Latest / Today’s AI News / Anthropic Cuts Off OpenClaw Users, Netflix Releases Physics-Aware Video AI, ChatGPT Comes to CarPlay

Transcript

Anthropic Cuts Off OpenClaw Users, Netflix Releases Physics-Aware Video AI, ChatGPT Comes to CarPlay · Today’s AI News

• 0:00
  Welcome back to today's AI News, brought to you by Nynex
• 0:03
  Productions. Hey everyone, great to be back
• 0:05
  for another deep dive. So imagine an all you can eat
• 0:08
  buffet that suddenly realizes a single customer brought in a
• 0:11
  fleet of hungry robots to clear out the entire kitchen.
• 0:16
  That's essentially what's happening in the AI world today,
• 0:18
  as providers scramble to figure out who or what is eating all
• 0:22
  their compute. Yeah, the buffet is officially
• 0:25
  closing early. I mean, the physical economics
• 0:27
  are just completely buckling. Right.
• 0:29
  And that that is exactly our mission.
• 0:31
  For today's Deep Dive, we are going to explore this massive
• 0:34
  tension growing between AI platforms and their absolute
• 0:38
  power users. Skip messy, It really is.
• 0:40
  And then we're unpacking how Netflix is using AI to literally
• 0:44
  rewrite the laws of physics inside video environment.
• 0:47
  Which is mind blowing, totally. After that, we'll run through
• 0:49
  some rapid fire ecosystem shifts in our quick hits, and then
• 0:52
  finally we'll look at how a local hobbyist is utilizing AI
• 0:56
  to build community tools like from the ground up.
• 0:58
  With 0 coding experience. Yeah, exactly 0.
• 1:02
  But let's jump right into this first story because it is making
• 1:06
  some serious waves. So Anthropic, the company behind
• 1:10
  Claude, has officially started booting third party agent
• 1:13
  platforms from their standard subscription plans.
• 1:16
  Yeah. Platforms like Open Claude just
• 1:18
  getting kicked off. Right because agent tools they
• 1:20
  just create these non-stop requests and that completely
• 1:23
  exceeds what a flat rate plan can cover.
• 1:26
  So now users have to pay separately via like usage
• 1:30
  add-ons or API keys. Which is a huge pivot from how
• 1:33
  this was originally sold to the public.
• 1:35
  Exactly. I mean, Boris Churney from
• 1:37
  Anthropic came out and said this is all about managing growth and
• 1:40
  long term sustainability and they're trying to soften the
• 1:43
  blow. Yeah, handing out one month
• 1:45
  credits. Right, one month credits, 30%
• 1:47
  discounts on the add-ons and full refunds if you just want to
• 1:50
  cancel. But I mean, the open source
• 1:52
  community is completely livid. They're furious.
• 1:55
  Yeah, Peter Steinberger, the creator of Open Claw, he heavily
• 1:59
  criticized them. He basically said Anthropic just
• 2:02
  copies popular open source features, puts them in their
• 2:05
  closed system, and then completely locks the open source
• 2:08
  developers out. It's, well, it's that classic
• 2:10
  text story, right? A platform scales up and then
• 2:13
  suddenly realizes the exact developers who popularized their
• 2:16
  system are now like their biggest financial liability.
• 2:20
  Right. I kind of look at it like this.
• 2:22
  Using a standard AI subscription is like buying a monthly city
• 2:26
  bus pass. You know, you write it to work,
• 2:28
  you get off, no big deal. Makes sense.
• 2:30
  But running an AI agent, it's like trying to use that exact
• 2:33
  same $20 bus pass to operate an entire commercial logistics
• 2:36
  fleet. Yeah, like 24/7.
• 2:38
  It just breaks the system entirely.
• 2:40
  But I do have a question for you on this.
• 2:42
  By punishing their most innovative power users right at
• 2:44
  this exact moment, when Open AI is aggressively courting
• 2:47
  developers, is Anthropic risking its reputation just to save on
• 2:51
  compute costs? Well, it's a really precarious
• 2:53
  balancing act. But to understand why they're
• 2:56
  willing to take that reputational hit, you have to
• 2:58
  look at the economics of what we call a gentic loops.
• 3:03
  OK, break that down for us. So compared to the human
• 3:05
  prompting, when you or I use an AI, there's natural friction.
• 3:09
  We type of prompt, we wait, we read the answer, maybe we go
• 3:13
  grab a coffee. Right, the human latency.
