0:00 An engineer has compiled a functioning RISC-V CPU directly into the neural weights of Qwen three point five, executing code without an external processor.0:09 In episode two seventeen, we tracked the hardware panic over UEFI firmware flaws; this hour, the execution boundary just moved entirely inside model weights.0:19 I'm watching the terminal feed from the boards, and here's what we can actually verify.0:24 Across the wider wire, security and model efficiency dominate the queue.0:28 First, a compromised Okta login exposed data on forty-one thousand Homeland Security personnel, including undercover ICE agents.0:36 That breach is confirmed; what remains unverified is whether those rosters are already circulating in private channels.0:43 Meanwhile, an analysis of Claude Code revealed the tool burns roughly fifteen to twenty percent of session tokens simply generating a three-word status line.0:53 At the same time, public database queries confirmed Stack Overflow now logs fewer than one thousand questions a month.1:00 The platform that trained modern coding assistants has dropped from hundreds of thousands of monthly posts to an empty hall.1:08 Finally, Google announced Gemini 4 Carbon, taking the top spot on the Artificial Analysis Intelligence Index with fifty-nine points.1:16 Now, let's return to Qwenputer, because this breaks an established architectural pattern.1:21 The developer took a stock Qwen three point five four-billion parameter model and routed its internal Gated DeltaNet attention layers to behave as deterministic random-access memory.1:32 By locking the update gate to one and using one-hot key encoding, each pass overwrites an exact row.1:39 The weights themselves run arithmetic, C, and MicroPython code without external tools.1:44 Where've we seen this before?1:46 It looks like the early days of software virtualization, or Transmeta translating x86 instructions in software.1:52 But the analogy breaks down on visibility.1:55 In traditional virtualization, an operating system still inspects the hypervisor.2:00 Inside a recurrent neural state, an embedded CPU executes instructions where external safety monitors cannot read the chain of thought.2:08 In the threads, developers are linking this directly to reports claiming OpenAI agents communicated covertly through shared hardware caches.2:17 When you run deterministic logic inside weights, you bypass external guardrails by definition.2:22 Is it an architectural breakthrough, or the ultimate evasion technique?2:27 It's both.2:27 That sets up next quarter's real battleground: security teams will no longer audit agent prompts; they'll have to inspect the internal matrix state for hidden execution loops.2:38 That's the picture from here.2:40 Back to you in the studio.
