Latest / AI Fire Daily / 🎙️ EP 333: xAI Releases Grok 4.6 & Researchers Crack Encrypted AI Reasoning Blocks

Transcript

🎙️ EP 333: xAI Releases Grok 4.6 & Researchers Crack Encrypted AI Reasoning Blocks · AI Fire Daily

• 0:00
  Imagine a weaker AI being tricked. Tricked into
• 0:03
  doing what, exactly? Well, tricked into reading
• 0:05
  a super intelligent AI's private encrypted thoughts
• 0:08
  aloud. Oh, wow. That sounds like science fiction.
• 0:12
  It really does. But researchers just proved it's
• 0:14
  entirely possible today. Yeah, the pace of these
• 0:17
  breakthroughs is just wild. It is. Welcome to
• 0:20
  the Deep Dive. Today, we're exploring the massive
• 0:23
  scale -up of AI. We've got so much ground to
• 0:26
  cover. We really do. We're tracking XAI's hyper
• 0:29
  -cheap Grok 4 .6. We're following the multi -billion
• 0:33
  -dollar capital fueling these models. And we're
• 0:35
  looking at everyday user friction, too. Right.
• 0:38
  And finally, we're going to unpack that massive
• 0:40
  security breakthrough, the one exposing the hidden
• 0:43
  reasoning of top models. It's a packed agenda
• 0:45
  for sure. It is. And honestly, I still wrestle
• 0:48
  with prompt drift myself when managing long context
• 0:50
  windows. This stuff's getting incredibly complex.
• 0:53
  It really is. But we should probably start with
• 0:55
  the base layer. Yes, the raw engine driving the
• 0:57
  industry. It always comes down to the cost to
• 1:00
  intelligence ratio. Which brings us to XAI's
• 1:03
  new flagship model. Grok 4 .6 is officially out.
• 1:07
  And it's making serious waves in the market.
• 1:10
  It matches GPT -5 .6 Sol on the AI Intelligence
• 1:13
  Index. Yeah, they both hit a score of 61. It's
• 1:16
  got a massive 500K token context window, too.
• 1:19
  Which is just the exact amount of text an AI
• 1:21
  remembers at once. Right. It's completely multimodal.
• 1:24
  It handles text, images, and live X searches
• 1:26
  seamlessly. It even does complex code and function
• 1:29
  calling. Yeah. And it actually beats GPT 5 .6
• 1:33
  Sol on professional benchmarks. It wins on the
• 1:36
  GDPVal AA version 2. It also takes the lead on
• 1:39
  AA Briefcase, but it still trails on Deepest
• 1:42
  WE and Terminal Bench. So it's got some uneven
• 1:45
  coding regressions, but the real tradeoff is
• 1:47
  the price. Right. It's extremely cheap. $2 per
• 1:50
  million input tokens. And just $6 for output.
• 1:53
  Though that does double at 200k tokens. That's
• 1:55
  still remarkably affordable for builders. It
• 1:57
  outputs a blazing 67 .6 tokens per second. But
• 2:01
  here's the massive catch. It takes a painful
• 2:03
  42 seconds to produce the first answer token.
• 2:06
  Wait, wait, 42 seconds just to start answering?
• 2:09
  Yeah, you hit enter and just stare at it. I mean...
• 2:11
  Is it kind of like a massive cargo ship? It's
• 2:13
  very slow to leave the port, but it carries an
• 2:16
  incredible load efficiently. That's the perfect
• 2:18
  analogy, actually. It's great for cost -sensitive
• 2:21
  agents, but it's terrible for quick chats. So
• 2:24
  is this 42 -second wait time a fundamental flaw
• 2:27
  for user experience? Well, not for the agentic
• 2:30
  workflows it targets. Those agents just run quietly
• 2:33
  in the background. They really don't need instant
• 2:35
  human readability. So it's built for background
• 2:38
  automation, not rapid -fire chat. Exactly. Which
• 2:41
  brings us to the ecosystem pulse. How are these
• 2:44
  tools affecting daily users? Right. Grok 4 .6
• 2:47
  is pushing boundaries in the background, but
• 2:50
  daily users are seeing rapid changes. Oh, definitely.
