Latest / Elon Musk Podcast / A wild week in AI

Transcript

A wild week in AI · Elon Musk Podcast

• 0:00
  Welcome back to the Elon Musk Podcast.
• 0:02
  I'm thrilled to share some exciting news with you over the
• 0:05
  next two weeks. We're evolving.
• 0:07
  We'll be broadening our focus to cover all the tech Titans
• 0:11
  shaping our world. And with that, our show will
• 0:14
  become Stage 0. You'll still get the latest
• 0:17
  insights on Elon Musk, plus so much more, so stay tuned for our
• 0:22
  official relaunch at Stage 0 coming soon.
• 0:26
  So for the past four or five years, I've been bringing you in
• 0:28
  depth, no nonsense insights from the world of Elon Musk.
• 0:33
  But I need your help to keep the show alive and growing.
• 0:36
  If you'd love what you hear, consider supporting Stage 0 on
• 0:40
  patreon@patreon.com/stage 0 News.
• 0:45
  By joining our Patreon community, you'll get exclusive
• 0:48
  content, early access to some episodes, and a chance to shape
• 0:53
  future topics. Everyone has a voice, and your
• 0:56
  support goes directly into making this show absolutely
• 0:59
  better, and it helps me keep bringing you the cut that you
• 1:02
  enjoy every single day. If you're getting value from
• 1:06
  Stage 0 news, becoming a patron is the best way to make sure
• 1:10
  this journey keeps going. So let's make the next five
• 1:14
  years even bigger together. There's a link in the show notes
• 1:20
  just for you. Why would Google train an AI to
• 1:23
  talk to dolphins? That's not what you hear every
• 1:26
  day, and it's exactly the kind of question that sums up what
• 1:31
  might be the most unpredictable, chaotic, and fascinating week in
• 1:36
  AI so far this year. Not only are researchers now
• 1:40
  attempting to decode animal communication using lightweight
• 1:43
  neural networks, but we also saw humanoid robots, runner of
• 1:47
  literal half marathon, AI tools that can animate pets into
• 1:51
  dancers or bring comic panels to life with one click, and Open AI
• 1:55
  launching its most intelligent models yet.
• 1:58
  Every one of these stories is a question in itself.
• 2:02
  Why now, How does it work? And what does this mean for the
• 2:06
  future of interaction, creativity, and intelligence
• 2:09
  itself? Now to start, Google made
• 2:12
  headlines this week with something called Dolphin Gemma,
• 2:16
  a compact AI model trained to understand it even generate
• 2:20
  dolphin vocalizations. Now, what makes this unusual
• 2:24
  isn't just the implication, it's that the model runs directly on
• 2:28
  your phone. Researchers use Google Pixel
• 2:31
  devices to process real time dolphin chatter.
• 2:34
  Using a framework based on audio tokenization, they recorded
• 2:39
  every sound dolphins make, clicks, squawks, whistles, and
• 2:42
  converted them into discrete audio tokens using Google's
• 2:46
  Soundstream codec. This tokenized data set was then
• 2:50
  used to train a smaller variant of Google's JEMA model, which
• 2:54
  comes in at around 400 million parameters.
• 2:57
  That's small enough to run efficiently on mobile hardware
• 3:01
  without external computing and beyond understanding.
• 3:05
  The model can also synthesize new dolphin like sounds, which
• 3:08
  is a potential breakthrough for researchers aiming to translate
• 3:12
  interspecies communication into something that humans can
• 3:17
  eventually understand. Now the significance of this
• 3:20
  extends well beyond dolphins. The architecture is general
• 3:25
  enough that it could be retrained to understand and
• 3:27
  emulate vocalizations of other animals.
• 3:30
  And in theory we could eventually support real time two
• 3:33
  way communication with certain animal species.
• 3:37
  Could you imagine talking to your own pets?
• 3:40
  Now that poses the question into strange territories.
• 3:45
  If an AI can mimic a language like structure in dolphin
• 3:48
  chatter, are we on the verge of developing machine driven cross
• 3:52
  species translators now? While AI was speaking to marine
• 3:58
  life, animation tools were speaking to the internet's
• 4:02
  favorite content creators. Uni Animate DIT, which is a new
• 4:05
  plugin built for the open source model Animate DIF 1.2 allows
• 4:11
  users to animate any character image with a motion reference
• 4:14
  video up like a static image or any photo of a person, a cartoon
• 4:19
  or even a pet, and combine it with a short clip of someone
• 4:22
  dancing or moving around. The tool extracts the pose data
• 4:25
  from the video, then applies it to the static image, producing a
• 4:29
  fully animated clip with smooth transitions.
• 4:32
  And what stands out is that the model can guess unseen angles
• 4:37
  like the back of the character and animate flowing fabric or