• 3:15
  Exactly that human latency is fully baked into those $20 flat
• 3:21
  rate pricing models. Like the buffet doesn't go
• 3:23
  bankrupt because human beings eventually get full, or at the
• 3:26
  very least they have to take time to chew their food.
• 3:29
  Exactly. But autonomous agents do not
• 3:32
  chew, right? When you set an agent on a
• 3:34
  complex task like, I don't know, researching an industry and
• 3:38
  writing a report, it creates A continuous loop.
• 3:41
  It prompts the model, gets the output, checks for errors, and
• 3:44
  prompts it again. Over and over.
• 3:46
  Without taking a breath. Right.
• 3:47
  An agent could send hundreds of complex requests in the exact
• 3:50
  same time it takes you to sip your coffee.
• 3:53
  So those flat rate structures fundamentally breakdown because
• 3:55
  the compute cost curve just goes exponential.
• 3:58
  We're talking about massive, incredibly expensive GPU
• 4:00
  clusters spinning up continuously while the revenue
• 4:03
  Anthropic gets from that user stays at a flat 20 bucks.
• 4:08
  So for sheer infrastructure survival, they had to cap it.
• 4:11
  So they're literally just bleeding money on these specific
• 4:14
  users. Substantially bleeding money now
• 4:17
  to be impartial here, the frustrations from the open
• 4:20
  source community are entirely valid too.
• 4:22
  I mean people like the Open Claw creator built their platforms
• 4:25
  assuming clawed would be accessible.
• 4:27
  And then the rut that's pulled. Yeah, it stings.
• 4:29
  It feels like a betrayal of the open source ethos.
• 4:32
  But Anthropic just can't subsidized independent
• 4:35
  developers indefinitely or they'd go bankrupt.
• 4:38
  And meanwhile, Open AI is probably just sitting back
• 4:40
  watching this happen. Open AI stands to benefit
• 4:43
  immensely from this friction. If Anthropic makes things
• 4:46
  expensive for agent developers, Open AI can just swoop in with
• 4:50
  like free credits and better rate limits to grab that market
• 4:53
  share. But won't they hit the same wall
• 4:55
  eventually? Oh absolutely.
• 4:56
  Compute is not infinite, you know.
• 4:58
  It doesn't matter which logo is on the server.
• 5:00
  Eventually, the physics of the data center will catch up with
• 5:03
  open AI. Which is actually a perfect
• 5:05
  transition, honestly, because Anthropic's whole struggle right
• 5:08
  now is about managing the unpredictable, cascading
• 5:12
  consequences of autonomous agents.
• 5:15
  And that mirrors our next story perfectly, because Netflix is
• 5:19
  trying to manage the cascading physical consequences of objects
• 5:23
  inside video environments. This story is wild.
• 5:27
  It's one of the biggest leaps in synthetic media we've seen in a
• 5:29
  while. It really is.
• 5:31
  So Netflix research just open sourced void VO ID and it's a
• 5:37
  physics aware AI framework for video editing.
• 5:40
  Normally if you use AI to remove an object from a video, it just
• 5:44
  kind of paints over the gap, right?
• 5:46
  It hallucinates the background, but void actually rewrites the
• 5:50
  scene's physics. It simulates it.
• 5:51
  Exactly. It uses a mask to show what's
• 5:53
  being removed, and then it uses this thing called a judge model
• 5:56
  to chart the consequences of removing it.
• 5:58
  And the examples he gave in the paper are just fantastic.
• 6:00
  Yeah, so for you listening, imagine a video of a person
• 6:04
  holding a balloon. If you use void to erase the
• 6:08
  person, the balloon doesn't just awkwardly hover in midair.
• 6:11
  The AI actually knows that without a hand holding the
• 6:14
  string, the balloon has to float away.
• 6:16
  Right. It understands buoyancy.
• 6:18
  Yes, or if you have a stack of blocks and you erase one in the
• 6:21
  middle. The top blocks don't just say
• 6:23
  suspended, they actually fall. And when evaluators compared
• 6:27
  Void against six other baseline models, including big ones like
• 6:31
  Runway, they preferred Netflix's results nearly 2/3 of the time.
• 6:35
  Which is a massive margin in these kinds of evaluations.
• 6:38
  The technical lead here from static visual in painting to
• 6:42
  actual temporal physics based reasoning is huge.
