Latest / Today’s AI News / Meta Is Back in the AI Race, HeyGen’s Hyper-Realistic Avatar, Anthropic’s Agent Builder Opens Up

Transcript

Meta Is Back in the AI Race, HeyGen’s Hyper-Realistic Avatar, Anthropic’s Agent Builder Opens Up · Today’s AI News

• 0:00
  Welcome back to today's AI News, brought to you by 9X
• 0:02
  Productions. Glad to be here for this one.
• 0:04
  Yeah, we've got a massive deep dive for you today.
• 0:07
  So imagine buying the world's most expensive high performance
• 0:10
  engine for $14.3 billion and dropping it right into a family
• 0:15
  sedan. That's quite the visual.
• 0:17
  Right. Well, today we finally get to
• 0:19
  see how fast that car can go because Meta just fired up the
• 0:22
  engine on their brand new super intelligence lap.
• 0:25
  And that is just the start. Exactly.
• 0:27
  Our mission today is to explore that massive new proprietary
• 0:31
  model from Meta. Plus, we'll look at how AI
• 0:33
  avatars just solved their biggest visual glitch.
• 0:37
  An all in one studio for AD generation and tropics new
• 0:40
  invisible back end. And we'll cap it off with the
• 0:42
  latest quick hits and a really brilliant community workflow.
• 0:45
  It's a huge agenda. And you know, the through line
• 0:48
  for all of this today is scale. We're seeing intelligence scale
• 0:51
  up in these massive server rooms and then simultaneously scale
• 0:55
  right down into your hands. Well, OK, let's start with that
• 0:58
  massive scale. Mehta is making the biggest
• 1:00
  splash in the industry right now.
• 1:02
  We finally get to see what Mark Zuckerberg bought with his
• 1:05
  checkbook. And it was a very large check.
• 1:07
  Yeah, like $14.3 billion large HE scale AI about nine months
• 1:14
  ago, right? Right, he bought the company and
• 1:16
  put Alexandra Wang in charge of a completely new division, the
• 1:20
  Meta Super Intelligence Labs. Right.
• 1:22
  And they didn't just tweak things.
• 1:24
  Wang and his team basically rebuilt Meta's AI stack from
• 1:28
  scratch. From the ground up, yeah.
• 1:30
  And the result of that nine month Sprint is Musespark.
• 1:34
  Musespark. So it's a multimodal reasoning
• 1:36
  model. Exactly.
• 1:37
  Meaning it handles voice, text, and images all natively.
• 1:41
  But the the really fascinating part here, the engine from your
• 1:44
  analogy is what they're calling contemplating mode.
• 1:47
  OK, let's unpack this because contemplating mode sounds,
• 1:50
  honestly, it sounds like a slick marketing buzzword.
• 1:52
  Like, what is it actually doing? Is it just, I don't know,
• 1:55
  running slower? No, it's a totally different
• 1:57
  architecture. So you know how standard models
• 1:59
  work, right? Next token prediction.
• 2:01
  Right. They just guess the next word in
• 2:02
  the sequence. Yeah, like very smart auto
• 2:04
  complete. Exactly, but contemplating mode
• 2:07
  breaks that completely when you give it a hard problem.
• 2:11
  It basically pits multiple internal AI agents against each
• 2:14
  other. Wait, really?
• 2:16
  Like they argue. Basically, yeah.
• 2:18
  One agent drafts a solution, and then a second agent plays
• 2:21
  devil's advocate to find logical flaws in it, and then maybe a
• 2:24
  third one checks the math. Wow.
• 2:26
  So they debate and iterate before they ever show you the
• 2:29
  final answer. Which would explain why the
• 2:31
  benchmarks are so wild, like it's running competitive with
• 2:34
  Frontier models Opus 4.6 and GPT 5.4 specifically on reasoning.
• 2:40
  It's very strong there, especially in health reasoning
• 2:43
  actually. Which makes sense for Meta's
• 2:44
  whole personal super intelligence mission.
