Latest / AI Fire Daily / 🎙️ EP 339: Mysterious Model Rules OpenRouter & TARS Releases Embodied Model AWE 3.5

Transcript

🎙️ EP 339: Mysterious Model Rules OpenRouter & TARS Releases Embodied Model AWE 3.5 · AI Fire Daily

• 0:00
  We are looking at a true Phantom Titan today.
• 0:02
  Yeah, a massive top -tier AI model just dropped
• 0:05
  out of nowhere. For free. And nobody knows who
• 0:08
  built it. It is completely shaking up the developer
• 0:10
  community right now. Welcome to this deep dive.
• 0:12
  We have a truly fascinating journey ahead for
• 0:15
  you today. We really do. We are exploring this
• 0:18
  mysterious model known as OxAlpha. It is a massive
• 0:22
  digital enigma. Right. And then we will unpack
• 0:25
  our exploding digital ecosystem. From autonomous
• 0:28
  agents writing complex code to some massive company
• 0:32
  valuations. Like a staggering $13 billion valuation
• 0:36
  for Hugging Face. We will also explore everyday
• 0:39
  consumer AI tools. You know, some of the recent
• 0:43
  controversies, too. Finally, we will examine
• 0:45
  a massive breakthrough in physical robotics.
• 0:48
  AI is finally getting a unified physical body.
• 0:50
  The physical world really is the ultimate frontier.
• 0:53
  P -Dan, let's start with this Phantom Titan.
• 0:56
  Let's do it. The sheer scale of Oxalpha is staggering.
• 0:59
  It appeared on Open Raider just last week. Yeah,
• 1:01
  it was a completely stealth launch. It came totally
• 1:04
  out of the blue. Open Router explicitly states
• 1:07
  it is from an anonymous third party. Nobody has
• 1:10
  publicly confirmed who built it yet. Developers
• 1:12
  are definitely speculating online, though. Oh,
• 1:14
  absolutely. They're guessing it might be Z .AI
• 1:17
  or GLM. Some think it could be Microsoft. Right,
• 1:20
  but there's currently no verified attribution
• 1:23
  anywhere. The underlying specs are absolutely
• 1:27
  insane. They really are. It is built heavily
• 1:30
  for complex coding. It excels at long -running
• 1:33
  agent tasks. And it handles deep reasoning beautifully.
• 1:37
  Plus it has a massive context window. We define
• 1:39
  that as the memory span an AI has for a single
• 1:42
  conversation. Exactly. And it sits at over 1
• 1:45
  million tokens. 1 ,048 ,576 tokens, to be precise.
• 1:50
  Think of that like holding several thick novels
• 1:52
  in its memory at once. That is a staggering amount
• 1:55
  of recall. And its output capacity is equally
• 1:58
  wild. It can output up to 131 ,072 tokens. Two
• 2:03
  sec silence. Let's just pause on that. That means
• 2:06
  it doesn't just write a single function. It can
• 2:08
  write an entire application in one go. Right.
• 2:11
  And it is highly versatile, too. It accepts text,
• 2:14
  images and video as input. It is a true multimodal
• 2:19
  system. Meaning it. processes text images and
• 2:22
  video simultaneously it also supports complex
• 2:25
  tool calling and it handles structured outputs
• 2:28
  perfectly and right now it is completely free
• 2:31
  during its preview phase that is honestly the
• 2:34
  wildest part of this story open code advertised
• 2:36
  a massive pool capacity for it we are talking
• 2:39
  up to 100 trillion tokens per day that is near
• 2:43
  unlimited usage for the global developer community
• 2:46
  the compute cost for that is astronomical it's
• 2:49
  like some Someone dropping a state -of -the -art
• 2:51
  Ferrari in a public parking lot with the keys
• 2:53
  in the ignition and no license plates. That is
• 2:55
  the perfect analogy. And, you know, people are
• 2:57
  definitely taking it for a spin. It is getting
• 2:59
  serious real -world traction. Stripe CEO Patrick
• 3:02
  Collison publicly praised its performance recently.
• 3:05
  Early independent testing is also yielding very
• 3:07
  interesting results. A full deep SWE benchmark
• 3:10
  run was completed. SWE stands for software engineering.
• 3:14
  This tests real -world coding problems. And it
• 3:17
  solved 66 out of 113 software engineering tasks.
