Latest / Today’s AI News / Elon and Anthropic Team Up on Compute, Murati Says Altman Lied, Google DeepMind Goes to Space

Transcript

Elon and Anthropic Team Up on Compute, Murati Says Altman Lied, Google DeepMind Goes to Space · Today’s AI News

• 0:00
  Imagine Batman suddenly renting out the bat cave to the Joker
• 0:04
  just because they both happen to hate Superman.
• 0:07
  That's exactly what it feels like today.
• 0:09
  As two massive AI rivals forged an incredibly unexpected
• 0:14
  alliance, the landscape of AI computing power just shifted
• 0:18
  beneath our feet. It really did.
• 0:20
  I mean, we're talking about a total realignment of the
• 0:22
  industry. Yeah, exactly.
• 0:25
  Welcome back to today's AI News, brought to you by NYNEX
• 0:27
  Productions. Today we are taking a massive
• 0:29
  stack of tech reporting legal depositions and product release
• 0:34
  notes, and going on a deep dive to give you a rapid,
• 0:37
  comprehensive briefing on these huge shifts.
• 0:40
  Because honestly, computing power to run these frontier
• 0:42
  models is, well, it's the new oil, right?
• 0:45
  And what we are seeing today is this frantic scramble to secure
• 0:48
  that resource. The alliance is forming to
• 0:50
  control this infrastructure. Are you know they're redrawing
• 0:53
  the map of the entire tech sector?
• 0:55
  Really are. And we are looking right at
• 0:56
  Anthropic and SpaceX for that. Or SpaceX AI as it's apparently
• 1:00
  being called. Now, yeah, that new branding,
• 1:01
  right? They just signed a massive new
• 1:04
  compute deal. So Anthropic is leasing all of
• 1:07
  Colossus One. And for you listening, that is
• 1:09
  Spacex's 300 plus MW supercluster down in Memphis.
• 1:13
  Which is just staggering. It's huge.
• 1:16
  We're talking about more than 220,000 NVIDIA GPU's coming
• 1:21
  online for Anthropic like within the month.
• 1:23
  Yeah, and to put 220,000 GPU's into perspective, that level of
• 1:28
  silicon requires, you know, dedicated industrial scale power
• 1:32
  grids. Yeah, and massive liquid cooling
• 1:35
  systems. It literally dictates whether an
• 1:38
  AI company can actually serve its users or if it's models just
• 1:42
  sort of choke under the daily traffic.
• 1:43
  Which we've all experienced, but we are already seeing the
• 1:46
  immediate benefits of this deal for the end user, right?
• 1:48
  Anthropic announced that clog codes 5 hour usage caps are
• 1:52
  doubling across all paid tiers. Exactly.
• 1:54
  So no more getting throttled during peak hours when you are
• 1:57
  like right in the middle of a complex coding project?
• 1:59
  Yes, and plus they're opening up additional rate increases via
• 2:02
  the API. It's just a sudden massive
• 2:04
  injection of bandwidth. OK, let's unpack this for a
• 2:07
  second. Is Elon Musk actually changing
• 2:09
  his mind about anthropics philosophy?
• 2:12
  Or is this just an enemy of my enemy is my friends situation?
• 2:15
  That's the big question. Right, Like is this specifically
• 2:17
  designed to starve open AI of compute power while feeding its
• 2:22
  biggest rival? Because, I mean, just a few
• 2:25
  months ago Musk was publicly posting that Anthropic should be
• 2:28
  renamed Misanthropic. Yeah, he did say that.
• 2:31
  He literally claimed they hate western civilization and now now
• 2:36
  he's tossing them the keys to a 220,000 GPU supercluster.
• 2:40
  What's fascinating here is the sheer pragmatism of the
• 2:43
  underlying mechanics. Musk recently tweeted that
• 2:46
  SpaceX will rent compute to AI companies, taking the right
• 2:49
  steps to ensure it is good for humanity.
• 2:52
  Right, the public. Pivot.
• 2:53
  Exactly. But looking past the public
• 2:55
  statements, this accomplishes several strategic goals for
• 2:57
  SpaceX AI. First, it helps patch a glaring
• 3:00
  compute deficit for Anthropic, who you know just happens to be
• 3:03
  Open AI's biggest direct competitor.