• 2:47
  Though I mean to be fair it does lag a bit in coding and the
• 2:50
  ARCAG I2 tests. Yeah, it's not perfect across
• 2:53
  the board yet. But here's the thing that really
• 2:55
  stood out to me. Use Spark is a completely
• 2:57
  proprietary. Yeah, that is the earthquake the
• 3:00
  industry is feeling right now. I mean it's a huge departure.
• 3:04
  It's like Meta spent all these years building a free community
• 3:07
  garden with the llama models and now they've suddenly erected
• 3:11
  AVIP only greenhouse right in the middle of it.
• 3:14
  That's a great way to put it. So does this signal the end of
• 3:17
  Meta's open source era? Are they just alienating all
• 3:20
  those developers? Well, what's fascinating here is
• 3:22
  the financial reality of it. When you drop $14.3 billion on
• 3:28
  an acquisition, the boardroom dynamics shift entirely.
• 3:31
  Right, they want their money back.
• 3:33
  Exactly. Shareholders want immediate ROI.
• 3:35
  You don't just open source a $14 billion advantage on day one.
• 3:39
  Even though they said they hoped to open source future versions,
• 3:42
  there's no timeline. Right.
• 3:44
  And you have to remember Meta's ultimate advantage, it's not
• 3:46
  just the model, it's their distribution.
• 3:48
  They have over 3 billion daily users.
• 3:52
  The Facebook, Instagram, WhatsApp, just massive data
• 3:55
  streams. Exactly deploying this
• 3:57
  contemplating model across that proprietary ecosystem is how
• 4:01
  they win. They're building an invisible
• 4:03
  super brain for their own network.
• 4:06
  OK, so if meta is building this invisible contemplating brain,
• 4:10
  it brings up a good question like how does that intelligence
• 4:13
  actually talk to you it? Needs an interface.
• 4:16
  Right, a text box feels way too primitive for a multi agent
• 4:19
  reasoning engine. It needs a physical face.
• 4:23
  And that leads us perfectly to Hagen's latest update.
• 4:26
  Avatar V. Yes, Hagen just released Avatar
• 4:29
  V and they are claiming it's the most realistic AI avatar model
• 4:34
  in the world. It's a bold claim, but they
• 4:36
  actually solved a massive technical hurdle here.
• 4:38
  The big glitch? Identity drift.
• 4:41
  OK yeah, so for anyone who hasn't messed around with AI
• 4:44
  video, identity drift is that really annoying thing where the
• 4:47
  generated face slowly morphs over time.
• 4:49
  It just slowly stops looking like the person.
• 4:51
  Exactly. The eyes get a little further
• 4:53
  apart, the jaw changes. It's like a photocopy of a
• 4:55
  photocopy. Yeah, but Hagen claims Avatar V
• 4:58
  fixes this completely. And the crazy part is how little
• 5:01
  data it needs. Just a 15 second phone
• 5:03
  recording. Right from 15 seconds it maps
• 5:05
  your real facial details, your micro gestures, all of it.
• 5:08
  How are they doing that without it drifting?
• 5:10
  So in older models, the AI would bake everything together.
• 5:15
  Your face, your shirt, the lighting in the room.
• 5:17
  It was all one big data BLOB, entangled exactly.
• 5:21
  But Avatar V untangles it. It separates your identity from
• 5:25
  your appearance. Here's where it gets really
• 5:26
  interesting for me. If it separates the identity,
• 5:29
  that means you just record that 15 second clip once.
• 5:31
  Yep, one time. And then you could swap out your
• 5:34
  outfits, change your background to a beach or a studio without
• 5:37
  ever actually filming it again. It's basically a digital stunt
• 5:40
  double that can change its entire wardrobe without stepping
• 5:43
  foot in a trailer. And their internal tests back
• 5:45
  this up. It's beating Google's VO 3.1 on
• 5:49
  lip sync and accuracy, and it's crushing cling and sea dance.
• 5:53
  But I mean, let me push back a little here.
