cost actually00:08:05
matter. Another major advantage is00:08:06
versatility. VLJa can handle generation,00:08:10
classification, retrieval and00:08:12
discriminative visual question answering00:08:14
using the same architecture. There are00:08:16
no task specific heads and no separate00:08:18
models. For open vocabulary00:08:20
classification, candidate labels are00:08:22
encoded into embeddings and compared to00:08:25
the predicted embedding. The closest00:08:26
match wins. For texttovide retrieval,00:08:29
the text query is encoded and videos are00:08:31
ranked by similarity. For discriminative00:08:34
VQA, all candidate answers are embedded00:08:36
and the nearest one is selected. They00:08:38
evaluate this across a wide set of00:08:40
benchmarks on eight video classification00:08:42
data sets and eight texttovideo00:08:44
retrieval data sets. The base VLJA model00:08:47
with 1.6 billion parameters and only00:08:50
about 2 billion training samples00:08:51
outperforms clip SIG lip 2 and00:08:54
perception encoder on average. Some of00:08:56
those baselines have seen up to 8600:08:58
billion samples. After supervised00:09:01
fine-tuning, the VLJA SFT model improves00:09:04
even further. It's no longer strictly00:09:06
zeroot, but as a single generalist00:09:08
model, it approaches specialist systems00:09:10
that are tuned individually for each00:09:12
data set. On visual question answering,00:09:15
the results are especially telling. They00:09:17
evaluate on GQA for compositional00:09:19
reasoning, tally QA for complex00:09:21
counting, and POPE and POPE v2 for00:09:23
hallucination detection. Vla SFT with00:09:26
1.6 6 billion parameters, lands in the00:09:29
same range as models like Instruct Blip00:09:31
and QuenVL, many of which rely on much00:09:34
larger backbones and multi-stage00:09:35
instruction tuning. It doesn't dominate00:09:37
every benchmark, but the fact that it's00:09:39
competitive at all is important because00:09:41
it's not a classic generative VLM. It00:09:43
answers questions by comparing meaning,00:09:45
not by generating free form text. Then00:09:48
there's the world modeling experiment.00:09:49
Here the model is shown an initial image00:09:52
and a final image and has to choose00:09:54
which action caused the transition from00:09:56
four candidate video clips. This is00:09:58
closer to understanding physical00:10:00
causality than language generation. Vla00:10:02
SFT reaches 65.7%00:10:05
accuracy setting a new state-of-the-art.00:10:08
It outperforms larger vision language00:10:10
models and even beats frontier language00:10:12
models like GPT40, Claude 3.5, and00:10:16
Gemini 2, which rely on captioning and00:10:18
textbased reasoning. That result00:10:20
matters. It suggests that directly00:10:22
predicting latent semantics can be more00:10:24
effective than narrating the world in00:10:26
words and reasoning over those words00:10:28
afterward. They also analyze the quality00:10:30
of the text embeddings themselves using00:10:32
hard negative benchmarks like Sugar00:10:34
Crate++ and Vizla. They test whether the00:10:36
Y encoder can detect subtle semantic00:10:39
changes like swapped attributes or00:10:41
altered relationships. The base VLJA Yen00:10:44
encoder outperforms clip SIGLIP 2 and00:10:47
perception encoder indicating a sharper00:10:50
and more structured semantic space.00:10:52
Finally, they stress test the system00:10:54
through ablations. When the large00:10:55
captionbased pre-training stage is00:10:57
removed, performance drops sharply,00:10:59
especially for classification and00:11:01
retrieval. Breezing the Y encoder hurts00:11:03
alignment. Overly simple training00:11:05
objectives weaken learning. Larger00:11:07
predictors help, particularly for VQA.00:11:10
Visually aligned text encoders00:11:11
consistently boost retrieval and00:11:13
classification. The pattern is00:11:15
consistent. When components that support00:11:17
semantic learning are strengthened, the00:11:19
model improves. When they're removed, it00:11:22
degrades. That kind of behavior is00:11:24
exactly what you want to see from a00:11:25
system that's meant to scale. VLJA isn't00:11:28
trying to replace language models00:11:30
everywhere. Tasks like deep reasoning,00:11:32
tool use, and agent style planning still00:11:35
favor token-based systems. But for00:11:37
perceptionheavy problems, especially00:11:39
those involving video, real-time input,00:11:41
and continuous understanding of the00:11:43
world, this approach fits naturally. It00:11:45
shifts the center of gravity from00:11:47
language to meaning. Words become an00:11:49
output option, not the core mechanism of00:11:51
intelligence. And that shift is what00:11:53
makes this work feel like more than just00:11:55
another model iteration. Thanks for00:11:56
watching, and I will catch you in the00:11:58
next one.Description:
AI is starting to move in a very different direction from what we’ve gotten used to. Instead of chasing bigger language models and better text generation, the focus is shifting toward systems that operate on meaning itself. Words stop being the center. Semantics, vision, video, and real-time understanding take priority. In this video, we break down a new AI architecture from Meta FAIR, led by Yann LeCun and his team, that takes this approach head-on and shows why it feels like what comes after LLMs. 📩 Brand Deals and Partnerships: me@faiz.mov ✉ General Inquiries: airevolutionofficial@gmail.com 🧠 What You’ll See • The paper: https://arxiv.org/abs/2512.10942 • VL-JEPA architecture explained in simple terms • Why predicting meaning beats predicting words • How this model works without token-by-token generation • Why it performs better on vision and video tasks • What this means for real-time AI systems 🚨 Why It Matters AI has been moving fast, but most progress has been tied to generating better text. This shift moves AI closer to real-world understanding, where systems react to what they see, track change over time, and operate with lower latency and lower cost. When meaning becomes the core output instead of words, AI starts becoming something that can actually run continuously in the world.
