Latest / Today’s AI News / Anthropic’s Secret Mythos Model Leaks, Altman vs Amodei’s Decade-Long Feud, Anthropic Wins in Court

Transcript

Anthropic’s Secret Mythos Model Leaks, Altman vs Amodei’s Decade-Long Feud, Anthropic Wins in Court · Today’s AI News

• 0:00
  Welcome back to today's AI News, brought to you by 9X
• 0:02
  Productions. So imagine a master illusionist,
• 0:06
  right? And they just accidentally
• 0:08
  dropped their notebook of ultimate secrets right outside
• 0:11
  of Rivals Theater. Oh, just a completely accidental
• 0:14
  drop, I'm sure. Right, totally an accident.
• 0:16
  Well, that's exactly what AI Lab Anthropic did this week with
• 0:19
  their top secret Mythos model. And I mean the timing couldn't
• 0:23
  be more. Perfect.
• 0:24
  It really couldn't. So today we are going on a
• 0:27
  massive deep dive into this entire situation.
• 0:30
  We are exploring the mechanics of Anthropics data leak and
• 0:34
  really looking at how the bitter like decade long personal feud
• 0:39
  between the leaders of Anthropic and Open AI is secretly
• 0:42
  dictating the pace of this whole industry.
• 0:44
  Which is a huge factor that a lot of people overlook.
• 0:46
  Exactly. And from there, we'll track how
• 0:48
  that massive frontier technology is trickling down to you,
• 0:51
  looking at practical ways everyday people are automating
• 0:53
  their lives right now. Then we'll run through a rapid
• 0:55
  fire update on the rest of the AI landscape and close out with
• 0:58
  a truly touching community workflow.
• 1:01
  Let's just jump right into this anthropic situation.
• 1:03
  Yeah. So we are looking at a major
• 1:05
  configuration error here. Anthropics content management
• 1:08
  system. Their CMS basically had a
• 1:10
  staging environment that was left completely exposed.
• 1:13
  Which is just wild for a company of that scale.
• 1:15
  Right. It resulted in thousands of
• 1:18
  unpublished internal assets just sitting there in a publicly
• 1:21
  accessible data cache. And right in the middle of that
• 1:24
  cache was a draft blog post detailing a new model called
• 1:27
  Claude Mythos. And the leak specifically places
• 1:30
  Mythos in a brand new tier. They are calling it the Cappy
• 1:33
  Bar tier which first of all, great name.
• 1:36
  A very anthropic name, yeah. Very much so.
• 1:39
  But second, it completely disrupts the current hierarchy
• 1:42
  of frontier models. Capybara sits entirely above
• 1:45
  their current top tier Opus class.
• 1:47
  And Opus is already a massive, highly capable model.
• 1:51
  Exactly. To create a tier above it means
• 1:53
  we are talking about a significant leap in parameter
• 1:56
  count and training compute. It's larger, and the internal
• 1:59
  documents note it's significantly more expensive to
• 2:02
  run. But the internal notes are where
• 2:04
  the real story is. The leak claims that Mythos is,
• 2:07
  quote, currently far ahead of any other AI model in cyber
• 2:10
  capabilities. Yeah, and Anthropics own
• 2:12
  developers specifically warned that this model could help
• 2:15
  hackers outpace defenders. Fortune actually confirmed
• 2:19
  independently that they are testing this new model,
• 2:22
  highlighting major advances in reasoning, coding and
• 2:24
  specifically cybersecurity. OK, but I have to ask, wait, is
• 2:27
  it really an accident when a safety focused lab like
• 2:31
  Anthropic leaves his most powerful paradigm shifting
• 2:34
  launch materials in a public data store?
• 2:36
  That is the big question, because we.
• 2:38
  Saw this exact same playbook with Open a eyes incredibly
• 2:42
  conveniently timed Q star and strawberry leaks.