• 3:22
  That is a 58 .4 % success rate. That is highly
• 3:28
  impressive for a totally unknown model. It navigates
• 3:31
  actual code bases and fixes deep bugs. But the
• 3:34
  core mystery still remains. Beat. Why would a
• 3:38
  developer spend millions to build this and release
• 3:40
  it anonymously? I strongly suspect it is a stealth
• 3:43
  stress test. A stress test. Yeah. A massive company
• 3:46
  wants to see how this model handles immense,
• 3:48
  unpredictable user loads. They want to test it
• 3:51
  before attaching their brand name to it. Exactly.
• 3:53
  They need to know it won't break. So we are basically
• 3:55
  free beta testers for a tech giant secret weapon.
• 3:59
  We are doing their global quality assurance for
• 4:02
  free. It makes sense. Now, if OxAlpha is the
• 4:05
  mysterious giant entering the software world,
• 4:08
  the environment it's entering is rapidly shifting.
• 4:10
  The whole ecosystem is moving toward autonomous
• 4:12
  agents. And the market valuations are absolutely
• 4:15
  exploding. Let's talk about the startup called
• 4:17
  Inherent. The team consists of DeepMind alumni.
• 4:20
  They built an agent called Faraday. This agent
• 4:23
  recently beat both Anthropic and OpenAI. It excelled
• 4:27
  at reproducing complex research papers. And it
• 4:30
  did this despite running on a much smaller quen
• 4:32
  model. It proves that architecture often matters
• 4:35
  more than raw model size. It is all about how
• 4:38
  the agent's logic is structured. I mean, I still
• 4:41
  wrestle with prompt drift myself. So seeing an
• 4:44
  agent handle that level of complexity autonomously
• 4:46
  is humbling. It really changes our perspective
• 4:48
  on what is possible. And the infrastructure for
• 4:51
  these agents is expanding rapidly. Construct
• 4:53
  Computer is a great example of this trend. They
• 4:55
  are literally giving AI agents their own virtual
• 4:58
  computer. The agents can install specific skills.
• 5:01
  They can build custom tools for your whole team
• 5:03
  to use. It creates a dedicated digital workspace
• 5:06
  for the AI. Then you have Google Anti -Gravity.
• 5:09
  They just updated it to support remote control.
• 5:11
  This means you can keep working with your agents
• 5:13
  from another device. You can start a complex
• 5:15
  task on your desktop. Then you can continue it
• 5:18
  seamlessly on your phone. It is true cross -device
• 5:21
  integration. We are also seeing heavy innovation
• 5:24
  in dedicated coding environments. Fetch Sandbox
• 5:28
  MCP is a perfect example of this. It uses over
• 5:31
  70 different API sandboxes. And it runs directly
• 5:34
  inside... cursor or cloud code. It doesn't just
• 5:38
  write code for you. It fixes your code and proves
• 5:40
  the solution actually works in a sandbox. These
• 5:43
  platforms are becoming incredibly valuable infrastructure.
• 5:47
  Look at the current situation with Hugging Face.
• 5:49
  They are reportedly exploring some massive acquisition
• 5:52
  offers. The numbers being discussed are staggering.
• 5:55
  The offers are worth at least $13 billion. That
• 5:58
  is nearly three times its valuation from 2023.
• 6:01
  It was valued at $4 .5 billion back then. This
• 6:04
  shows how incredibly valuable AI developer platforms
• 6:07
  have become. The foundational layer of AI development
• 6:10
  is a goldmine. The hustle is very real everywhere
• 6:13
  you look, honestly. Just look at the Grok .bot
• 6:15
  domain situation. That story is a perfect example
• 6:18
  of the current frenzy. A buyer grabbed the Grok
• 6:21
  .bot domain weeks before Grok's latest launch.
• 6:25
  It is totally unaffiliated with the actual Grok
• 6:27
  bot. Now they are trying to flip the domain for
• 6:30
  a profit. They are asking for a wild price. People
• 6:33
  are aggressively trying to cash in on the rapid
• 6:35
  pace of change. The consumer side is changing
• 6:38
  just as fast, though. Grok Imagine recently got
• 6:41
  an updated discovery page. It removes a lot of
• 6:43
  user friction. You can jump between features
• 6:45
  much more easily now. You can do photo editing
• 6:48
  and image resizing completely seamlessly. Me
• 6:52
  Too One also released something highly impressive
• 6:53
  for consumers. It is called Long Cat Video Avatar
• 6:56
  1 .5. It creates incredibly stable talking avatars.
