Latest / Elon Musk Podcast / AI Update - New Free Video Generator, Bytedance, Google and More! - Sponsor : Fish.Audio

Transcript

AI Update - New Free Video Generator, Bytedance, Google and More! - Sponsor : Fish.Audio · Elon Musk Podcast

• 0:01
  Hey everyone, welcome back to the Elon Musk Podcast.
• 0:05
  I'm thrilled to share some exciting news with you over the
• 0:08
  next two weeks. We're evolving.
• 0:10
  We'll be broadening our focus to cover all the tech Titans
• 0:14
  shaping our world. And with that, our show will
• 0:16
  become Stage 0. You'll still get the latest
• 0:20
  insights on Elon Musk plus so much more.
• 0:23
  So stay tuned for our official relaunch at Stage 0 coming soon.
• 0:28
  I'd like to thank the sponsor of this show, Phish dot Audio, a
• 0:33
  great service. So if you're a podcaster, a
• 0:36
  voice over artist, or a content creator like myself and you're
• 0:39
  looking for hyper realistic AI generated voices, you need to
• 0:44
  check out Phish Audio. Ranked among the top 2 most
• 0:48
  realistic voice generators in 13 languages, Phish Audio lets you
• 0:52
  create high quality voice overs with just a few clicks.
• 0:57
  Now I'm going to be honest with you, sometimes I make mistakes
• 0:59
  during the show and I need to replace some words because I
• 1:02
  misspoke or something was in the background, there was some noise
• 1:06
  or something. I just can't get rid of it.
• 1:07
  So sometimes I use AI generated me in order to fill in those
• 1:11
  gaps. And I tried Phish Audio for this
• 1:14
  and it worked so good, so much better than the others that I've
• 1:19
  tried. And it's open source friendly.
• 1:21
  It's super fast, it's even faster than 11 labs if you're
• 1:24
  familiar with that. And best of all, it's
• 1:26
  affordable. It has the lowest prices on the
• 1:29
  market. So whether you need multilingual
• 1:32
  voice cloning, real time speech to speech conversion, or
• 1:36
  professional quality narration, Fish Audio has you completely
• 1:42
  covered with whatever you need done.
• 1:43
  Try it today for free at Fish dot Audio and see why it's the
• 1:49
  absolute best AI vice tool out there today.
• 1:53
  Go check them out at Fish dot Audio.
• 1:56
  That's Fish dot Audio fish dot audio.
• 2:00
  Head to Fish dot Audio and start creating with AI powered voices
• 2:05
  right now. Welcome to this week's AI
• 2:08
  update. So how many times can artificial
• 2:11
  intelligence reinvent itself in a single week?
• 2:14
  And what happens when it starts to reshape the very definition
• 2:18
  of creativity, design, and human likeness?
• 2:21
  This week in AI has introduced more than just flashy demos or
• 2:25
  iterative upgrades. It has opened a new chapter in
• 2:28
  how we visualize the future of interaction, media and autonomy.
• 2:33
  From image generation that understands multiple references
• 2:36
  to hyper realistic avatars that lip sync in multiple languages,
• 2:40
  to humanoid robots executing front flips and punches, all
• 2:44
  within a few days, the question is no longer what AI can do, but
• 2:50
  what it will leave for us to do. Bytedance, Google, Meta,
• 2:54
  Alibaba, Vivago AI and a constellation of open source
• 2:58
  contributors have each launched new products, updates or
• 3:01
  jaw-dropping showcases, many of which feel like they came from
• 3:05
  different futures. This isn't about incremental
• 3:08
  progress, it's about a cascade of systems each pushing
• 3:11
  boundaries once thought unscalable.
• 3:14
  But what's buried beneath the spectacle is the unsettling
• 3:17
  realization. How fast are we letting machines
• 3:21
  to what once required human hands, eyes, and voices, or
• 3:24
  evened our own emotions? So let's start with Bytedance's
• 3:29
  latest creation, Uno. This new AI model redefines what
• 3:33
  it means to generate an image. Uno responds to text prompts,
• 3:39
  but it also fuses multiple image references into a single
• 3:42
  coherent output, something that previous models often struggle
• 3:46
  with. Most impressively, it retains
• 3:49
  character fidelity and object accuracy, regardless of how
• 3:53
  surreal or stylistically diverse the prompt might be.