• 3:05
  Convenient, right? Very, by empowering Anthropic,
• 3:09
  he indirectly pressures open AI without having to out build them
• 3:13
  directly in the consumer market right this second.
• 3:16
  So he is basically arming the rival faction to keep his main
• 3:20
  adversary busy. That's one way to look at it,
• 3:22
  yeah. But what does this mean for
• 3:24
  SpaceX? A is own ambitions.
• 3:26
  It's because they are still building their own models,
• 3:28
  right? They are and that is the most
• 3:29
  critical shift for the long term industry structure here.
• 3:33
  It signals a completely new business model for them.
• 3:35
  Like what? Well, think about Amazon back in
• 3:38
  2006. Oh, with AWS.
• 3:40
  Exactly. They had massive internal server
• 3:43
  capacity to run their e-commerce store, and they realized that
• 3:47
  renting out that raw infrastructure was actually a
• 3:50
  more lucrative fundamental business than just selling
• 3:52
  books. That makes a lot of sense.
• 3:54
  So SpaceX AI is stepping into that exact dual role.
• 3:57
  Right, they are pushing their own model, Grok to reach the
• 4:01
  frontier of capabilities, but simultaneously they are acting
• 4:04
  as a compute landlord to their direct rivals.
• 4:07
  And the scale of the money required to be that kind of
• 4:10
  landlord is just, it's hard to comprehend.
• 4:12
  It's astronomical. Literally.
• 4:14
  The Information just reported that Anthropic is also
• 4:17
  committing to a $200 billion five GW compute deal with Google
• 4:22
  Cloud over the next five years. 5 gigawatts, yeah. 5 gigawatts
• 4:26
  is enough to power a major metropolitan city several times
• 4:30
  over. It just illustrates that
• 4:31
  securing raw compute is the single biggest bottleneck in the
• 4:35
  AI industry right now. You can design the most elegant
• 4:38
  neural network architecture in the world, but.
• 4:40
  But without the physical infrastructure to run it, it
• 4:42
  doesn't matter. Exactly.
• 4:44
  Without the power and the chips to train it, your models remain
• 4:47
  purely theoretical. So we are seeing this massive
• 4:50
  external war over securing physical compute, but there is
• 4:55
  an equally intent internal war happening right now over
• 4:57
  corporate governance. Yes, absolutely.
• 4:59
  Let's transition from Musk's strategic business moves against
• 5:02
  Open AI to his literal, ongoing legal battles with them in the
• 5:06
  courtroom. The trial is giving the public a
• 5:08
  very rare, unvarnished look inside the leadership dynamics
• 5:12
  of a major AI lab at a really critical juncture.
• 5:15
  It really is, Ex Open AI Chief Technology Officer Mira Murati
• 5:20
  just testified via video deposition in Musk's lawsuit
• 5:23
  against the company. And the details coming out of
• 5:25
  this deposition are highly specific.
• 5:27
  Yeah, she accused CEO Sam Altman of lying about a model skipping
• 5:31
  a safety review. Which is a huge claim.
• 5:33
  Right. The mechanics of this claim are
• 5:36
  that Altman told her the legal team had cleared a specific
• 5:39
  model to bypass standard safety protocols, but Murati went and
• 5:43
  verified this directly with their counsel, Jason Kwan, and
• 5:46
  found out that no such clearance was ever actually given.
• 5:49
  And, you know, in an industry where red teaming and safety
• 5:52
  protocols are the primary mechanisms for preventing
• 5:55
  catastrophic model failures. It's basically all we have to
• 5:58
  keep these things safe. Exactly.
• 6:00
  So the claim that ACEO might unilaterally bypass them and
• 6:03
  allegedly misrepresent legal clearance to do so, it goes to
• 6:07
  the very heart of corporate governance.
• 6:09
  It really suggests a breakdown in the structural checks and
• 6:11
  balances of the organization. Yeah.
• 6:13
  She also highlighted that Altman was giving conflicting
• 6:15
  directions to different executives.
• 6:17
  She stated it was creating chaos across the leadership team and
• 6:21
  undermining her authority as CTO.