• 2:45
  It just feels like the new standard playbook for generating
• 2:48
  free hype, you know? The timing definitely resets the
• 2:51
  industry narrative. I mean, just as open AI is
• 2:54
  dominating the news cycle, Entropic suddenly reminds
• 2:56
  everyone they have this absolute behemoth waiting in the wings.
• 2:59
  Very. But looking past the marketing
• 3:01
  mechanics, that specific focus on cyber capabilities is a
• 3:05
  massive escalation. A model that is far ahead in
• 3:08
  cybersecurity doesn't just act as a firewall.
• 3:11
  To truly understand how to defend a network, a system has
• 3:14
  to fundamentally understand how to break into it.
• 3:17
  Because it's a dual use technology if it can like audit
• 3:21
  millions of lines of code to find a vulnerability.
• 3:23
  It knows exactly where the weak points are precisely right.
• 3:26
  It spots zero day exploits by understanding the logical flow
• 3:30
  of the code rather than just looking for known malware
• 3:32
  signatures. If you release a Capybara tier
• 3:35
  model that can execute those exploits at superhuman speed,
• 3:39
  you are pushing the entire digital infrastructure into
• 3:42
  completely uncharted territory. The double edged sword.
• 3:45
  Exactly. You're handing over a tool that
• 3:48
  fundamentally shifts the balance of power between offensive
• 3:51
  hackers and defensive cybersecurity teams.
• 3:54
  But, you know, these frontier models don't just build
• 3:55
  themselves. The sheer velocity of these
• 3:58
  releases and the willingness to push into that precarious
• 4:01
  territory is really being driven by a very intense, deeply
• 4:05
  personal rivalry operating behind the scenes.
• 4:07
  Yeah, the human architecture behind the neural networks.
• 4:10
  Perfectly said The Wall Street Journal just published a massive
• 4:13
  report tracing this power struggle all the way back to a
• 4:16
  specific San Francisco group house in 2016.
• 4:19
  The lore goes deep. It really does.
• 4:21
  The animosity between the leaders of Anthropic and Open AI
• 4:25
  is historic. Daria Almode was at Open AI from
• 4:28
  2016 to 2020, and Daniella Amode was there from 2018 to 2020.
• 4:33
  The Journal details major early friction, particularly with Open
• 4:37
  AI Co founder Greg Brockman. Right, There were some
• 4:39
  fundamental disagreements early on.
• 4:41
  Yeah, at one point, Brockman allegedly floated the idea of
• 4:44
  selling artificial general intelligence directly to UN
• 4:47
  Security Council nuclear powers. And Dario Amodei reportedly
• 4:50
  considered that proposal, quote, tantamount to treason.
• 4:54
  I mean, that is an incredibly severe characterization of a
• 4:56
  strategic disagreement. It really highlights how early
• 4:59
  the ideological fracturing began.
• 5:01
  And the fracturing only accelerated from there.
• 5:04
  The WSJ report notes that Open AI CEO Sam Allman accused the
• 5:07
  Amedes of plotting against him to the board in a private
• 5:10
  meeting. But then, when confronted about
• 5:12
  that accusation, Altman allegedly denied making it.
• 5:15
  Which just adds to the drama. The report also brought to light
• 5:18
  some highly intense private comparisons made by Amade.
• 5:21
  He reportedly likened the ongoing lawsuit between Sam
• 5:24
  Altman and Elon Musk to Hitler versus Stalin.
• 5:27
  That is intense. Extremely.
• 5:29
  He also referred to open AI as Big tobacco and he labeled a
• 5:33
  political donation that Greg Brockman made to a pro Trump
• 5:36
  pack as evil. Now before we go further, for
• 5:40
  you listening, we have a strict impartiality rule here.
• 5:43
  The sources contain some highly political rhetoric.
• 5:46
  UN nuclear powers pro Trump packs historical dictator.
• 5:50
  Right, we are just the messengers here.
