Latest / AI Fire Daily / 🎙️ EP 341: Autonomous Russian Drone Attack & Thomson Reuters Launches Legal Model

Transcript

🎙️ EP 341: Autonomous Russian Drone Attack & Thomson Reuters Launches Legal Model · AI Fire Daily

• 0:00
  There is this quiet threshold in modern technology.
• 0:03
  It is the exact moment a machine stops waiting.
• 0:08
  It stops waiting for our human instructions.
• 0:10
  Right. It just starts deciding. Exactly. Instead
• 0:12
  of waiting, it just makes its own decisions.
• 0:15
  We are crossing that invisible line right now.
• 0:17
  We really are. Welcome to today's deep dive into
• 0:20
  the source material. We are looking at AI's massive
• 0:24
  leap into full autonomy. It is a pretty profound
• 0:27
  transition for all of us. Yeah. It is. We will
• 0:30
  start with a chilling physical reality today.
• 0:32
  We are tracking autonomous drones deployed in
• 0:35
  Ukraine. Which is just a fascinating engineering
• 0:38
  shift. Then we are going to see how cheap computing
• 0:40
  scales globally. We will look at highly specialized
• 0:43
  new agent tools. And finally, we'll examine a
• 0:46
  legacy corporate giant. Right, Thomson Reuters.
• 0:48
  Yes, they just built a $40 million private AI
• 0:52
  brain. This is a massive shift for everyday technology
• 0:55
  users like you. We are moving past AI as a simple
• 0:58
  flight assistant. We are truly entering an era
• 1:00
  of autonomous decision makers. Let us start with
• 1:03
  the first section of our sources. We are looking
• 1:06
  at a July 6th New York Times report. And to be
• 1:10
  clear, right up front, we are strictly looking
• 1:12
  at the technology here. We maintain complete
• 1:15
  neutrality regarding the geopolitical conflict.
• 1:17
  Exactly right. We are just examining the raw
• 1:20
  technological facts reported. We are not taking
• 1:22
  sides. The report details a very specific event.
• 1:26
  in zaparizhia ukraine yeah a russian ai guided
• 1:29
  drone reportedly killed three civilians there
• 1:33
  and the physical hardware inside this drone is
• 1:36
  what is truly shocking it really is investigators
• 1:39
  found something very specific inside the wreckage
• 1:42
  the drone had absolutely no antenna for remote
• 1:45
  control right and that is a massive engineering
• 1:47
  shift it is huge no antenna means there is no
• 1:49
  remote pilot steering it it means the drone cannot
• 1:52
  receive external commands at all so what was
• 1:54
  driving it Well, inside they found an NVIDIA
• 1:57
  Jetson Orin mini computer. And that specific
• 2:00
  component is a widely available commercial product.
• 2:02
  Yeah, you can buy one for just a few hundred
• 2:04
  dollars. You know, it is often used by students
• 2:07
  for basic robotics projects. Exactly. It is off
• 2:10
  -the -shelf consumer tech. This normalizes something
• 2:13
  that is, frankly, quite terrifying. It is literally
• 2:17
  like giving a calculator a loaded weapon. We
• 2:20
  are putting lethal autonomy on cheap consumer
• 2:23
  hardware. And the software side is equally complex
• 2:26
  and troubling because it lacks an antenna. It
• 2:29
  processes everything locally. Right. The drone
• 2:32
  uses its own cameras to view the ground. It compares
• 2:35
  that live video to stored terrain images. And
• 2:39
  the AI uses those images to navigate entirely
• 2:42
  independently. Earlier AI models only handled
• 2:45
  the very final approach flight. Yeah. A human
• 2:47
  operator always chose the actual target first.
• 2:51
  But Ukrainian experts say this new drone is fundamentally
• 2:53
  different. It really is. It can actually choose
• 2:55
  its final target by itself. It does not need
• 2:57
  a human to click on a screen. No, it can identify
• 3:00
  and strike specific fuel tanks autonomously.
• 3:02
  It recognizes the visual signature of the target
• 3:05
  and acts. And both Russia and Ukraine are actively
• 3:07
  testing these systems. The New York Times report
• 3:10
  highlights a critical evolution here. AI is no
• 3:13
  longer just helping these drones fly straight.
• 3:16
  It is potentially letting drones decide what
• 3:18
  to actually attack. The decision -making loop
• 3:20
  is entirely closed off from humans. Two -sec
• 3:22
  silence. It removes the human completely from
• 3:25
  the final decision. So I have to ask you this.