• 7:00
  You only need one image and a single voice clip
• 7:03
  to generate it. The lip sync is significantly
• 7:05
  tighter now. This matters because it bridges
• 7:08
  the uncanny valley. It makes the digital avatars
• 7:10
  feel much more trustworthy. Everyday digital
• 7:12
  friction is also being smoothed out by localized
• 7:15
  AI. OpenLogi is a fascinating example of this
• 7:18
  trend. It is a free, local -first alternative
• 7:21
  to Logitech Options Plus. It lets you remap your
• 7:24
  mouse buttons and control DPI. The best part
• 7:27
  is it requires no accounts. It has zero tracking.
• 7:29
  It processes everything locally for maximum privacy.
• 7:32
  Then there is Axiomote. It applies AI. To your
• 7:36
  daily commute. It turns your raw car data into
• 7:40
  easy to understand insights. It tracks over 20
• 7:42
  different vehicle metrics constantly. It monitors
• 7:45
  fuel costs and long -term driving trends. AI
• 7:48
  is optimizing literally everything we touch.
• 7:51
  But with this rapid growth comes inevitable friction.
• 7:54
  We have to look at the recent industry controversies.
• 7:57
  Yeah, the intersection of AI and commercial art
• 7:59
  is still very messy. Look at the recent Marvel
• 8:01
  debate. Spider -Man. Brand New Day got caught
• 8:05
  in a major AI art debate. Fans flagged some wonky
• 8:08
  concept art in the issue. A designer stated that
• 8:11
  Marvel actively screened for AI generation. Despite
• 8:14
  their best efforts, controversial art still slipped
• 8:17
  through the cracks. It highlights how difficult
• 8:19
  detection has become. The regulatory landscape
• 8:21
  is also shifting dramatically. We are seeing
• 8:24
  major reversals in corporate policy positions.
• 8:27
  OpenAI is a prime example of this shift. They
• 8:29
  previously opposed California's SB53 bill strongly.
• 8:32
  Now they actually want the law strengthened.
• 8:34
  They want it to include monitoring models during
• 8:37
  both training and evaluation. This is a complete
• 8:39
  reversal from their past opposition. It shows
• 8:42
  how quickly the industry's perspective on regulation
• 8:45
  is changing. As models become more powerful.
• 8:48
  Companies want different safeguards. We are just
• 8:51
  observing the changing dynamics here. Two sec
• 8:53
  silence. So, do these sandbox tools mean human
• 8:57
  coders are transitioning into mere supervisors?
• 9:01
  I think humans will focus entirely on system
• 9:03
  architecture rather than basic syntax. We're
• 9:05
  trading the keyboard for the conductor's baton.
• 9:08
  Exactly. We direct the overall flow of logic
• 9:10
  now. mid -roll sponsor break read provided separately
• 9:13
  welcome back we just discussed digital sandboxes
• 9:16
  and cross -device remotes but what happens when
• 9:18
  agents get a physical body this is where things
• 9:21
  get truly science fiction pars robotics just
• 9:24
  introduced something massive to the field it
• 9:27
  is called awe 3 .5 tars describes it as an embodied
• 9:31
  native foundation model We should unpack that
• 9:34
  specific term. It simply means an AI built directly
• 9:37
  for a physical robot body. It isn't just software
• 9:40
  inside a metal shell. The core philosophy driving
• 9:43
  this is fascinating. They call it born as one.
• 9:46
  Let's explore how this differs from traditional
• 9:48
  robotics. Historically, robotics involved connecting
• 9:51
  separate modules together later. You had one
• 9:54
  distinct module for visual perception. You had
• 9:56
  another completely separate module for path planning.
• 10:00
  And you had a third module for physical control.
• 10:02
  The latency between those modules caused constant
• 10:05
  errors. Exactly, because the systems weren't
• 10:07
  truly unified. Born as one changes that entirely.
• 10:11
  Instead of teaching a brain a set of eyes and
• 10:13
  an arm separately and hoping they cooperate,
• 10:15
  it's like growing a complete connected nervous
• 10:18
  system all at once. That is exactly what they
• 10:21
  are doing. It combines perception, action, geometry,
• 10:24
  and touch inside one system from the very beginning.