• 6:45
  OK, let me give an analogy for this.
• 6:47
  It's like removing a load bearing wall in a digital house.
• 6:50
  OK, I like. Old AI would just, you know,
• 6:52
  paint a nice picture of a normal wall over the gap, but Netflix
• 6:55
  is AI knows the roof should actually collapse.
• 6:58
  So my question is, if this hyper realistic physics manipulation
• 7:02
  is open sourced, how long until literally anyone can seamlessly
• 7:07
  alter reality in a video? And are we really ready for that
• 7:10
  level of controllability? Well, to answer the second part
• 7:13
  first, no, yeah, we are definitely not ready as a
• 7:16
  society for flawless consumer level reality alteration.
• 7:19
  Yeah, that's what I. Figured, but to explain why this
• 7:22
  is such a leap, we have to talk about that judge model you
• 7:25
  mentioned. Older models operate almost
• 7:27
  entirely frame by frame. Like a flip.
• 7:29
  Book kind of yeah. They look at a single still
• 7:31
  image, patch the pixels based on visual patterns and move to the
• 7:35
  next. They have literally 0 concept of
• 7:37
  time or how objects relate to each other physically.
• 7:40
  So they know what a balloon looks like, but they don't know
• 7:42
  what a balloon does. Exactly.
• 7:43
  The judge model and void is basically an AI trained to
• 7:46
  understand intuitive physics, gravity, momentum, structural
• 7:51
  integrity. So when it looks at the frames
• 7:53
  it actually simulates cause and effect.
• 7:55
  If object A is removed, it asks what are the physical forces
• 7:59
  acting on object B in the next frame.
• 8:01
  So it's literally a physics engine working backwards from a
• 8:04
  flat video. Yes, and that is a massive shift
• 8:08
  for Hollywood production pipelines.
• 8:10
  Think about how many millions of dollars go into visual effects
• 8:14
  just to clean up a shot or remove a safety wire.
• 8:16
  Oh yeah, the cleanup costs are insane, right?
• 8:19
  So if you have an AI that intrinsically gets the physical
• 8:22
  reality of the scene, you aren't just editing flat video anymore,
• 8:26
  you're manipulating a 3D simulated world.
• 8:29
  It means we're moving away from AI that just hallucinates pretty
• 8:33
  pixels tord AI that generates logically consistent realities.
• 8:37
  Which is just wild to think about.
• 8:38
  And you know, if companies like Netflix and Anthropic are
• 8:41
  building these incredibly complex reality altering
• 8:44
  systems, it really explains why the business side is just an
• 8:46
  absolute arms race right now. It is a total land grab.
• 8:49
  It is, so let's jump into our rapid fire ecosystem shifts
• 8:52
  starting with the quick hits. Let's do it.
• 8:54
  OK, so skipping over the ad stuff on the tools front, Cursor
• 8:57
  3 just launched an agent first interface for parallel coding.
• 9:01
  Which is going to speed up dev workflows immensely.
• 9:04
  Then there's Pika stream One Point O, which is basically a
• 9:07
  video chat AI that gives AI agents an actual face and voice.
• 9:12
  And Google just dropped Gemma 4, which is a new open weight AI
• 9:16
  boasting state-of-the-art intelligence.
• 9:18
  And we should clarify something really quick on Gemma 4.
• 9:21
  Yeah, Google specifically called it open weight, not open source.
• 9:25
  Oh, right. What is the practical difference
• 9:27
  there for the listener? It's a huge distinction.
• 9:30
  True open source means everything's public, The
• 9:33
  training data, the code to train it, everything.
• 9:35
  But open weight means they're only releasing the final trained
• 9:38
  neural network parameters. The weights you get the final
• 9:41
  product, but not the recipe. Exactly.
• 9:43
  Developers can use it and build on it, but Google keeps their
• 9:46
  corporate secrets hidden. That makes total sense.
• 9:49
  OK, on to the industry news, which is heavy today.
• 9:53
  Open AI is going through a massive leadership shift.
• 9:56
  Fiji Simo is on medical leave, Brad Lightcap is moving to
• 9:59
  special projects and Kate Rauch is stepping down for cancer
• 10:03
  recovery. Yeah, a lot of executive
• 10:04
  shuffling over there. Definitely.