• 6:24
  And all of this culminated in that chaotic 2023 board drama,
• 6:28
  right? Which Murati noted put the
• 6:30
  entire company at risk of falling apart.
• 6:33
  We also saw testimony from former board member Helen Toner,
• 6:37
  who reportedly stated that Murati was afraid to stick her
• 6:40
  neck out during that time. Oh wow.
• 6:42
  Afraid of what? Due to fear of blowback for her
• 6:44
  career, it outlines a highly fractured internal culture at
• 6:48
  the exact moment the company was scaling its most powerful
• 6:51
  models. Here's where it gets really
• 6:53
  interesting. To understand the legal weight
• 6:55
  of this, we have to look beyond the interpersonal conflicts.
• 6:58
  Definitely. It's like watching an open heart
• 7:00
  surgery on a corporate entity where the lawyers are trying to
• 7:02
  dissect whether it operates like a nonprofit charity or a hyper
• 7:06
  capitalist tech giant. That's a great way to put it.
• 7:09
  And we have to be completely objective here.
• 7:13
  And remember, these are just claims being made in a courtroom
• 7:15
  for a jury to ultimately decide. But why does this specific
• 7:20
  testimony about internal chaos matter so much to Musk?
• 7:25
  Score legal argument. From a legal strategy
• 7:27
  perspective, this testimony is ammunition for the foundational
• 7:31
  premise of Musk's lawsuit. Which is what?
• 7:33
  Exactly. The core claim is that Altman
• 7:36
  and Greg Brockman were trying to steal a charity back in 2017
• 7:41
  when they shifted the corporate structure from a pure nonprofit
• 7:44
  to a capped profit model. So Murati's testimony about
• 7:47
  misdirection and bypassing safety protocols provides must
• 7:51
  legal team with evidence to argue that the current
• 7:53
  leadership operates with a lack of transparency.
• 7:55
  Right. A lack of transparency that
• 7:57
  violates the foundational trust and safety mission of that
• 7:59
  original nonprofit charter. Exactly.
• 8:02
  Well, if open AI is struggling with internal chaos at the
• 8:04
  leadership level, Anthropic, the rival Musk is now backing, is
• 8:08
  actually releasing tools designed to make your daily work
• 8:11
  less chaotic. Yeah, a nice pivot to something
• 8:13
  practical. Yeah, right.
• 8:14
  Let's look at how AI is being used to structure massive
• 8:18
  amounts of messy human data. We have a very practical guide
• 8:22
  on how to use clawed design to optimize workflows.
• 8:25
  The specific use case here is taking a raw unstructured CSV
• 8:29
  file like say exported YouTube channel analytics or Facebook ad
• 8:34
  campaign results. Oh, those are always a mess to
• 8:36
  look at. Totally, and turning it into a
• 8:38
  cohesive strategy deck. Wait, I'm skeptical.
• 8:41
  Every time I've tried to make an AI build a slide deck in the
• 8:43
  past, it looks like a middle schooler's PowerPoint from 2004.
• 8:46
  Yeah, just terrible formatting. It usually just slaps random
• 8:50
  bullet points onto a gradient background.
• 8:53
  What is actually different about the underlying architecture here
• 8:56
  that makes it usable for a professional presentation?
• 8:59
  The difference is really in the semantic parsing and how it maps
• 9:03
  data relationships to visual hierarchy.
• 9:05
  OK, what does that mean in practice?
• 9:07
  Well, older tools just summarize text and pasted it onto a
• 9:09
  template. Claw design actually analyzes
• 9:12
  the quantitative relationships in your CSV.
• 9:14
  The workflow is to go to Claude dot a design, select the slide
• 9:17
  deck option, and crucially toggle on the speaker notes.
• 9:20
  Oh, the speaker notes Coggle is key.
• 9:22
  Yes, when you upload your data, you don't just ask for a
• 9:25
  presentation, you direct the semantic weight.
• 9:29
  So the prompt strategy is the secret sauce here.
• 9:31
  You tell it. Turn these files into a strategy
• 9:35
  deck on performance, analyze the results by item, extract best
• 9:39
  practices, use charts and rankings, and provide concrete
• 9:42
  recommendations. Exactly.