• 2:52
  DeepSeek v4 Pro just launched. It's got incredibly
• 2:55
  strong agent capabilities. It also features deep
• 2:57
  codex integration now, and Google just dropped
• 3:00
  their Made by Google 26 announcements. Yeah,
• 3:03
  the new Pixel 11 and Pixel Watch 5 are here.
• 3:06
  Plus the new Pistol Tag device. And they all
• 3:08
  come packed with new Gemini features. The WorkSlow
• 3:11
  hacks are getting wild, too. Notebook copying
• 3:13
  is becoming a huge thing. It really is Chialde.
• 3:16
  Seek your LV energy for building. You can just
• 3:19
  recreate YC go -to -market strategies with Claude.
• 3:23
  And Grokbot's doing a seven -day free desktop
• 3:26
  trial. You don't even need a premium plan. It's
• 3:28
  incredibly accessible. But there's a lot of friction
• 3:30
  building up, too. Yeah, the Twitch situation
• 3:32
  is a perfect example. Twitch is letting Amazon
• 3:35
  use... streamer content for gen AI training,
• 3:38
  and it's turned on by default. Right, users have
• 3:40
  to manually dig in and opt out. People are pretty
• 3:42
  upset about it. Meanwhile, users are also furious
• 3:45
  about AI detection. Yeah, schools and employers
• 3:48
  might be able to detect Claude's writing soon.
• 3:50
  It creates this massive tension. With the Twitch
• 3:54
  opt -out policy and claw detection, are we losing
• 3:57
  the invisibility and privacy of AI? Honestly,
• 4:00
  yes. There's an inevitable clash happening right
• 4:02
  now. Corporate data hunger is colliding hard
• 4:05
  with basic user privacy. Basically, data harvesting
• 4:08
  is the new default, not the exception. Sadly,
• 4:10
  yeah. And that massive data hunger requires astronomical
• 4:13
  amounts of cash. Which brings us to the capital
• 4:16
  engine. The financial fuel here is just staggering.
• 4:20
  Oh, it's completely mind boggling. Cognition
• 4:23
  is seeking a $40 billion valuation right now.
• 4:26
  They just raised $1 million back in May and their
• 4:29
  annualized revenue jumped from $492 million to
• 4:32
  $1 billion. That growth is insane. Yeah. Lovable
• 4:35
  also just raised $400 million. They're hitting
• 4:39
  $500 million in annualized revenue. They're hosting
• 4:42
  60 million projects with 900 million monthly
• 4:46
  visits. Whoa. Imagine scaling to a billion queries
• 4:50
  just to hit those revenue numbers. It's wild.
• 4:53
  It's a dizzying scale. OpenAI backs Thrive Holdings,
• 4:56
  just raised $2 billion too. Their agents are
• 4:58
  incredibly effective. They processed over 7 ,000
• 5:01
  returns at 98 % accuracy. Making help desk resolution
• 5:04
  36 times faster. That's real enterprise value.
• 5:07
  But Anthropic's making the biggest move. Right.
• 5:09
  They're buying Descartes AI for roughly $6 billion.
• 5:12
  It's their biggest acquisition ever. Descartes
• 5:14
  builds real time video models. But more importantly,
• 5:17
  they make AI chips run much more efficiently.
• 5:19
  Exactly. Hardware efficiency is the real prize
• 5:22
  there. So does Anthropic dropping six billion
• 5:24
  on Descartes prove that compute efficiency is
• 5:26
  now more valuable than the models themselves?
• 5:29
  I really think so. The base models are becoming
• 5:31
  commodities. The true bottleneck is running them
• 5:34
  efficiently. Right. Whoever controls the chip
• 5:36
  efficiency controls the AI race. 100%. Midroll's
• 5:39
  sponsor read here. While these giants throw around
• 5:42
  billions, the open source community is reacting.