• 5:55
  Do you really want an AI twin that is completely
• 5:58
  indistinguishable from you? Like the trust issues there are
• 6:01
  wild. A lot of people definitely scoff
• 6:03
  at the idea of an AI twin. It feels invasive.
• 6:06
  But if you look at the content creation landscape, it's
• 6:09
  changing irrevocably. Well, separating identity from
• 6:12
  appearance is a massive leap for enterprise.
• 6:15
  Imagine you're a global brand. You need an ad in 20 languages
• 6:19
  with 20 different localized backgrounds.
• 6:21
  By tomorrow you can't shoot that physically, but with this, the
• 6:25
  CEO records for 15 seconds and the AI does the rest perfectly.
• 6:30
  The efficiency is just undeniable.
• 6:32
  OK, so you have your flawless digital stunt double, but
• 6:36
  standing in a void isn't an ad campaign.
• 6:38
  You still need like a set, lighting, maybe a music track.
• 6:42
  A whole production crew. Exactly.
• 6:45
  So what if you could just generate the whole campaign in
• 6:47
  one browser window? 11 Labs just launched Flows and
• 6:52
  it does exactly that. This is a really elegant
• 6:54
  solution to tool fragmentation. It's an all in one studio.
• 6:57
  So for the listener, let me break down how simple this is.
• 6:59
  You open 11 labs, go to 11 Creative and click New Flow.
• 7:02
  It's very visual node based. Right, so you drag an image
• 7:05
  generation node onto the screen. You upload a quick product shot
• 7:08
  from your phone. Then you prompt it.
• 7:10
  Product on a white pedestal. Soft morning light.
• 7:12
  Simple enough. Then you literally just connect
• 7:15
  that node to a video generation node and you prompt the camera
• 7:18
  movement like slow cinematic push in.
• 7:20
  And then you just run it. Yeah, hit run till here and if
• 7:23
  you don't like it you just tweak the text and iterate.
• 7:27
  It's incredibly fast. And the pro tip is you can add
• 7:30
  text to speech nodes or music nodes and wire them all into a
• 7:33
  mix audio node to finish the ad. So I have to ask, what does this
• 7:39
  mean for traditional creative agencies?
• 7:42
  It's highly disruptive. It feels like we just handed
• 7:45
  someone a box of magic Lego blocks and every single block is
• 7:49
  an entire Hollywood department. That analogy is spot on, but we
• 7:53
  do need to caution people because if we connect this to
• 7:56
  the bigger picture, the barrier to entry is basically gone now.
• 8:00
  Anyone can make a high quality video.
• 8:02
  Right. So the differentiator isn't your
• 8:03
  budget anymore. It's not about affording the
• 8:06
  best camera or lighting crew. So what?
• 8:08
  Is it? It's pure creative vision, good
• 8:11
  taste when anyone can generate A cinematic shot.
• 8:14
  The only thing that makes your ads stand out is whether the
• 8:16
  concept actually connects with people.
• 8:18
  OK, so we've got the digital faces, we've got the video ads.
• 8:22
  But behind the scenes, running these AI systems used to be a
• 8:25
  nightmare. You needed a massive engineering
• 8:28
  team to handle the back end. Oh, the plumbing was terrible,
• 8:32
  right? But Anthropic is fixing that.
• 8:34
  They just opened a public beta for Claude Managed Agents.
• 8:38
  This is huge for developers. They are making the back end
• 8:41
  completely invisible. Yeah, you can go from an idea to
• 8:44
  a live working agent in days instead of months.
• 8:48
  The developer just picks a task the software tools the AI can
• 8:51
  use and sets the guardrails. And Anthropic handles the
• 8:55
  security, the infrastructure, all of it.
• 8:57
  But the real breakthrough here is that the agents don't drop
• 9:00
  state. Right, OK, explain dropping
• 9:02
  state for us because that's a big deal.
• 9:03
  So previously if you had an AI running a task for hours, it
• 9:07
  would eventually forget what it was doing, its memory would fill
• 9:10
  up and it would drop the context or.
• 9:11
  Just lose the plot. Exactly a drop state, but
• 9:14
  Anthropic solved this. These managed agents can run
• 9:17
  solo for hours without forgetting their instructions.
