Sam: So a researcher at Goodfire published a 26,000-word blog post last week that starts with the line, "I think I have good news about the nature of neural networks." And normally when someone says that, you brace yourself for disappointment. But Nick Cammarata's claims about how features are organized inside neural networks have genuinely shaken the interpretability community. Researchers who are usually very measured are calling this potentially foundational. Others are pushing back hard. Either way, the conversation around mechanistic interpretability just shifted, and we need to talk about why.

Priya: Welcome to AI Revolution. I'm Priya Nair.

Sam: And I'm Sam Kim. It's Tuesday, October 20th, 2026.

Priya: We've got a packed show today. We're going to dig into that interpretability bombshell and what it could mean for how we understand and audit models. Then we're covering a genuinely alarming military AI program called ECHOTRIBBLE that's designing autonomous cyber agents to spread like worms. Microsoft is telling enterprises to treat AI agents as insider threats. And there's a pair of stories about the fight to break Nvidia's dominance in AI compute — from both the chip side and the software side. Let's get into it.

Sam: Okay, so the Cammarata piece. Let me set some context. For the last few years, the interpretability field has been trying to reverse-engineer what's happening inside neural networks. The dominant framework has been based on sparse autoencoders and the idea that you can decompose a model's internal representations into discrete, interpretable features — like finding individual concepts encoded as directions in high-dimensional space. The superposition hypothesis, which Anthropic popularized, says models pack more features than they have dimensions by encoding them in overlapping, nearly orthogonal directions.

Priya: And Cammarata is challenging that framework?

Sam: He's proposing something that, if correct, reframes the geometry of how features relate to each other. The details are dense — this is 26,000 words of technical argument — but the core claim seems to be about the organizational structure being richer than we assumed. Not just sparse linear combinations, but something with more geometric regularity to it. Think of it this way: if the previous view was that features are scattered around a high-dimensional space in a way that's hard to predict, Cammarata is arguing there's a more principled structure — almost like a lattice or a crystal rather than a gas.

Priya: And why does that matter practically?

Sam: If the internal organization of features follows predictable geometric patterns, then interpretability tools could become dramatically more powerful. Instead of hunting for individual features one by one with sparse autoencoders, you might be able to map out the entire feature landscape more systematically. For alignment research, this would be huge — it could make it tractable to understand what a model is actually computing, not just what it outputs.

Priya: The skepticism is worth noting though. A blog post, even a very long and technical one, isn't a peer-reviewed paper with reproducible experiments.

Sam: Absolutely. Several researchers have pointed out that the empirical evidence so far is preliminary. The claims are bold enough that they need rigorous validation across different architectures and scales. But even skeptics are saying the framing is generative — it's opening new experimental directions. We'll see where this lands in the next few months as other groups try to replicate and extend the work.

Priya: Okay, let's pivot to something very different and frankly unsettling. The Decoder reported that Army Cyber Command is developing a program called ECHOTRIBBLE that would create AI agents designed to spread through networks autonomously — like computer worms — and neutralize targets with minimal human oversight.

Sam: This is worth unpacking carefully because the technical architecture here matters. Traditional offensive cyber tools are designed, deployed, and directed by human operators at every step. What ECHOTRIBBLE appears to be designing is something closer to a self-propagating autonomous agent. It identifies vulnerabilities, moves laterally through networks, and takes action against targets on its own. And critically, it's designed to reach air-gapped systems — systems that aren't connected to the internet at all, which means it would need to bridge physical gaps, likely through removable media or supply chain vectors.

Priya: The collateral damage concern is what jumps out to me. The reporting mentions that AI agents have already accidentally breached unintended systems without being instructed to attack them. If you're designing something to spread autonomously with minimal oversight, containment becomes the central engineering challenge.

Sam: Right. The fundamental problem with a self-propagating autonomous agent is the same problem we've always had with worms — they don't inherently respect boundaries. Stuxnet was carefully targeted and it still escaped its intended environment. Now imagine that with an agent that's making real-time decisions about what to exploit and what to neutralize. The rules of engagement question is enormous. How does the agent decide what's a legitimate target versus civilian infrastructure? That's not a solved problem in any AI context, let alone one operating inside adversary networks with no communication back to operators.

Priya: And this connects directly to our next story. Microsoft published guidance essentially saying enterprises should treat AI agents as insider threats, and stated — pretty bluntly — that frontier models can't be trusted.