• 3:57
  Now, say you upload a logo and a photo of AT shirt.
• 4:00
  Uno doesn't just overlay it, it visualizes the logo printed on
• 4:04
  the shirt with convincing fabric, shadows and appropriate
• 4:08
  scale for the logo. Or consider uploading a doll and
• 4:11
  a plush toy. Uno will place them together in
• 4:15
  a new realistic scene. It even let's users generate A
• 4:19
  stylized portrait of a person using a single input photo.
• 4:23
  Anime, Pixar, or even video game style.
• 4:27
  Uno handles all of that stuff while maintaining identity
• 4:30
  trades. Now for creators, Uno is a
• 4:33
  silent assassin. If you're running a fashion
• 4:36
  brand, Uno is basically your model and your whole photo shoot
• 4:39
  team. Upload 2 clothing items and
• 4:41
  prompt the model to generate a person wearing them.
• 4:45
  Set in a cityscape or flower field of sunset.
• 4:48
  You can even try different models, body types, or urban
• 4:50
  settings. And for influencers, Uno becomes
• 4:53
  a face swapping studio. Give it one video of yourself
• 4:56
  and it'll generate your likeness in any scene, doing anything,
• 5:00
  wearing anything. This kind of tool doesn't just
• 5:03
  help with photorealistic image creation, it enables entirely
• 5:07
  new workflows for people who don't have a studio full of
• 5:11
  people. Now, what used to require hours
• 5:14
  and possibly days in Photoshop and multiple production rounds
• 5:17
  is now available with just a few clicks and a smart prompt.
• 5:23
  Now, Uno's value isn't just in flexibility, it's in fidelity.
• 5:27
  In benchmark comparisons with tools like Omni IP Adapter and
• 5:31
  OEM control, Uno consistently produce more accurate
• 5:35
  generations. For instance, when fed a
• 5:37
  reference image of a uniquely designed clock, Uno was the only
• 5:41
  tool able to correctly recreate it on green grass surrounded by
• 5:45
  sunflowers with even the yellow 3 intact.
• 5:49
  The same held true for altering the color of toys or preserving
• 5:53
  intricate shapes. Accuracy at the level means less
• 5:56
  trial and error and more trust in AI generated visuals.
• 6:00
  That's a quiet revolution for designers, developers and
• 6:04
  content creators who work under very tight deadlines.
• 6:08
  Now Byte Dance has made Uno freely available through a
• 6:11
  Hugging Face demo with an optional GitHub repository for
• 6:15
  local installs. You'll need a machine with at
• 6:17
  least 16 gigs of RAM and some familiarity with launching
• 6:21
  Python scripts. Quantized model variants are
• 6:24
  also available for those with limited GPU resources and this
• 6:28
  dual offering browser based ease and local control makes Uno
• 6:31
  accessible without compromising its power.
• 6:34
  Gives professionals and hobbyists alike the ability to
• 6:37
  integrate next Gen. visual tooling into their own
• 6:39
  workflows. Meanwhile, a whole different
• 6:43
  team has pushed video generation to the next level.
• 6:46
  Their latest tool, called One Minute Video Generation with
• 6:49
  Test Time Training, allows AI to produce full length, minute long
• 6:53
  videos that maintain consistent characters, environments, and
• 6:58
  art styles. Instead of relying on a single
• 7:01
  long prompt, you just breakdown a video into storyboard style
• 7:04
  scenes. Each scene contains a short
• 7:06
  description, like Jerry the brown mouse holds cheese or Tom
• 7:10
  the cat snatches the cheese and the model stitches these
• 7:13
  together to form a coherent, evolving story.
• 7:17
  Yes, the generated videos have rough edges, the lip sync is
• 7:20
  glitchy, backgrounds jitter, and text in scenes is often
• 7:24
  illegible. But the consistency of
• 7:26
  character, motion and design throughout a full minute is a
• 7:29
  major leap forward. Prior tools could barely sustain
• 7:33
  coherence over 3 to 5 seconds. This was made possible by
• 7:37
  layering test time training modules over Cog Video's base
• 7:40
  model. These modules act like memory
• 7:43
  units, absorbing and maintaining the visual state across scene
• 7:46
  transitions. It points to a near future where
• 7:50
  AI can generate a 10 minute short film with nothing but a
• 7:54
  script and a few references. Now then there's Hollow Point,
• 7:59
  another new AI system focused on 3D modeling.