• 10:28
  Let's unpack the actual training loop for this
• 10:30
  model. It is a highly complex five -part process.
• 10:34
  The first part is the embodied native architecture
• 10:36
  itself. Then comes the critical dual prior pre
• 10:39
  -training phase. During this pre -training, the
• 10:41
  model learns two core priors before it even moves.
• 10:44
  First, it learns basic action patterns. Second,
• 10:47
  it learns spatial structure and fundamental physical
• 10:50
  laws. Think of it like giving a toddler a basic
• 10:52
  sense of physics. It is learning how gravity
• 10:55
  works before it even tries to walk. It learns
• 10:58
  that walls are solid early on. Then comes the
• 11:02
  intensive post -training phase. This involves
• 11:04
  something called the AI World Engine. This engine
• 11:07
  actively predicts possible future states in real
• 11:11
  time. It constantly compares what different physical
• 11:14
  actions may lead to. It uses those predictions
• 11:16
  to choose better actions instantly. It simulates
• 11:19
  the physics before executing the movement. Whoa.
• 11:23
  Imagine scaling to a billion real -world interactions
• 11:26
  being predicted in real time. It is entirely
• 11:29
  mind -bending. The whole system is continuously
• 11:31
  predicting, acting, and adjusting. The training
• 11:34
  loop includes two more vital parts. It uses scaling
• 11:37
  validation and continuous data feedback. The
• 11:40
  real -world results speak for themselves. TARS
• 11:42
  says AWE 3 .5 is incredibly stable. It can sustain
• 11:46
  several minutes of long -horizon closed -loop
• 11:49
  reasoning. That means it can perform a multi
• 11:51
  -step physical task. without losing its focus.
• 11:54
  That is a massive achievement in modern robotics.
• 11:56
  TARS also claims a huge jump in overall efficiency.
• 11:59
  They report roughly two times higher task execution
• 12:02
  efficiency than Pi 0 .5. They're positioning
• 12:05
  AWE 3 .5 as a general purpose model for physical
• 12:10
  AI. They are betting their future on this unified
• 12:12
  approach. They strongly believe more capable
• 12:15
  robots will come from training one single model.
• 12:19
  This model must understand, predict, and act
• 12:22
  in the physical world seamlessly. Beat. But what
• 12:25
  happens when a born -as -one model encounters
• 12:28
  an environment that violates its pre -trained
• 12:30
  physical laws? The continuous data feedback loop
• 12:33
  instantly corrects its assumptions based on new
• 12:36
  input. It updates its own internal physics engine
• 12:39
  on the fly. Precisely. It adapts to new realities
• 12:42
  like a living organism. Two -sex silence. Let's
• 12:45
  pull this all together. We have covered a massive
• 12:47
  amount of brown today. We really have. The pace
• 12:49
  of modern innovation is completely relentless.
• 12:52
  We started our journey with the digital mystery
• 12:53
  of aux alpha. We explored its massive context
• 12:56
  window and its free global preview. Then we moved
• 12:59
  through a rapidly expanding software ecosystem.
• 13:01
  We saw AI agents automating complex coding tasks
• 13:04
  seamlessly. We saw Hugging Face reach a $13 billion
• 13:08
  valuation. We saw the rise of everyday consumer
• 13:11
  AI tools, removing daily friction. Finally, we
• 13:14
  went straight into the physical world. TARS is
• 13:17
  literally giving AI a unified robotic nervous
• 13:20
  system. The AWE 3 .5 model is pushing physical
• 13:24
  AI forward aggressively. Beat. I want to leave
• 13:29
  you with a final thought today. It is something
• 13:31
  to really ponder deeply. We saw that anonymous
• 13:33
  developers can successfully deploy world -class
• 13:36
  digital brains like Oxalpha for free. We also
• 13:39
  saw companies proving that robotic bodies can
• 13:41
  be trained born as one. The digital and physical
• 13:44
  technologies are converging incredibly rapidly
• 13:46
  now. So how long before we see open source anonymous
• 13:49
  physical robotics models walking around in the
• 13:52
  real world? And who is responsible when an anonymous
• 13:55
  robot makes a mistake? That is the ultimate societal
• 13:57
  question. We have to face soon. Thank you for
• 13:59
  taking this deep dive with us today Keep questioning
• 14:02
  the rapid changes in the world around you 80
• 14:04
  row music