• 9:20
  And they even have a preview of a coordination mode right where
• 9:23
  one manager agent farms out smaller tasks to a bunch of
• 9:27
  subordinate agents. Yes, agent swarms and it's so
• 9:30
  accessible because of the pricing model.
• 9:33
  Yes, it's not some massive enterprise license, it's usage
• 9:37
  based $0.08 per hour per session. $0.08 an hour for an
• 9:41
  autonomous worker. It's crazy and early adopters
• 9:44
  are already jumping on it. Notion, Asana Sentry.
• 9:48
  But the recruiting example is wild.
• 9:51
  The deployment speed was incredible there.
• 9:53
  They deployed agents across 5 different departments and it
• 9:56
  only took them about a week for each one.
• 9:58
  It's literally like hiring a brilliant project manager who
• 10:01
  automatically hires, manages, and pays a team of temp workers
• 10:05
  for exactly $0.08 an hour. And what this shows is that the
• 10:08
  true bottleneck for AI adoption hasn't actually been the models
• 10:12
  themselves. It.
• 10:12
  Was the implementation exactly? It's the back end headaches,
• 10:15
  maintaining servers, keeping the state active, and Fropic is
• 10:18
  actively eliminating all of that friction.
• 10:21
  All right, so having explored these deep technical shifts,
• 10:24
  let's zoom out a bit. Let's look at how this is
• 10:26
  shaking up the business and political landscape.
• 10:29
  In our Quick hit segment, there's a.
• 10:30
  Lot moving quickly today. Rapid fire time.
• 10:33
  First, some trending tools. We've got Scrunch, which is an
• 10:36
  AI site interpreter and auditor, and Clicky which is this open
• 10:40
  source AI teacher that just lives next to your cursor.
• 10:43
  Very handy utility tools. Yeah, and over the business
• 10:46
  world, Canva just acquired SIM Theory and Ordo to push their
• 10:49
  end to end campaign workflows. They're.
• 10:51
  Building a massive ecosystem. But here's the real business
• 10:54
  news Perplexity just hit a massive $450 million in
• 11:00
  estimated annual recurring revenue.
• 11:02
  That is a staggering number. It's a 50% monthly jump driven
• 11:07
  entirely by their computer agentic system and usage based
• 11:10
  pricing. I want to pause here because we
• 11:12
  talk so much about the hype cycle, but a 50% monthly jump in
• 11:16
  real revenue proves people are actually paying for these
• 11:19
  agentic systems to do their work.
• 11:20
  It's not just hype anymore. No, the value is proven, and
• 11:24
  this actually raises an important question about the
• 11:26
  talent war happening right now. No, the poaching.
• 11:29
  Exactly. Jeff Bezos has this secretive AI
• 11:32
  startup called Prometheus and they just poached Kyle Kosek.
• 11:36
  He. Was a former XAI Co founder
• 11:37
  right? Yeah, and he was recently at
• 11:39
  Open AI, so he's jumping between all the major players.
• 11:41
  And it highlights that the real bottleneck in AI isn't just
• 11:45
  compute power anymore. It's human capital.
• 11:47
  Yes, it's the incredibly small pool of humans who actually know
• 11:52
  how to build and scale these massive of systems.
• 11:55
  To be fair, the compute war is definitely still happening.
• 11:58
  Elon Musk just revealed X AI is training 7 new models on the
• 12:02
  Colossus 2 supercomputer. Including massive 6 trillion and
• 12:06
  10 trillion parameter systems. Just behemoths.
• 12:09
  And Speaking of Musk, there's major movement on the regulation
• 12:12
  and legal side too. First, Open AI published a child
• 12:14
  safety policy blueprint pushing for updated US laws on AI
• 12:18
  generated C Sam. Which is a very necessary
• 12:21
  regulatory push. Absolutely.
• 12:23
  And on the legal front, and I want to be very clear to the
• 12:26
  listener, we are remaining entirely neutral on this.