• 3:28
  What happens to global security when advanced
• 3:32
  autonomous targeting is available on hardware
• 3:34
  that costs less than a smartphone? Honestly,
• 3:37
  it completely destabilizes traditional military
• 3:40
  defense strategies worldwide. Cheap autonomous
• 3:43
  weapons mean any small group can scale attacks
• 3:45
  indefinitely. Cheap tech completely redefines
• 3:49
  global asymmetric warfare forever. But this democratization
• 3:52
  of cheap power is not staying on the battlefield.
• 3:55
  No, it definitely is not. It is bleeding directly
• 3:57
  into our digital infrastructure globally. We
• 3:59
  are seeing it scale rapidly in cyberspace. It
• 4:02
  is accelerating in robotics and advanced hardware,
• 4:04
  too. Our sources show it is even moving into
• 4:06
  orbit. Yeah, the scaling is happening faster
• 4:08
  than most people realize. Just look at what is
• 4:10
  happening in the cybersecurity world. Chinese
• 4:13
  hackers are actively using deep seek language
• 4:15
  models. They're using this cheap AI to scale
• 4:18
  real cyber operations. Which proves a very unsettling
• 4:22
  point about AI models generally. Attackers do
• 4:24
  not need the smartest model to cause major damage.
• 4:28
  Right. They do not need the most expensive reasoning
• 4:29
  engine. They just need a model that acts quickly
• 4:31
  and cheaply. Yeah. Sheer volume and speed often
• 4:34
  beat pure intelligence. Hackers can automate
• 4:37
  massive spear phishing campaigns almost instantly
• 4:40
  now. And the hardware. Supporting this speed
• 4:42
  is upgrading rapidly. OpenAI just shared results
• 4:45
  for their new jalapeno chip. Which is an inference
• 4:48
  chip. Right. Hardware built specifically to run
• 4:51
  trained AI models efficiently. OpenAI claims
• 4:54
  it offers 4 .1 times better performance. That
• 4:57
  is specifically for highly interactive, complex
• 5:00
  workloads. It processes vast amounts of data
• 5:02
  with incredible efficiency. And this computing
• 5:05
  power is literally leaving the planet now. This
• 5:07
  is the craziest part to me. SpaceX AI plans to
• 5:10
  launch orbital AI compute systems soon. They
• 5:13
  are aiming for the fourth quarter of 2027. They
• 5:17
  will use NVIDIA Vera CPUs aboard their StarMine
• 5:20
  satellites. Whoa! Imagine scaling to a billion
• 5:23
  queries straight from space. It is wild. Data
• 5:27
  processed in orbit could bypass terrestrial network
• 5:29
  latency entirely. You skip the miles of underwater
• 5:32
  fiber optic cables. Exactly. You just beam the
• 5:35
  answers straight down. Meanwhile, corporate investments
• 5:38
  in physical AI are exploding downward, too. Back
• 5:41
  down here on Earth. General intuition just raised
• 5:44
  money at a $6 billion valuation. That comes just
• 5:48
  weeks after a massive previous funding round.
• 5:51
  They are betting big on AI for real physical
• 5:53
  robotics. They want AI that interacts with the
• 5:56
  messy physical world. Meta is also making massive
• 5:59
  moves in this space currently. Yeah, they are
• 6:01
  planning to launch their Hatch AI agent very
• 6:03
  soon. It will feature deep web service integrations
• 6:05
  and premium paid plans. They also have a mysterious
• 6:08
  major model internally called Watermelon. That
• 6:11
  new model is scheduled to arrive this coming
• 6:13
  October. But it is not just giant corporate robotics
• 6:16
  and massive models. No, everyday daily use of
• 6:19
  AI is changing how we actually work. A new clawed
• 6:22
  skill called ELI -5 is going viral internally.
• 6:26
  It explains complex topics using big visuals
• 6:29
  and very few words. Users are getting incredibly
• 6:32
  sophisticated with these everyday tools. One
• 6:34
  Redditor used Codex to create a completely new
• 6:37
  video. They turned the classic disaster girl
• 6:40
  meme into a 20 -second clip. They used a surprisingly
• 6:43
  detailed five -step prompt workflow to do it.