• 9:43
  You can even tell it to match your creative images to the CSV
• 9:47
  data using the actual file names.
• 9:49
  That is super helpful, yeah. And it builds the structure in
• 9:52
  about 10 to 15 minutes, which you then export directly to
• 9:55
  PowerPoint or Google Slides. Right.
• 9:57
  And that there is a powerful workflow technique here too,
• 9:59
  which is duplicating the project.
• 10:01
  How so? You can layer in multiple
• 10:03
  different data sources, say your ad spend data cross referenced
• 10:07
  with your website conversion data.
• 10:10
  Claude will synthesize the relationships between those
• 10:12
  different data sets into a single cohesive visual
• 10:15
  narrative. OK, so does this completely
• 10:17
  replace the junior analyst or is it just a formatting shortcut to
• 10:21
  get past the blank page syndrome?
• 10:23
  It completely inverts the traditional ratio of Labor.
• 10:26
  Typically, an analyst spends 80% of their time formatting slides,
• 10:30
  cleaning data, and aligning charts right.
• 10:33
  Yeah, big groundwork. Exactly, leaving 20% for
• 10:35
  strategic thinking. By handling the data structuring
• 10:39
  and visual layout, this architecture forces the human
• 10:43
  worker to focus entirely on the So what implications of the
• 10:47
  data. I love that you become the pilot
• 10:49
  directing the strategy and the AI is merely the navigator
• 10:53
  executing the map. Precisely.
• 10:54
  So we just talked about using AI to map out past human behavior
• 10:58
  in a spreadsheet, but DeepMind is trying to figure out how AI
• 11:01
  can predict future human behavior.
• 11:03
  Yeah, and their approach is wild.
• 11:05
  Do that. They're putting it inside a
• 11:07
  virtual Galaxy. Google DeepMind just bought a
• 11:10
  minority stake in a studio called Fenris Creations.
• 11:12
  Which is a huge deal for gaming and AI.
• 11:15
  Right, They are a spinoff from CCT Games, the creators of the
• 11:18
  23 year old space simulation game EV Online.
• 11:22
  DeepMind is going to use an offline clone of EVE as a
• 11:25
  massive AI testbed. This represents A fundamental
• 11:28
  leap in how we evaluate artificial intelligence beyond
• 11:31
  simple language tasks. EVE Online is a highly unique
• 11:35
  digital environment. Yeah, I've heard crazy stories
• 11:38
  about that game. Unlike most massive multiplayer
• 11:41
  games that segment players into different servers, EVE runs on a
• 11:45
  single continuous server. Meaning everyone is in the same
• 11:48
  universe together. Exactly.
• 11:50
  And the players dictate the entire universe.
• 11:52
  They form complex corporate structures, run highly realistic
• 11:56
  supply chain economies, set market prices based on resource
• 11:59
  scarcity, and engage in massive fleet battles.
• 12:02
  Fleet battles that routinely destroy digital assets worth
• 12:05
  hundreds of thousands of real world dollars?
• 12:08
  Yes. It is a notoriously ruthless
• 12:10
  game and Deepmine is going to run AI agents in this
• 12:13
  environment to test how they reason over long timelines,
• 12:16
  retain memory, and learn in a space where there is no clear
• 12:19
  finish line. Demis Hassadis, the head of
• 12:21
  Deepmine, has previously used games like Atari, Alphago and
• 12:25
  StarCraft to train AI, but the CEO of Fenders Creations pointed
• 12:29
  out that EVE behaves like a living world.
• 12:32
  Which is a huge distinction from those other games.
• 12:35
  So what does this all mean? We used to teach AI to beat us
• 12:37
  at chess. Now it sounds like we are
• 12:39
  putting them into a digital cut throat Ant farm of space
• 12:42
  capitalists just to see how they evolve.
• 12:45
  Why is this necessary? If we connect this to the bigger
• 12:48
  picture, it exposes the limitations of current AI
• 12:51
  training environments. Chess or Go are win State Games.
• 12:55
  Right. You either win or lose.
• 12:56
  They have. Rigid rules, perfect information
• 12:58
  on the board, and a binary outcome.
• 13:00
  The real world is an open-ended, unpredictable, highly complex
• 13:05
  societal simulation with hidden variables.