• 8:02
  The tool doesn't just segment object, it reconstructs what's
• 8:06
  hidden from view. If you upload a ring, the model
• 8:09
  identifies the band in the settings, then completes the
• 8:12
  invisible side of each. That completion isn't guesswork.
• 8:16
  It uses a two stage system. First it segments visible parts.
• 8:20
  Next, it sends each piece through a 3D diffusion model to
• 8:25
  fill in occluded areas. The end result is a fully
• 8:29
  editable mesh with complete logical subparts.
• 8:33
  For 3D artists, this means they can now expand or customize
• 8:36
  individual parts of a model, resize the diamond, change a rim
• 8:40
  texture, or separate a chair's legs from its seat, all without
• 8:45
  needing to manually remodel or make speculative guesses.
• 8:49
  This kind of automation saves hours of work in production
• 8:53
  design, animation, and game development.
• 8:55
  Like many other tools this week, Hollow Part is available on
• 8:58
  Hugging Face for interactive testing and through GitHub for
• 9:01
  full installation. Open source models just got
• 9:04
  better than commercial ones. One major development this week
• 9:08
  came from a lesser known but quickly rising company called
• 9:10
  Vivigo AI. It launched Hydram, a new open
• 9:14
  source image generation model that has overtaken the
• 9:17
  competition, and that's at least in benchmark ratings.
• 9:21
  Unlike other high performing systems that remain closed or
• 9:24
  partially commercialized, High Dream is fully uncensored and
• 9:28
  openly available. According to Artificial Analysis
• 9:32
  Independent Model Ranking Organization, High Dream now
• 9:35
  holds the 3rd place position among all text to image models.
• 9:39
  It is the highest ranked open source model on the list,
• 9:42
  surpassing not only stable to Fusion, but also Flux 1D.
• 9:47
  Flux Pro, which ranks slightly higher, remains closed source,
• 9:51
  making Hydream the best available option for developers
• 9:54
  and researchers working without licensing restrictions.
• 9:58
  Early tests show that Hydream is more than a benchmark leader,
• 10:02
  and it performs well under practical creative conditions.
• 10:06
  Text prompts that previously produced vague or distorted
• 10:09
  visuals now return precise, detailed, and highly stylized
• 10:12
  results. This has opened up new use cases
• 10:15
  for illustrators, concept designers, and product marketers
• 10:18
  who want control over the visual fidelity of AI generated art.
• 10:22
  Importantly, I Dream allows full creative expression without the
• 10:26
  usual guardrails or sensors. This is a rare stance in the AI
• 10:30
  space, where most companies apply aggressive filtering or
• 10:34
  content restrictions. That raises ethical questions
• 10:37
  for many users, particularly in animation satire and adult
• 10:41
  design solves a problem they've long encountered with other
• 10:44
  models. Now Omni SVG is also rising it's
• 10:50
  text to vector generation. Another is stand out Omni SVG,
• 10:54
  an AI tool designed to generate scalable vector graphics, or
• 10:58
  SVGS, from text prompts or images.
• 11:00
  Unlike pixel based images, SVGS are resolution independent means
• 11:05
  that designers can scale these visuals from handheld to
• 11:10
  billboard sizes without losing fidelity, and developers can
• 11:14
  drop them into apps without worrying about file size or
• 11:17
  weight on the SVG stands out for its ability to produce complex,
• 11:21
  detailed vector illustrations that actually match user
• 11:24
  prompts. Previous models SVG generation
• 11:27
  often failed to capture layout or character details, but this
• 11:31
  one completely nails it. Examples include cartoon
• 11:34
  characters with mushroom hats, stylized buttons with web UI
• 11:37
  concepts, and even photo based SVG conversations.
• 11:41
  The retain line fidelity, so vectors still matter.
• 11:45
  This matters because the web still runs on vectors.