• 12:29
  We are simply reporting the filings from both sides.
• 12:32
  We're not endorsing either viewpoint here.
• 12:34
  Right, just looking at the facts of the case.
• 12:36
  Yes. So Elon Musk amended his lawsuit
• 12:39
  against opening AI to redirect all potential damages to their
• 12:42
  nonprofit arm, and he's pushing to get Sam Altman off the board.
• 12:46
  And open eyes response was pretty aggressive.
• 12:48
  Very. They responded by officially
• 12:50
  calling the amended lawsuit a harassment campaign.
• 12:55
  So yeah, the legal battles are heating up just as fast as the
• 12:57
  server farms. It's a deeply complex landscape
• 13:00
  right now. It really is.
• 13:02
  OK, so we've talked about billionaires fighting massive
• 13:05
  supercomputers and multibillion dollar buyouts, but how is this
• 13:09
  actually impacting you, the listener in your day-to-day
• 13:12
  life? Let's look at the community.
• 13:13
  I love this segment. It grounds the technology.
• 13:16
  Exactly. Today we're highlighting a
• 13:17
  reader named Lynn B from the UK. She's a serious amateur
• 13:21
  photographer. And she does some really complex
• 13:23
  in camera techniques, right? Yeah, she creates these
• 13:26
  beautiful expressionist multiple exposure images.
• 13:30
  The issue is she teaches these techniques to people over Zoom
• 13:33
  and a lot of her students just don't have cameras capable of
• 13:36
  multiple exposure's. There's an accessibility issue
• 13:38
  for her students. Right O she rocesses her own
• 13:40
  stuff in Hotosho, but writing a script to automate that process
• 13:44
  for her students is tough if you don't code.
• 13:48
  So she turned to ChatGPT. Exactly, she used a ChatGPT to
• 13:52
  write the code for custom scripts within Photoshop to
• 13:55
  simulate the whole in camera multiple exposure process.
• 13:59
  That's incredibly clever. And Photoshop error codes could
• 14:02
  be really temperamental, so she uses the AI to debug them.
• 14:06
  I love what she said about it, she said.
• 14:07
  The AI matches my tenacity with a challenging task.
• 14:10
  Matches my tenacity. That's a great phrase.
• 14:13
  Isn't it? It just points out that AI is
• 14:14
  the ultimate patient collaborator.
• 14:16
  It doesn't get tired, it just keeps helping you iterate.
• 14:19
  And what's so beautiful about Lynn's workflow is how it
• 14:22
  bridges the physical and digital worlds.
• 14:24
  She's using a language model, text based AI to democratize a
• 14:28
  very niche physical photography technique for her friends.
• 14:32
  It's amazing. It really is, and it leaves us
• 14:34
  with a really fascinating contrast to Mull over today.
• 14:37
  What's that think about the scale we've covered?
• 14:39
  On one hand, you have Meta spending $14.3 billion just to
• 14:45
  build a contemplative super brain.
• 14:48
  You have Elon Musk building 10 trillion parameter behemoths on
• 14:53
  supercomputers. The macro scale.
• 14:54
  Exactly the macro corporate level, but on the exact same day
• 14:58
  an amateur photographer in the UK is using that exact same
• 15:02
  underlying technology just to solve a niche Photoshop script
• 15:06
  for her class. The micro personal level, right?
• 15:08
  Intelligence is staling simultaneously at both of those
• 15:11
  extremes, and the question I want to leave the listener with
• 15:14
  is which one of those scales will actually change your life
• 15:17
  more? Oh, that is a great thought to
• 15:19
  end on because honestly, Meta might be building the $14
• 15:21
  billion engine, but people like Lynn are the ones figuring out
• 15:25
  how to actually drive the car in everyday life.
• 15:27
  Well, that is all we have for today's deep dive.
• 15:29
  Make sure to subscribe so you can stay updated on all these
• 15:32
  massive shifts. And if you enjoy the show,
• 15:34
  please rate it five stars. It really helps us out.
• 15:38
  We'll see you back for tomorrow's episode.