• 6:46
  It is impressive how people are chaining these
• 6:48
  commands. Honestly? I still wrestle with prompt
• 6:51
  drift myself. Oh, sure. We all do. Seeing someone
• 6:54
  execute a perfect five -step workflow is pretty
• 6:56
  humbling. Well, the tools are definitely getting
• 6:58
  better at managing user context. Claude Cowork
• 7:01
  can now actually remember what you told past
• 7:03
  chats. You do not have to brief the same project
• 7:05
  twice anymore. You can just view, edit, or delete
• 7:09
  those specific memories. It makes the whole interaction
• 7:11
  feel much more fluid. Two sec silence. If hackers
• 7:15
  are using cheap, accessible models to scale attacks,
• 7:18
  is our cyber defense adapting fast enough? Honestly,
• 7:21
  traditional defense is struggling to keep pace
• 7:23
  today. Defenders have to adopt AI just to filter
• 7:26
  through the noise. They need AI to spot the AI
• 7:29
  -generated attacks. Defense most move from human
• 7:31
  reaction to automated prediction. Sponsor mid
• 7:34
  -roll placeholder text. Welcome back. We have
• 7:37
  seen how chaotic AI scaling is getting globally.
• 7:40
  The sheer volume of automated actions is honestly
• 7:43
  overwhelming. Now let us look at the highly specialized
• 7:46
  tools emerging. Right, the tools built specifically
• 7:49
  to harness and control this data. The ecosystem
• 7:52
  of empowered AI tools is growing incredibly fast.
• 7:56
  We are moving beyond simple chatbots into complex
• 7:58
  multi -tool agents. WAN 3 .0 is a perfect example
• 8:02
  of this shift. It turns almost any basic input
• 8:05
  into an AI video clip. You can use text, videos,
• 8:08
  documents, or even boring spreadsheets. And it
• 8:11
  generates a highly polished 30 -second video
• 8:13
  almost instantly. That completely changes how
• 8:16
  we consume basic information daily. Why read
• 8:18
  a spreadsheet when AI can animate the data trends?
• 8:21
  Then we have massive research tools like Keynable
• 8:24
  entering the market. Keenable gives AI agents
• 8:27
  incredibly fast web search capabilities natively.
• 8:30
  It scans across 100 billion different documents
• 8:33
  seamlessly. It also includes a fascinating feature
• 8:36
  called Time Machine. This lets your AI explore
• 8:39
  how the web looked in the past. It pulls historical
• 8:42
  context directly into its current reasoning process.
• 8:45
  Another fascinating tool in our source material
• 8:47
  is Actidot Pro. It gives AI agents direct access
• 8:51
  to private company data safely. It offers four
• 8:54
  times the normal research depth for enterprise
• 8:57
  users. And it provides over 100 signals for much
• 9:01
  smarter business outreach. It is like stacking
• 9:03
  Lego blocks of data perfectly. Yeah, exactly.
• 9:06
  You build a completely custom knowledge base
• 9:07
  for your agent. It builds a solid foundation
• 9:10
  for enterprise AI workflows. But managing all
• 9:12
  these advanced tools requires immense processing
• 9:15
  power. That brings us to a new optimization tool
• 9:17
  called DietCloud. DietCloud tracks your Quad
• 9:20
  API usage in real time. It monitors exactly how
• 9:24
  much computing power you are burning. It optimizes
• 9:26
  your token used to save you significant money.
• 9:29
  But here is the most interesting part of its
• 9:31
  workflow. When you hit your usage limits, it
• 9:34
  moves your conversation automatically. It seamlessly
• 9:36
  shifts you to another language model completely.
• 9:39
  You do not even notice the underlying model changed.
• 9:41
  It is a brilliant way to manage constant workflow
• 9:44
  interruptions. And never actually stop working
• 9:46
  when you hit a strict wall. The tool just reroutes
• 9:49
  your thoughts to a different synthetic brain.
• 9:51
  Beat. But this raises a very serious question
• 9:54
  for us. Are we actively managing these AI tools
• 9:59
  in our daily lives? Or with tools like Diet Clod
• 10:03
  moving our conversations around autonomously,
• 10:05
  are they starting to manage us? It is a blurry
• 10:08
  line that keeps getting blurrier every day. We
• 10:10
  are outsourcing our basic workflow logistics
• 10:12
  to invisible algorithms. The AI decides which
• 10:16
  AI is best suited for your prompt. Do these ultra
• 10:20
  -specific tools risk fragmenting our workflow?
• 10:23
  Or are they the only way to get real utility
• 10:25
  out of AI? Without specialized tools, large models
• 10:28
  are really just parlor tricks. These specific
• 10:30
  tools actually ground the AI in measurable real
• 10:33
  -world productivity. Hyper -specific tools are
• 10:36
  necessary bridges to broader AI systems. Yeah,
• 10:39
  they really are. This brings us to the final
• 10:41
  section of our sources. We are seeing a massive
• 10:43
  divergence in how companies use AI. Instead of
• 10:46
  stitching together external tools like Diet Clod
• 10:49
  or Keenable, some are pivoting. Legacy giants
• 10:52
  are choosing to build their own proprietary AI
• 10:55
  brains. They are building everything entirely
• 10:57
  in -house. Which is a massive strategic shift
• 11:01
  in the enterprise landscape. Thomson Reuters
• 11:03
  just built their very first proprietary AI model.