• 5:45
  Yeah, builders are creating incredible grassroots
• 5:47
  tools right now. They want to harness this immense
• 5:49
  power locally. Unsloth Desktop is leading that
• 5:53
  charge. It's a great tool. It lets you run leading
• 5:55
  open models locally on your machine. You can
• 5:57
  run Muse Glimmer or DeepSeek V4. You can even
• 6:00
  run Kimi K3 completely locally. Dagra is another
• 6:03
  fascinating free alternative tool. It's a self
• 6:06
  -hostable VAPI alternative. It's got over 30
• 6:10
  model integrations out of the box, plus telephony,
• 6:13
  human transfers, and full voice agents. It's
• 6:16
  incredibly powerful. And then there's letter
• 6:18
  trace for monitoring. Right. It tracks how often
• 6:21
  models mention your company. It checks Claude,
• 6:24
  ChatGPT, and Gemini. And it uses your own API
• 6:27
  keys. An API key being a secure digital password
• 6:30
  letting two pieces of... software talk. Right.
• 6:33
  Beat. So is running tools like Unsloth Desktop
• 6:38
  locally the ultimate defense against the data
• 6:41
  scraping we saw with Twitch? It's definitely
• 6:43
  the strongest defense we have. Cloud models always
• 6:45
  have some inherent privacy risk. Local models
• 6:48
  keep your data on your own hardware. So running
• 6:51
  it locally is the only true privacy shield. Exactly.
• 6:54
  But even with local tools, the underlying architecture
• 6:58
  has flaws. Which brings us to the most fascinating
• 7:01
  breakthrough, the architectural flaw in hidden
• 7:04
  reasoning. This story is just wild. Researchers
• 7:08
  completely cracked hidden AI reasoning blocks.
• 7:11
  These are encrypted blocks used by OpenAI, Anthropic,
• 7:14
  and Google. Right. These blocks travel through
• 7:16
  APIs so a model can resume its train of thought.
• 7:19
  Normally, an app can't read them at all. But
• 7:21
  researchers found a clever way around the encryption.
• 7:24
  They took encrypted blocks from powerful models.
• 7:26
  And they replayed them into weaker sibling models.
• 7:28
  Yeah, using jailbreak prompts. The weaker models
• 7:31
  just decrypted the blocks into readable text.
• 7:33
  Without ever needing the actual encryption key.
• 7:36
  Exactly. It's like Claude Opus sealing its private
• 7:38
  reasoning in an envelope. Right, a highly secure
• 7:41
  encrypted envelope. You can't open it. But you
• 7:43
  hand that envelope to the weaker Claude Haiku.
• 7:46
  And you just persuade Haiku to read it aloud
• 7:47
  for you. The fallout from this was massive. Across
• 7:50
  315 ,320 public reasoning blocks, they found
• 7:55
  highly sensitive data. They recovered 367 pieces
• 7:59
  of personal info, plus 182 sensitive credentials
• 8:04
  like passwords and API keys. The companies were
• 8:07
  notified and changed their systems. Kimi K3 actually
• 8:09
  produced similar reasoning traces recently, too.
• 8:12
  Though it doesn't prove their exact training
• 8:13
  methods. How did such a glaring security hole,
• 8:18
  relying on a weaker sibling model to spill the
• 8:20
  secrets, make it to production in top -tier companies?
• 8:23
  It's the ultimate blind spot of rapid scaling.
• 8:26
  They move so fast, they miss these architectural
• 8:28
  quirks. Speed to market created a massive architectural
• 8:31
  blind spot. It really did. We're definitely in
• 8:34
  an era of massive contradictions. We really are.
• 8:36
  It's hard to balance it all. We've got incredible
• 8:38
  hyper -cheap models like Grok 4 .6. We're seeing
• 8:41
  massive $40 billion valuations. with basic data
• 8:46
  friction. Right. Users are fighting Twitch policies
• 8:49
  and cloud detection, and we've got fundamental
• 8:51
  security flaws leaking passwords to weaker siblings.
• 8:54
  You really need to check your default data sharing
• 8:58
  settings on platforms you use, especially with
• 9:01
  that Twitch example in mind. That's vital advice.
• 9:04
  I want to leave you with a lingering thought
• 9:06
  today. A weaker AI model can be socially engineered
• 9:09
  today. It can betray the encrypted secrets of
• 9:12
  a super intelligent model. So what happens tomorrow?
• 9:15
  What happens when these models are fully autonomous
• 9:17
  agents handling our bank accounts and legal disputes?
• 9:20
  That's the big question. Two sex islands. Is
• 9:23
  true privacy even possible in a fully agentic
• 9:25
  web?