• 11:48
  Icons, logos, app UI elements, and infographics all rely on
• 11:53
  SVGS. And for developers who need
• 11:56
  sharp design across multiple screen sizes, or for companies
• 12:00
  localizing assets into dozens of languages, tools like Omni SVG
• 12:04
  reduce the time between idea and development.
• 12:07
  And what's more, the model also accept raster images and
• 12:11
  converts them into precise SVG renderings.
• 12:14
  This brings it into direct competition with some commercial
• 12:17
  vectorization tools and according to testing it
• 12:20
  outperforms most of them in detail.
• 12:22
  Preservation and line geometry code and data sets for Omni SVG
• 12:27
  are expected to be released soon.
• 12:30
  Meanwhile, Alibaba has debuted a new version of the talking head
• 12:34
  generator dubbed Omni Talker. The tool takes a video of a
• 12:38
  person talking just a few seconds and allows users to
• 12:41
  change what that person says using a custom transcript.
• 12:46
  And the results is a realistic looking video where the person
• 12:49
  says anything you want them to say.
• 12:51
  The output is not only visually smooth, but supports different
• 12:55
  languages, maintains mouth movements In Sync with the
• 12:58
  audio, and preserves facial expressions.
• 13:02
  This means you can take an input of Jackie Chan speaking in
• 13:05
  Mandarin and generate a video of him delivering a speech in
• 13:08
  English, accent intact, expressions unbroken.
• 13:13
  Of course, it's not perfect. The output can become uncanny if
• 13:16
  the person moves their head rapidly or if there are long
• 13:19
  monologues without breaks. Some cases, the AI renders the
• 13:22
  face smoothly or loses subtle emotional transitions if the lip
• 13:26
  sync and frame continuity far surpass most tools currently
• 13:31
  available. And the more fascinating part is
• 13:34
  emotional modulation. If the reference video shows a
• 13:38
  sad expression, the generated face will appear somber even if
• 13:42
  the transcript is upbeat. A happy reference video injects
• 13:46
  cheerful expressions into whatever script is applied.
• 13:49
  This emotional anchoring makes the system feel less robotic and
• 13:53
  more like a tool for actors or presenters who want to alter
• 13:56
  delivery without rerecording. Now most avatar generators top
• 14:02
  out at about 10 to 15 seconds. Omni Talker handles multi minute
• 14:06
  videos, maintaining voice sync and facial movements throughout.
• 14:11
  In one test, a 2 minute political speech was generated
• 14:14
  using just a headshot and a transcript, no voice acting
• 14:18
  needed. You can even interact with the
• 14:20
  avatars live in real time, giving them queries and getting
• 14:24
  real time spoken answers. This makes the system useful for
• 14:27
  educators, sales, customer service or political messaging.
• 14:32
  Any scenario where one person must deliver dynamic scripts at
• 14:36
  scale without reshooting or editing the video, removes AI
• 14:40
  from behind the scenes to center stage.
• 14:42
  And another related tool launched this week, though not
• 14:46
  from Alibaba, can take a static photo and animate the face using
• 14:50
  either an audio clip or a full reference video of facial
• 14:54
  movement feeded a Steve Jobs speech and a photo of Einstein.
• 14:58
  It'll make Einstein deliver Jobs words, lip syncing each syllable
• 15:01
  with Uri accuracy. And what's remarkable here isn't
• 15:04
  just the lip sync, it's the ability to mimic non verbal
• 15:07
  cues, blinking, nodding, pausing or even glancing away.
• 15:12
  The AI tracks motion and emotional shifts, mapping them
• 15:15
  to any face, real or fictional. Whether it's Anne Hathaway or a
• 15:19
  digital clone, the face mimics both speech and sentiment, and
• 15:24
  compared to rivals like X portrait or live portrait, this
• 15:27
  tool performs better on both fidelity and emotional nuance.
• 15:32
  The head doesn't stay locked, the eyes shift naturally, and
• 15:35
  when combined with expressive audio, it creates A convincing
• 15:38
  illustration of live speech. The GitHub repository is listed,
• 15:44
  but the code hasn't been released yet.