• 13:07
  So reinforcement learning struggles with.
• 13:09
  That exactly. It often fails in the real world
• 13:12
  because the policies it learns are too brittle for human
• 13:14
  unpredictability. EVE Online is the closest
• 13:17
  digital approximation we have to a living, breathing economy
• 13:21
  driven by human ecology, betrayal, alliances and market
• 13:25
  manipulation. Wow, so the agents have to learn
• 13:27
  how to survive supply chain disruption and digital space
• 13:30
  pirates before we can trust them to manage a real world logistics
• 13:33
  network? That's the idea.
• 13:35
  OK, let's zoom out from the digital Galaxy of DeepMind and
• 13:39
  do a rapid fire scan of the rest of the AI universe with today's
• 13:42
  quick hits. Let's do it.
• 13:44
  First up, Anthropic launch new features for managed agents
• 13:47
  including Outcomes, Multi Agent Orchestration and Dreaming.
• 13:51
  This allows agents to study past sessions, grade their own work,
• 13:54
  and split up complex tasks. That dreaming function is really
• 13:58
  interesting. It's a form of offline
• 14:00
  reinforcement learning. Like how we sleep to process the
• 14:03
  day. Exactly.
• 14:04
  It allows the agent to consolidate memories and
• 14:06
  optimize its pathways without actively burning compute on live
• 14:10
  user interactions. That is wild over at Open AI.
• 14:13
  They just made GPT 5.5 instant the new default model across
• 14:18
  ChatGPT for everyone. Nice upgrade for the free tier.
• 14:21
  Definitely. They also announced a massive
• 14:23
  hardware partnership, teaming up with AMD, Intel, NVIDIA,
• 14:26
  Microsoft and Broadcom to launch MRC.
• 14:29
  It is an open source tool designed to prevent hardware
• 14:32
  failures from ruining giant training runs mid session.
• 14:35
  Which is crucial when you're dealing with these massive
• 14:37
  clusters. But wait.
• 14:38
  OK, MRC prevents hardware failures, but how does it
• 14:42
  actually do that when you are running like 100,000 GPU's at
• 14:45
  once? Well.
• 14:46
  When you are running a cluster that large, the statistical
• 14:50
  probability of a chip burning out during a month long training
• 14:53
  run is essentially 100%. Guaranteed failure.
• 14:56
  Yeah, so MRC acts as a distributed checkpointing
• 14:59
  system. It continuously saves the state
• 15:02
  of the model across the cluster, so if a single GPU fails, it
• 15:05
  doesn't corrupt the entire matrix.
• 15:06
  You simply swap the hardware in resume.
• 15:08
  Right, saving millions of dollars in lost compute time.
• 15:11
  Total Sense Next, a startup called Sub Q debuted a new model
• 15:16
  claiming a 12 million token context window paired with A52X
• 15:20
  speed boost on long tasks compared to rivals. 12 million,
• 15:24
  that's just. I know, think of 12 million
• 15:27
  tokens as feeding the AI the entire text of the Harry Potter
• 15:30
  series 15 times over in a single breath.
• 15:33
  That's a great visual. How is a normal tech follower
• 15:35
  supposed to keep up with these numbers though?
• 15:37
  Honestly, you don't need to memorize the model names or the
• 15:40
  specific token counts. The practical take away for you
• 15:43
  is that the barrier to analyzing massive data sets has basically
• 15:46
  vanished. Meaning you can just dump
• 15:48
  everything in. Yeah, a 12 million token window
• 15:50
  means you aren't just uploading APDF.
• 15:52
  You can upload an entire corporation's historical
• 15:55
  archives or a massive library of software code in one single
• 15:59
  prompt, and the 52X speed boost means you get the analysis
• 16:03
  without the crazy latency. That is powerful.
• 16:07
  Moving on, we also saw the launch of real time TTS 2A new
• 16:11
  voice AI that actually listens to the user to dynamically match
• 16:15
  tone and emotion. Which is going to make
• 16:17
  conversational agents so much more natural.
• 16:19
  Chinese AI startup DeepSeek is reportedly nearing a funding
• 16:22
  round, valuing them at $45 billion.