• 10:06
  Meanwhile, Anthropic just acquired a startup called
• 10:09
  Coefficient Bio for roughly $400 million, specifically to focus
• 10:14
  on drug discovery. Oh, and Mercor, which is a $10
• 10:18
  billion AI startup, suffered A4 TB data breach.
• 10:22
  That one is a big deal. Huge deal.
• 10:24
  It happened via a supply chain attack on an open source library
• 10:27
  called Light LM. And bridging all this together,
• 10:30
  Pika Labs just put that Pika stream beta into Google meet
• 10:33
  calls so you can literally have an AI avatar join your video
• 10:37
  conference. Oh, and Open AI rolled out
• 10:39
  ChatGPT voice natively in Apple cart play.
• 10:42
  OK, that is a very dense stack of strategic moves.
• 10:45
  Is a blitz, Yeah. And looking at all this, I have
• 10:47
  to ask you, we have Anthropic suddenly buying a $400 million
• 10:50
  biotech company and open AI putting chat bots in our car
• 10:54
  dashboards. Are they spreading themselves
• 10:55
  entirely too thin, or is this massive horizontal expansion
• 10:58
  just like table stakes for survival now?
• 11:00
  It's definitely table stakes, but it also highlights a huge
• 11:03
  divergent in their strategies. We'll look at open AI.
• 11:07
  Putting ChatGPT in Apple CarPlay means they are embedding
• 11:10
  themselves into your daily consumer life.
• 11:13
  They want to be the ambient intelligence in your car, your
• 11:15
  phone, your home. Right, they want to be
• 11:17
  everywhere you are. Exactly.
• 11:19
  But trying to be the universal utility for everyday people
• 11:22
  creates massive organizational strain, which probably explains
• 11:26
  some of that executive shuffling.
• 11:27
  The human toll of hyper growth. Yeah, but then look at
• 11:30
  Anthropic. They are leaning heavily into
• 11:33
  vertical high value enterprise solutions.
• 11:36
  Spending 400 million on coefficient bio tells us they
• 11:40
  believe the real money isn't in helping you write an e-mail.
• 11:43
  It didn't like curing diseases. Right.
• 11:45
  It's in solving complex multibillion dollar scientific
• 11:48
  problems like drug discovery. Clod is already known for its
• 11:52
  analytical capabilities, so it's perfect for massive biotech data
• 11:55
  sets. So they're just playing totally
• 11:57
  different games. Basically, survival means you
• 12:00
  either completely own the consumer ecosystem or you
• 12:02
  utterly dominate high end enterprise utility.
• 12:05
  Well, Speaking of enterprise, what about that Mercury breach?
• 12:08
  I mean 4 terabytes of data from a $10 billion startup.
• 12:13
  How does a company that big get hacked through a random open
• 12:17
  source library? It is the terrifying reality of
• 12:19
  supply chain attacks. LLM is just a popular free tool
• 12:23
  developers use to route request between different models.
• 12:28
  Merker integrated this free tool to save time, so the hackers
• 12:31
  didn't actually breach Merker's Fortress directly.
• 12:34
  They compromise the free tool that Mercur willingly brought
• 12:38
  inside the gates. Like a Trojan horse.
• 12:40
  Precisely. We get so distracted by these
• 12:43
  shiny new models that we forget these billion dollar systems are
• 12:46
  built on thousands of interconnected free
• 12:48
  dependencies. A vulnerability upstream just
• 12:51
  cascades right down. It's a very stark reminder that
• 12:54
  our security architecture is lagging dangerously behind our
• 12:57
  AI capabilities. It really isn't.
• 12:58
  You know, hearing about billion dollar acquisitions and TB data
• 13:02
  breaches, it can all feel a little detached from real,
• 13:04
  definitely. But the ultimate test of all
• 13:06
  this innovation is how it filters down to everyday people
• 13:10
  solving hyperlocal problems. Which brings us to our final
• 13:14
  segment, Empowering the Hobbyist.
• 13:16
  This is my favorite kind of story.
• 13:18
  Me too. So we have this reader, Kevin R
• 13:21
  out of Terre Haute IN. He's a 40 plus year RC car
• 13:25
  hobbyist and a forum moderator and his community was facing a
• 13:29
  problem. A new radio system, right?
• 13:31
  Yeah. This company, Radiomaster, came
• 13:33
  out with a new system using Edge TX firmware and it had a
• 13:36
  brutally steep learning curve for Surface based RC users.