• 5:51
  Exactly. We are simply reporting the
• 5:54
  documented historical grudges and rhetoric from the WSG source
• 5:57
  material. We are not taking any political
• 6:00
  sides or endorsing any of these viewpoints whatsoever.
• 6:02
  We just want to give you the full picture of what's driving
• 6:05
  this arms race. And when you look at those
• 6:07
  documented grudges, the stark differences between how open AI
• 6:10
  and anthropic operate actually make complete sense.
• 6:13
  This isn't just a polite disagreement over how to wait an
• 6:16
  oral network. Right.
• 6:17
  Or how to structure a safety team.
• 6:18
  Exactly. It is a fundamental breakdown of
• 6:21
  trust when the architects of these systems view their
• 6:24
  competitors through the lens of extreme historical and moral
• 6:27
  comparisons, like Big Tobacco and global dictators.
• 6:31
  Every model update, every leaked blog post, it becomes a battle
• 6:35
  in a much larger ideological war.
• 6:37
  But I have to ask you, does this intensely personal bad blood
• 6:42
  actually benefit the technology? I mean, you have these brilliant
• 6:45
  engineers who just want to out build the people they shared a
• 6:47
  house with 10 years ago. Does that fierce competition
• 6:50
  accelerate AI innovation? Or does this level of animosity
• 6:53
  make these companies dangerously reckless with things like
• 6:55
  cybersecurity models? Well, it creates the ultimate
• 6:58
  paradox of the current AI boom, really.
• 7:01
  Anger and rivalry are undeniably incredible motivators.
• 7:04
  That raw desire to prove a former colleague wrong is
• 7:06
  exactly why we are seeing these compressed timelines.
• 7:09
  You know, leaping from Opus to Capybara in record time.
• 7:12
  So it does push things faster. It does, but the shadow side of
• 7:16
  that acceleration is the erosion of caution.
• 7:19
  When your primary goal becomes beating a competitor that you
• 7:22
  view as morally compromised, well, standard safety protocols
• 7:27
  can suddenly look like unnecessary red tape.
• 7:29
  The feud fuels the engine, but it removes the guardrails.
• 7:33
  That makes a lot of sense, but you know, while the billionaires
• 7:36
  are locked in this ideological arms race over the future of
• 7:39
  AGI, regular users are taking the models that are currently
• 7:42
  available and completely rewiring their daily lives.
• 7:45
  Yeah, the practical power of AI on the ground right now is
• 7:48
  staggering. We are seeing a massive shift
• 7:51
  from AI as just a novelty to AI as foundational infrastructure
• 7:55
  for everyday tasks. Definitely, and we have two
• 7:57
  incredible stories from the community that illustrate this
• 8:00
  perfectly. The 1st is from Nate, a
• 8:02
  university educator. He was installing acoustic
• 8:04
  panels on his ceiling. OK, a classic DIY project.
• 8:07
  Right, and usually you take a measuring tape, guess the square
• 8:10
  footage, buy a few boxes, and end up with terrible cuts and a
• 8:14
  bunch of wasted material. But Nate didn't do that.
• 8:17
  He used clawed artifacts to create a digital mock up of his
• 8:20
  specific room and then had the AI layout the panels in various
• 8:23
  orientations. See, that's brilliant because
• 8:25
  the AI was engaging in complex spatial reasoning there.
• 8:28
  It wasn't just drawing a grid, it was optimizing the geometry
• 8:31
  of the to actively reduce the number of physical cuts Nate
• 8:34
  would have to make. Which saves time and money.
• 8:37
  Exactly that kind of calculation minimizes material waste,
• 8:41
  ensures accurate purchasing, and provides precise cost
• 8:45
  estimations across different layout patterns.
• 8:47
  It took a manual error prone home improvement task and turned
• 8:51
  it into an optimized populated simulation.
• 8:54
  It's amazing. The second story is from
• 8:56
  Shubham, an editor who needed a portfolio website and instead of
• 9:00
  using a standard template builder, he had clawed build the
• 9:03
  entire site from scratch in one single sitting.