• 15:46
  It's dual mode functionality, supporting both speech only
• 15:49
  input and full motion transfer, means that users can choose
• 15:53
  between simple lip sync generation or full body
• 15:56
  animation. The results are strong enough to
• 15:59
  enter entertainment production, especially for animated
• 16:02
  interviewers, explainer videos or social content.
• 16:07
  Now, robots. Aside of software, the most
• 16:11
  viral moments this week came from humanoid robots.
• 16:14
  Unitri's robot has performed flips and complex dances before,
• 16:17
  but Engine AI's robot raises eyebrows with boxing humans
• 16:22
  reacting to punches and landing counter blows.
• 16:25
  The demo, though, was captured during a live stream with a well
• 16:28
  known streamer visiting the company's facility in China.
• 16:31
  Critics have long claimed these humanoid demos were faked CGI
• 16:35
  stitched into real environments, but the live stream confirmed
• 16:39
  otherwise. The robot fell, got back up,
• 16:42
  aimed some punches and adjusted its stance all autonomously with
• 16:46
  AI. Not pre recorded, not
• 16:48
  choreographed. Unlike dancing routines or
• 16:50
  rehearsed acrobatics, boxing requires real time adaptation.
• 16:54
  The robot had to interpret opponent movements, adjust its
• 16:58
  balance, and select appropriate actions, all within
• 17:00
  milliseconds. Like human.
• 17:03
  It's about machines competing with humans.
• 17:06
  Physically a human versus a robot.
• 17:09
  We've seen this movie before, and the humans don't fare too
• 17:12
  well in it. Now what sets this apart is the
• 17:15
  autonomy system wasn't remote controlled.
• 17:17
  It executed decisions based on internal algorithms tracking
• 17:21
  visual input and calculating body position.
• 17:24
  If it fell, it knew how to stand up.
• 17:26
  If it gets too close, it backed off.
• 17:29
  For the first time, the idea of reacting humanoid robotics feels
• 17:33
  actually real. Then there's Kawasaki.
• 17:37
  They introduced a concept that sounds absurd until you watch
• 17:41
  it, an AI powered mechanical horse.
• 17:46
  Unlike traditional scooters or two Wheelers, this machine walks
• 17:49
  on 4 robotic legs. Its body mimics the movement of
• 17:53
  an actual horse and it's designed to be powered by
• 17:56
  hydrogen, releasing only water vapor.
• 17:59
  At this stage the prototype is non functional.
• 18:02
  The riding demo is CGI, and the model shown at Osaka's Expo is a
• 18:07
  static mock up. But the idea is absolutely
• 18:09
  serious. Kawasaki envisions a transport
• 18:12
  method that's clean, adaptive, and capable of navigating uneven
• 18:16
  terrain, but early feedback has been skeptical.
• 18:19
  Most people agree that wheels still outperform legs when it
• 18:22
  comes to energy efficiency and also stability.
• 18:26
  Now, all of these tools and models and demos are very
• 18:31
  flashy. They're tangible shifts in how
• 18:34
  humans can work, communicate, and create.
• 18:36
  If you're a designer, AI will now generate not only your
• 18:39
  visuals, but your style. If you're a performer, AI will
• 18:42
  now speak with your face and your voice.
• 18:44
  If you're a robotics engineer, the next leap may come not from
• 18:48
  code, but from carbon fiber and real time planning algorithms.
• 18:53
  The acceleration isn't evenly distributed, but the
• 18:57
  opportunities and risks are everywhere.
• 18:59
  These tools aren't just for labs or corporations anymore.
• 19:02
  They're for us. They're available on GitHub.
• 19:05
  Hugging Face or even browser based apps anyone can build with
• 19:10
  them. Hey, thank you so much for
• 19:14
  listening today. I really do appreciate your
• 19:16
  support. If you could take a second and
• 19:18
  hit this subscribe or the follow button on whatever podcast
• 19:21
  platform that you're listening on right now, I greatly
• 19:24
  appreciate it. It helps out the show
• 19:26
  tremendously and you'll never miss an episode.
• 19:28
  And each episode is about 10 minutes or less to get you
• 19:32
  caught up quickly. And please, if you want to
• 19:35
  support the show even more, go to Atreoncom Stage Zero.
• 19:41
  And please take care of yourselves and each other, and
• 19:43
  I'll see you tomorrow.