• 4:40
  hand movements convincingly. All this runs locally with a
• 4:44
  minimum of 14 gigs of VRAM, meaning creators can use a tool
• 4:48
  without relying on cloud services.
• 4:51
  And I've used it. The results are surprisingly
• 4:53
  good. Even characters with complex
• 4:55
  appearances or unusual anatomy, like fictional anime designs or
• 5:00
  animals, can be animated with minimal artifacting.
• 5:04
  And since everything is released open source, artists and
• 5:07
  animators now have access to a new kind of puppetry.
• 5:10
  This is accessible. That's downloading a GitHub repo
• 5:14
  now. The companion tool emerged this
• 5:16
  week from Tencent as well. It's called instant character.
• 5:20
  It's focused is accuracy in reference based generation.
• 5:24
  So you have an image of a fictional character.
• 5:27
  Instant character can place the same character down to the
• 5:30
  facial structure, outfit details and accessories into a new
• 5:34
  scene. You can render them in a studio
• 5:37
  playing piano or walking in a snowstorm in full anime style.
• 5:41
  And the model is based on Flux, one of the highest fidelity open
• 5:45
  source diffusion models available, and it uses Laura
• 5:48
  adapters to style outputs in everything from Studio Ghibli to
• 5:53
  Makato Shinkai's signature look. Now, unlike most existing
• 5:58
  character transfer models, this one keeps attributes consistent
• 6:02
  across varied scenes and does so across 2D3D and photo realistic
• 6:08
  styles. So why this matters to you isn't
• 6:12
  just for cosplay creators or fan artist.
• 6:15
  In a world increasingly dominated by visualizations and
• 6:20
  virtual characters, the ability to preserve identity across
• 6:24
  generated media becomes critical, especially as avatars,
• 6:28
  V tubers and AI generated influencers grow more complex
• 6:32
  and integrated into media ecosystems.
• 6:36
  Now there's a new thing called Parkfield.
• 6:39
  It's a project from NVIDIA. It's focused on a very different
• 6:41
  kind of segmentation. This time it's 3 dimensions.
• 6:46
  It's a part segmentation model for 3D objects capable of
• 6:49
  breaking complex meshes into individual label components.
• 6:54
  Now think about 3D model of a robot or a car with part field.
• 6:58
  Each part, arm, leg, wheel, mirror is isolated into its own
• 7:02
  labeled region, enabling texture swaps, physical simulations, or
• 7:07
  animations to be applied to each section independently.
• 7:11
  This is obvious implications for anyone building physical
• 7:13
  simulations, robotics, or gaming assets.
• 7:17
  Compared to previous segmentation models, it not only
• 7:20
  performs better, but is much faster completing tasks in a
• 7:23
  fraction of the time thanks to more efficient tokenization and
• 7:28
  inference architecture. Now one more thing.
• 7:32
  A slightly surreal real world scene 1/2 marathon for humanoid
• 7:36
  robots just took place in Beijing.
• 7:39
  More than 20 companies from across China participated,
• 7:42
  entering bipedal robots that walked, jogged and ran across
• 7:46
  and around a race track in full view of cheering spectators.
• 7:49
  Some entries were clunky, barely able to maintain balance, while
• 7:53
  others like Unitaries G1 and Beijing Humanoid Innovation
• 7:57
  Centers Chiang Jiang Ultra, manage smoother gates and even
• 8:01
  completed longer runs now. Footage of this shows some
• 8:04
  robots falling or freezing mid run, but others powering through
• 8:08
  the full event now. Qiangyan Ultra in particular
• 8:13
  drew attention for its speed and its stability, suggesting real
• 8:17
  progress and bipedal locomotion design.
• 8:21
  The event might sound like a novelty though, but it reflects
• 8:24
  a changing reality. Robotic mobility is evolving
• 8:27
  fast enough that humanoids may soon be deployed in public
• 8:31
  facing or physical labor rolls. Holding a marathon may just be
• 8:35
  like a publicity stunt, but it's also a benchmark for these
• 8:38
  robots. Tests, endurance, adaptability,
• 8:42
  and real world stability, these traits that are notoriously
• 8:46
  difficult for machines to master.
• 8:51
  Hey, thank you so much for listening today.
• 8:53
  I really do appreciate your support.
• 8:55
  If you could take a second and hit subscribe or the follow
• 8:58
  button on whatever podcast platform that you're listening
• 9:00
  on right now, I greatly appreciate it.
• 9:02
  It helps out the show tremendously and you'll never
• 9:05
  miss an episode. And each episode is about 10
• 9:08
  minutes or less to get you caught up quickly.
• 9:11
  And please, if you want to support the show even more, go
• 9:14
  to patreon.com/stage Zero and please take care of yourselves
• 9:20
  and each other and I'll see you tomorrow.