• 9:06
  Just one prompt session. Pretty much he provided his
• 9:09
  LinkedIn and social links, describes his aesthetic vision,
• 9:13
  and the AI handled the design, deployment, DNS configuration
• 9:16
  and even SEO indexing. 0 Coding required on Shubham's part.
• 9:21
  That is wild. But the truly wild part was the
• 9:24
  debugging process. When the deployment inevitably
• 9:27
  hit a snag, the AI used a desktop extension browser
• 9:31
  control to fix the website itself.
• 9:33
  OK, that mechanism is a massive leap forward.
• 9:36
  The AI wasn't just outputting a text block saying you know,
• 9:39
  here's a code to fix your DNS error, it was actively driving
• 9:43
  the browser. Like a remote IT guy.
• 9:45
  Yes, it manipulated the document object model.
• 9:47
  Read the console errors and real time rewrote the broken function
• 9:51
  and refresh the page. The AI acted as both the
• 9:54
  developer and the IT support system, interacting directly
• 9:57
  with the software. Environment Which brings us to
• 9:59
  how you listening right now can upgrade how you interact with
• 10:02
  these tools. Chatting with an AI in a web
• 10:04
  browser is fine, but building reusable skills using the codecs
• 10:07
  desktop application is really where the real leverage is.
• 10:11
  Yeah, Codecs Desktop represents A vital architectural shift.
• 10:14
  It is a local application that bridges your operating system
• 10:18
  and the large language model, allowing you to create a global
• 10:21
  macro layer for your intellectual work.
• 10:23
  And the setup is incredibly straightforward.
• 10:26
  You just install codecs on your PC, hit command or control plus
• 10:30
  O to add a new project and designate a folder as your
• 10:33
  skills playground. From there you just ask codecs
• 10:36
  to create a specific skill. The source material goes a
• 10:39
  really great example of this, actually asking the AI to build
• 10:42
  a skill that uses the AARR framework to evaluate startup
• 10:46
  ideas. Right.
• 10:47
  And for context, AAR stands for acquisition, activation,
• 10:50
  retention and Referral, the standard business metric
• 10:53
  framework. Exactly, and instead of
• 10:55
  explaining that framework to ChatGPT every single time you
• 10:58
  want to analyze a business, you build it into codecs once.
• 11:01
  Yes, if the new skill doesn't populate immediately, you just
• 11:04
  quit and reopen the app. After that, you type a forward
• 11:08
  slash, select your startup evaluation skill and run it.
• 11:12
  You can even use the slash skill creator command to ensure codecs
• 11:16
  builds these repeatable workflows globally across your
• 11:19
  entire system, or you can lock them to specific project
• 11:22
  folders. It's so efficient.
• 11:24
  It really feels exactly like building old school excel
• 11:26
  macros, but instead of just sorting numbers you are
• 11:30
  automating cognitive reasoning. Well, if.
• 11:31
  We look at the history of computing.
• 11:33
  Every major leap in productivity comes from abstraction, right?
• 11:37
  When you simply chat with a web-based AI, you are forced to
• 11:40
  start from a blank page every single session.
• 11:43
  You have to re establish the context, the formatting rules
• 11:46
  and the ultimate goal. It is highly inefficient.
• 11:49
  By utilizing codecs to build a work flow, a skill, you
• 11:52
  construct a custom cognitive machine.
• 11:55
  You build the logic engine once, and from then on you only need
• 11:57
  to feed it raw materials. It transitions AI from being a
• 12:01
  conversational partner to acting as an automated assembly line
• 12:03
  for your specific intellectual tasks.
• 12:05
  An automated assembly line for your mind.
• 12:08
  That is exactly what it is. Let's shift gears now and run
• 12:11
  through our quick heads. We're looking at a rapid fire
• 12:14
  overview of the broader AI ecosystem, tracking how this
• 12:17
  tech is rippling through software, HARMA, and global
• 12:21
  hardware. What's happening this week?