• 16:25
  And finally, Google announced A partnership with Tune core
• 16:28
  parent company Believe, bringing their Flow Music and Lyria 3 Pro
• 16:31
  AI models directly to musical artists to aid in production.
• 16:35
  It really highlights how rapidly multimodal generation, you know,
• 16:38
  audio, music and voice is catching up to text based
• 16:41
  models. It's happening fast.
• 16:43
  So to ground all of this abstract model news, let's look
• 16:47
  at a specific physical application in our Community AI
• 16:50
  Workflow segment. This comes from an anonymous
• 16:53
  listener who is currently recovering from a torn ACL.
• 16:56
  This is a master class in chaining different model
• 16:59
  architectures together to bypass traditional software
• 17:02
  development. I loved reading about this.
• 17:04
  It's so clever so their physiotherapist films custom
• 17:07
  exercises, narrating the sets, reps and leaving voice notes on
• 17:11
  proper form and sends these raw videos over WhatsApp.
• 17:14
  Pretty standard physical therapy experience.
• 17:17
  Right, but the listener uploaded those raw videos directly into
• 17:20
  Google's Gemini. They asked Gemini to analyze the
• 17:24
  video kinematics and the audio, to understand the workout plan,
• 17:27
  and to output a highly specific structured prompt detailing the
• 17:31
  logic of those exercises. And here's the anatomy of why
• 17:33
  this works so well. They used Gemini specifically
• 17:36
  for its superior spatial and temporal reasoning.
• 17:39
  Because it can actually see the movement.
• 17:41
  Exactly. It extracted the visual data
• 17:43
  like the angle of the knee, the duration of the hold, and
• 17:46
  converted that physical reality into structured text logic.
• 17:49
  And then they took that structured prompt generated by
• 17:52
  Gemini and fed it into Anthropics Claude code.
• 17:55
  Claude code then one shotted it, meaning it successfully compiled
• 17:59
  the code on the first attempt without errors, creating a
• 18:02
  custom tracking app. Which is just incredible, right?
• 18:06
  This bespoke app tracks their sets, reps, weights and daily
• 18:11
  notes over months of recovery and automatically exports a
• 18:14
  clean CSV file to send back to the physio.
• 18:17
  By feeding the visual reasoning of 1 model into the coding
• 18:20
  expertise of another, they completely bypass the need to
• 18:24
  hire a developer or settle for a generic fitness app.
• 18:27
  They built exactly what they needed.
• 18:29
  Yeah, they used AI as a personal infrastructure builder to create
• 18:32
  medical tracking software tailored perfectly to their
• 18:34
  specific injury, all from their phone in a matter of minutes.
• 18:38
  It just shows you don't have to wait for the market to build the
• 18:40
  tool you need, you can orchestrate the intelligence to
• 18:42
  build it for you. Well, that covers our deep dive
• 18:45
  for today. We started with massive
• 18:47
  industrial data centers in Memphis and ended with a custom
• 18:50
  app for a torn ACL. It's been a huge day for AI and
• 18:55
  as we wrap up, I want to leave you with a thought regarding
• 18:57
  that DeepMind project we talked about earlier.
• 18:59
  The EVE Online simulation. Yes, this raises an important
• 19:03
  question. If an AI agent successfully
• 19:06
  learns to survive, deceive and manipulate trade within a
• 19:10
  roofless 23 year old virtual society like EVE Online, what
• 19:15
  happens when we deploy that exact same agent architecture
• 19:18
  into our real world global financial markets?
• 19:21
  Oh wow, now there is a thought to keep you up at night.
• 19:25
  Let's just hope they don't treat the stock market like a space
• 19:27
  pirate flake. Seriously.
• 19:29
  Thank you for joining us on this deep dive into the latest
• 19:31
  developments in AI. If you want to make sure you
• 19:34
  never miss a shift in this fast moving landscape, be sure to hit
• 19:37
  subscribe to stay updated. And if you enjoyed unpacking all
• 19:40
  this with us, we'd love if you could rate the show 5 stars.
• 19:42
  It really helps us out. The Batcave might be under new
• 19:45
  management, but we'll be right here to help you make sense of
• 19:47
  it all. See you back for tomorrow's
• 19:49
  episode.