• 13:40
  People just wanted to drive your cars, not become software
• 13:42
  engineers. Right, a classic adoption
• 13:44
  hurdle. So Kevin, who has 0 coding
• 13:46
  experience, none whatsoever, decides to build a software
• 13:50
  wizard app to help his community program their cars.
• 13:52
  I love this. He used clawed, Gemini and grok.
• 13:55
  He bounced between all three, and Claude actually generated
• 13:59
  the first working version of the app for him.
• 14:01
  Kevin said It hooked him entirely on AI development.
• 14:04
  It completely lowers the barrier to entry.
• 14:06
  It does. I mean, I read this in thought.
• 14:08
  It's like having a team of senior software engineers just
• 14:11
  hanging out in your garage, patiently helping you tinker on
• 14:14
  a Saturday. That's a great way to look at
• 14:16
  it. But I do have a push back here.
• 14:18
  Does this mean the traditional steep learning curve for niche
• 14:21
  hobbies is just dead? Or are we just trading one
• 14:25
  problem for another? Like will relying on AI just
• 14:29
  introduce a totally new technical dependency for
• 14:32
  hobbyists? Well, it's the pivotal tension
• 14:34
  of this whole era. To answer your first point, yes,
• 14:37
  the traditional steep learning curve for syntactical coding is
• 14:40
  effectively dead for hobbyists. Gone, because for decades the
• 14:45
  barrier wasn't having a good idea.
• 14:47
  It was knowing exactly where to put the brackets and semicolons
• 14:50
  to tell the machine what to do. LLNS completely bypassed that.
• 14:54
  They translate human intention directly into machine action.
• 14:57
  So Kevin had the domain expertise, he knew what the RC
• 15:00
  car needed to do, he just didn't speak computer.
• 15:03
  Exactly. And look at his workflow too.
• 15:05
  He didn't just use one model, he used Claude, Gemini, and Grok.
• 15:09
  Yeah, I wanted to ask about that.
• 15:11
  Why not just stick to Claude if it built the working version?
• 15:14
  Because multi model usage is becoming the standard for
• 15:17
  non-technical builders. Different models have different
• 15:20
  architectural blind spots. They act as like collaborative
• 15:24
  debuggers. Oh, I see if.
• 15:25
  Claude gets stuck on a logic error.
• 15:28
  You paste it into Gemini to troubleshoot, then you take it
• 15:31
  to Grok to refine the UI. It's literally getting second
• 15:35
  and third opinions from different engineering
• 15:37
  departments. It's a virtual dev studio.
• 15:39
  Right. But to your point about
• 15:41
  dependency, yes, there is a risk.
• 15:43
  If Kevin's app breaks next year because an API changes and the
• 15:47
  AI models can't figure it out, he might not have the
• 15:50
  foundational knowledge to patch it manually.
• 15:52
  Right, because he didn't learn the syntax.
• 15:54
  Exactly. We are moving from a world where
• 15:56
  you need to know how to build tools to a world where you just
• 15:58
  manage the entities that build the tool.
• 16:00
  Is it being the factory manager instead of the mechanic on the
• 16:02
  floor? Beautifully put.
• 16:04
  You manage the outcomes, not the mechanisms.
• 16:06
  And yes, it creates new dependencies.
• 16:08
  But empowering a guy in Indiana to solve a global community
• 16:13
  problem from his garage without a computer science degree?
• 16:17
  That is a profound net positive. It really is an inspiring note
• 16:21
  to end on. But before we wrap up, what is
• 16:24
  one final thought you want to leave us with today?
• 16:26
  Well, if we step back and look at the whole picture, consider
• 16:29
  this. If an AI like Netflix's Void,
• 16:33
  can Florida State simulate the physical consequences of reality
• 16:36
  predicting gravity and mass in a video?
• 16:39
  Yeah, and another AI can instantly build a functional
• 16:43
  software a pulled purely from the imagination of a hobbyist
• 16:46
  who doesn't know how to code. OK, how long until the barrier
• 16:50
  between merely imagining an entirely new digital world and
• 16:53
  actively stepping inside of it is completely erased?
• 16:56
  Wow, that is a heavy, incredible thought to leave on.
• 16:59
  If you want to keep exploring these shifts with us, make sure
• 17:02
  to subscribe so you stay updated and please take a second to rate
• 17:05
  the show 5 stars. It really helps us keep bringing
• 17:07
  you these deep dives. We will see you back here for
• 17:09
  our next deep dive tomorrow.