• 12:22
  Oh, definitely tarting with some trending tools.
• 12:24
  We have Hoto Tudio, which is a new ersonalized HOTO generation
• 12:29
  model. Then there is Tribe V2, which is
• 12:32
  an AI designed to actually simulate brain responses to
• 12:35
  sights, sounds and language. And finally, Cohere transcribe a
• 12:39
  state-of-the-art open source speech recognition model.
• 12:42
  I have to say, the mechanism behind Tri V2 is fascinating.
• 12:45
  Oh yeah, yeah, Standal generative AI predicts the next
• 12:48
  token in a sequence based on vast amounts of text data.
• 12:50
  Right? But Tri V2 operates differently.
• 12:53
  It maps neural activity data. Think F MRI scans against
• 12:56
  audiovisual inputs to predict how a human biological cognitive
• 12:59
  system will respond onto specific stimuli.
• 13:01
  Wow. Yeah, we are moving from AI that
• 13:03
  simply mimics human output to AI that actually models human
• 13:06
  biological processing. That is mind blowing.
• 13:09
  Moving to corporate news that Open AI and Anthropic rivalry
• 13:14
  continues to surface in unexpected ways.
• 13:17
  According to internal Slack messages seen by Axios, Sam
• 13:21
  Altman reportedly told Open AI staff that he actively tried to
• 13:25
  save Anthropic during its standoff with the Pentagon over
• 13:28
  defense contracts. That is quite a claim.
• 13:31
  Right. And he apparently claimed this
• 13:32
  while Open AI was simultaneously locking in its own Pentagon
• 13:35
  deals. It just highlights the
• 13:37
  incredibly complex layers of strategy and posturing happening
• 13:40
  at the executive level right now.
• 13:42
  Also in personnel news, Ross Nordine has officially departed
• 13:45
  X AI. Aside from Elon Musk, Gordine
• 13:48
  was the very last remaining original Co founder at the
• 13:50
  startup. Losing the entire founding
• 13:52
  technical team represents a massive shift in the
• 13:54
  foundational DNA of that company.
• 13:56
  It will absolutely impact their development.
• 13:58
  Development trajectory moving. Forward.
• 14:00
  Definitely. Now.
• 14:00
  This next story is arguably the most significant of a week.
• 14:03
  Pharmaceutical giant Eli Lilly just entered a massive $2.75
• 14:07
  billion deal with Hong Kong based in silico medicine.
• 14:10
  Huge numbers. Massive.
• 14:12
  They are licensing in Silico's AI discovered drug pipeline,
• 14:16
  which currently boasts 28 different compounds in
• 14:19
  development. This is AI moving out of text
• 14:22
  generation and into multibillion dollar hard science.
• 14:26
  I mean, how big of a deal is this really?
• 14:28
  To understand and the magnitude of this deal, we have to look at
• 14:30
  how drug discovery actually works.
• 14:33
  Chemical space. The number of possible small
• 14:35
  molecules is estimated to be around 10 to the 60th power.
• 14:39
  OK, so an unfathomably large number.
• 14:42
  Exactly. Traditional pharma relies on
• 14:44
  high throughput screening where physical robots test thousands
• 14:48
  of actual chemical compounds in a physical lab to see if they
• 14:51
  bind to a disease target. It takes months, cost millions,
• 14:54
  and it is largely trial and error.
• 14:56
  You are physically searching for a needle in a haystack the size
• 14:59
  of the universe. That's a great way to put it.
• 15:01
  AI bypasses the physical bottleneck entirely.
• 15:04
  An AI model can handle the combinatorial explosion of
• 15:07
  molecular interactions by simulating the physics and
• 15:09
  binding affinities digitally. It can screen millions of
• 15:12
  molecular structures in hours, predicting which key will
• 15:16
  perfectly fit the biological lock of a disease.