Mediafile available in formats
* — If the video is playing in a new tab, go to it, then right-click on the video and select "Save video as..."
** — Link intended for online playback in specialized players
Questions about downloading video
How can I download "They Just Built a New Form of AI, and It’s Better Than LLMs" video?
http://univideos.ru/ website is the best way to download a video or a separate audio track if you want to do without installing programs and extensions.
The UDL Helper extension is a convenient button that is seamlessly integrated into YouTube, Instagram and OK.ru sites for fast content download.
UDL Client program (for Windows) is the most powerful solution that supports more than 900 websites, social networks and video hosting sites, as well as any video quality that is available in the source.
UDL Lite is a really convenient way to access a website from your mobile device. With its help, you can easily download videos directly to your smartphone.
Which format of "They Just Built a New Form of AI, and It’s Better Than LLMs" video should I choose?
The best quality formats are FullHD (1080p), 2K (1440p), 4K (2160p) and 8K (4320p). The higher the resolution of your screen, the higher the video quality should be. However, there are other factors to consider: download speed, amount of free space, and device performance during playback.
Why does my computer freeze when loading a "They Just Built a New Form of AI, and It’s Better Than LLMs" video?
The browser/computer should not freeze completely! If this happens, please report it with a link to the video. Sometimes videos cannot be downloaded directly in a suitable format, so we have added the ability to convert the file to the desired format. In some cases, this process may actively use computer resources.
How can I download "They Just Built a New Form of AI, and It’s Better Than LLMs" video to my phone?
You can download a video to your smartphone using the website or the PWA application UDL Lite. It is also possible to send a download link via QR code using the UDL Helper extension.
How can I download an audio track (music) to MP3 "They Just Built a New Form of AI, and It’s Better Than LLMs"?
The most convenient way is to use the UDL Client program, which supports converting video to MP3 format. In some cases, MP3 can also be downloaded through the UDL Helper extension.
How can I save a frame from a video "They Just Built a New Form of AI, and It’s Better Than LLMs"?
This feature is available in the UDL Helper extension. Make sure that "Show the video snapshot button" is checked in the settings. A camera icon should appear in the lower right corner of the player to the left of the "Settings" icon. When you click on it, the current frame from the video will be saved to your computer in JPEG format.
How do I play and download streaming video?
- in video formats, hover your mouse over "Streaming Video**";
- right-click on "Copy link";
- open VLC-player;
- select Media - Open Network Stream - Network in the menu;
- paste the copied link into the input field;
- click "Play".
- copy the video address (URL);
- select "Open Network Stream" in the "Media" item of VLC player and paste the link to the video into the input field;
- click on the arrow on the "Play" button and select "Convert" in the list;
- select "Video - H.264 + MP3 (MP4)" in the "Profile" line;
- click the "Browse" button to select a folder to save the converted video and click the "Start" button;
- conversion speed depends on the resolution and duration of the video.
For this purpose you need VLC-player, which can be downloaded for free from the official website https://www.videolan.org/vlc/.
How to play streaming video through VLC player:
To download streaming video via VLC player, you need to convert it:
Warning: this download method no longer works with most YouTube videos.
What's the price of all this stuff?
It costs nothing. Our services are absolutely free for all users. There are no PRO subscriptions, no restrictions on the number or maximum length of downloaded videos.