• 11:06
  They are officially calling this new proprietary
• 11:08
  model Thomson One. They did not build this entirely
• 11:11
  from scratch, though. No, they adapted an Alibaba
• 11:15
  -Quen -based foundation model first. They took
• 11:17
  that base... and injected their own proprietary
• 11:20
  legal data. They used decades of their own legal
• 11:23
  and professional content. The economics of this
• 11:25
  corporate build are really fascinating here.
• 11:26
  It took about two years of dedicated internal
• 11:29
  development time. It reportedly cost around $40
• 11:32
  million in staff and computing. That is a massive
• 11:35
  upfront investment for a legacy company. But
• 11:38
  the cost of their ongoing training is dropping
• 11:41
  dramatically. Their absolute latest training
• 11:43
  run only cost roughly $450 ,000. That changes
• 11:47
  the long -term financial math completely. Right.
• 11:49
  And that changes the math for the future entirely.
• 11:52
  Thomson Reuters claims internal tests show absolutely
• 11:56
  incredible performance results. They say Thomson
• 11:59
  One sits ahead of Claude, Gemini, and GPT. That
• 12:03
  is specifically for highly complex legal task
• 12:06
  evaluations. It is a bold claim. I have got to
• 12:09
  push back on those claims a bit, though. Fair
• 12:11
  enough. We really have to question those internal
• 12:13
  test results closely. Thomson Reuters is basically
• 12:16
  grading their own homework here. Unverified corporate
• 12:19
  testing. always carries an inherent bias naturally.
• 12:22
  Every single company claims their proprietary
• 12:25
  tool wins internal benchmarks. That is a very
• 12:28
  fair point to keep in mind. The results have
• 12:30
  not been independently verified by outside researchers
• 12:32
  yet. But what is truly wild is how little data
• 12:35
  they actually used. The model trained on less
• 12:38
  than 10 % of their library. That means they have
• 12:40
  massive room for future improvement ahead. The
• 12:43
  company also plans to release an open weights
• 12:45
  version soon. A model where core settings are
• 12:48
  public for outside tweaking. Exactly. CTO Joel
• 12:52
  Hohn explained the strategy behind this move
• 12:54
  clearly. The company wants to build long -term
• 12:56
  value directly from its own data. They are still
• 13:00
  working with external providers like Anthropic
• 13:02
  currently. They have integrations between Claude
• 13:05
  and Co -Counsel running very smoothly. But building
• 13:07
  specialized models on strong open models is the
• 13:10
  future. Constant internal training is becoming
• 13:13
  much cheaper. than heavy API usage. Two sec silence.
• 13:17
  Is spending $40 million to build a specialized
• 13:20
  model a luxury or survival requirement for data
• 13:23
  -rich companies? It is absolutely survival if
• 13:26
  your core product is proprietary knowledge. Renting
• 13:29
  intelligence indefinitely from open AI will eventually
• 13:32
  erode your profit margins. Owning your data pipeline
• 13:34
  is a non -negotiable survival tactic. Right.
• 13:37
  This brings us to the ultimate synthesis of our
• 13:39
  journey today. We are watching AI transition
• 13:42
  globally right before our eyes. We are watching
• 13:45
  it move from an API that assists us politely.
• 13:47
  It is becoming an autonomous agent that acts
• 13:50
  entirely for us. We see this everywhere across
• 13:53
  the technological spectrum today. We see $300
• 13:56
  chips guiding lethal drones autonomously. We
• 13:59
  see cheap deep -seek models scaling hacker operations
• 14:02
  globally. And we see $40 million models replacing
• 14:05
  generalist AI systems entirely. The technology
• 14:09
  is no longer waiting for our final click. No,
• 14:12
  it really is not. The line between human instruction
• 14:14
  and machine action is practically vanishing.
• 14:18
  It is a totally new paradigm. I want you to think
• 14:20
  about a lethal economist targeting mistake. Or
• 14:23
  think about a massive automated legal error from
• 14:27
  Thompson 1. If the human is entirely removed
• 14:30
  from the final decision loop. Yeah, who is responsible?
• 14:33
  Exactly. Who is actually held accountable when
• 14:35
  the machine inevitably breaks? It is a question
• 14:38
  we have to answer soon. Thank you for taking
• 14:40
  this deep dive into the source material. Keep
• 14:42
  questioning the information ecosystem you see
• 14:44
  around you. Outio Music.