• 15:18
  Unbelievable. Eli Lilly paying $2.75 billion
• 15:22
  for 28 of these digitally discovered compounds is proof
• 15:26
  that Big Pharma no longer views AI as just a neat research
• 15:29
  assistant. They view it as the primary
• 15:32
  engine for the next century of medical breakthroughs.
• 15:35
  It really is a complete paradigm shift for global health.
• 15:38
  On the legal front, Anthropic just one of a federal injunction
• 15:42
  blocking the Trump administration's supply chain
• 15:44
  risk designation against the company.
• 15:47
  To quote the judge presiding over the case directly, that
• 15:49
  designation was called Classic Illegal First Amendment
• 15:52
  Retaliation. And again, remaining strictly
• 15:55
  impartial on the politics here, legally speaking, that
• 15:58
  establishes a significant precedent regarding how frontier
• 16:02
  AI companies interact with federal regulatory designations
• 16:05
  and supply chain restrictions. Exactly.
• 16:07
  Rounding out the rapid fire updates, Google has officially
• 16:10
  expanded the roll out of its Live Translate feature to iOS.
• 16:13
  The software update basically turns any pair of headphones
• 16:16
  connected to an iPhone into a real time 70 plus language
• 16:19
  interpreter. The sci-fi Universal Translator
• 16:21
  is now just a standard smartphone feature.
• 16:23
  It really is. OK, let's bring this deep dive
• 16:26
  back down to the human element. We've talked about billion
• 16:29
  dollar pharma deals and global cybersecurity risks, but AI's
• 16:33
  impact is often most profound when it is deeply personal.
• 16:37
  Today's community workflow comes from a reader named David B out
• 16:40
  in Wichita, KS. This is such a beautiful
• 16:43
  application of the technology. It really is so.
• 16:46
  David's father passed away this past February, leaving behind a
• 16:49
  large cardboard box filled with handwritten letters from the
• 16:52
  1950s and 60s. Normally, the friction of
• 16:54
  digitizing something like that is just too high.
• 16:57
  The box goes into a closet and the history slowly fades.
• 17:00
  Right transcription takes forever.
• 17:01
  But David photographed all the letters and fed the bulk images
• 17:04
  directly to clawed code. The AI was able to perfectly
• 17:08
  read and transcribe the 70 year old fading cursive.
• 17:11
  It then actually helped David write the code to build a
• 17:14
  dedicated family history website, pairing the original
• 17:17
  photographs of the letters side by side with the clean,
• 17:19
  searchable text transcriptions. The real power here is the
• 17:23
  transformation of state. David took unstructured hidden
• 17:27
  analog data and remove the friction of preservation.
• 17:31
  By utilizing the AI to handle the heavy lifting of
• 17:34
  transcription and web development, he created an
• 17:36
  interactive archive. He took isolated physical
• 17:40
  memories that were trapped in a box and made them instantly
• 17:42
  accessible across generations. It perfectly encapsulates the
• 17:46
  ultimate promise of this technology, taking the
• 17:49
  unstructured data of our lives and making it meaningful.
• 17:52
  It turned a forgotten box into a living digital legacy for his
• 17:56
  grandkids. Which leads me to a thought I
• 18:00
  want to leave you with today. Thinking about David's family
• 18:02
  archive. No, no, just thinking about it.
• 18:05
  As AI becomes our ultimate archivist, how will future
• 18:08
  generations interact with our personal histories?
• 18:10
  Will they just passively read the text of our emails and
• 18:13
  letters? Or will they be able to use
• 18:15
  these models to truly converse with the memories and the data
• 18:17
  we leave behind? Definitely something to chew on.
• 18:20
  Thank you so much for joining us on this deep dive.
• 18:23
  If you want to stay updated on everything we uncover, make sure
• 18:25
  you hit that subscribe button. And if you enjoyed the show,
• 18:28
  please take a second to rate US five stars.
• 18:30
  It really helps us out. We will see you back here for
• 18:33
  tomorrow's show.