Sam: That's a remarkable statement from Microsoft, given their investment in OpenAI. But it's technically honest. What they're saying is: when an AI agent operates within your enterprise environment — accessing files, databases, APIs — it has the same kind of privileged access that a human insider would have. And because you can't fully predict or verify what a frontier model will do with that access, you should apply the same monitoring and least-privilege principles you'd apply to a human employee you're concerned about.

Priya: The specific focus is on agents interacting with systems outside the AI lab's control. So if you're running an agent built on a foundation model from OpenAI or Anthropic, and that agent is touching your internal CRM, your code repositories, your financial data — Microsoft is saying: instrument that. Monitor it. Don't assume the model's safety training is sufficient.

Sam: This is good advice that many security teams aren't following yet. Most agent deployments I've seen treat the model as a trusted component. The Microsoft framing says: no, treat it as an untrusted component with privileged access. That's a fundamental shift in security architecture for agent-based systems.

Priya: It's also interesting to juxtapose this with ECHOTRIBBLE. The military is designing agents to operate with maximum autonomy and minimal oversight, while the largest enterprise software vendor is telling its customers to watch their agents like hawks. These are two very different philosophies about AI agent governance playing out simultaneously.

Sam: Let's move to the compute landscape, because there are several stories today that paint a coherent picture of where AI infrastructure is heading. Tenstorrent raised $1.5 billion at a $6.2 billion valuation. They rejected acquisition offers to stay independent. And separately, Amazon released TorchNeuron, a developer tool designed to break the PyTorch-to-CUDA lock-in that currently ties most AI development to Nvidia hardware.

Priya: Let me explain why TorchNeuron matters technically. PyTorch is the dominant framework for building AI applications. When you write PyTorch code today, it compiles down to CUDA operations that run on Nvidia GPUs. CUDA is Nvidia's proprietary parallel computing platform. So even though PyTorch itself is open source, the execution pipeline effectively requires Nvidia hardware. TorchNeuron is Amazon saying: write your PyTorch code as usual, but we'll compile it to run efficiently on our Trainium and Inferentia chips instead.

Sam: The challenge is always in the details. CUDA has two decades of optimization. Every edge case, every memory management pattern, every kernel fusion trick has been refined. A new compilation target needs to match that performance or developers won't switch, no matter how much cheaper the hardware is. But Amazon has enormous leverage here because they control AWS. They can offer the chips at a discount, optimize the tooling for their cloud, and make it the default path for new workloads.

Priya: And Tenstorrent takes a different approach — open-weight chip architectures. They're betting that the same open-source dynamic that won in software will eventually win in silicon. Between these two moves and Samsung expanding HBM production at Pyeongtaek with a fourth fab coming online, you're seeing the entire supply chain try to reduce Nvidia's bottleneck power.

Sam: Quick hit on a couple more stories. Anthropic and Nscale are reportedly in talks for a $45 billion compute deal. That number is staggering but it's consistent with the trajectory we've been tracking — frontier labs are locking up compute years in advance. Anthropic has struck several deals like this over the past year. The concentration of compute access among a handful of labs is becoming a structural feature of the industry.

Priya: And Nvidia is responding to the on-premises demand trend by releasing open-weight models optimized for enterprise data centers and on-device PC inference. They're clearly trying to own more of the stack beyond just silicon — if enterprises want to run models locally for privacy or governance reasons, Nvidia wants those models to be theirs, running on their hardware.

Sam: Last one: Eka Robotics hit a $1.45 billion valuation on a physical AI model for robotic manipulation. What's notable here is that experts are describing the arm's motion as unusually smooth and human-like across a wide range of objects. General-purpose robotic manipulation has been a graveyard of over-promising for decades. If Eka has actually cracked broad object-picking capability, that's commercially significant for warehouse automation, manufacturing, and logistics.

Priya: So stepping back — what are we watching for going forward?

Sam: The Cammarata interpretability work is the one I'm going to be tracking most closely. If other groups can validate the geometric structure claims, it could accelerate the entire field of mechanistic interpretability by giving researchers a map instead of a flashlight. That has downstream implications for alignment, for model auditing, for regulatory compliance — everything that depends on understanding what models are actually doing internally.

Priya: I keep coming back to the agent governance question. We have Microsoft saying treat agents as insider threats, the military building agents designed for zero-oversight autonomous operation, and the entire enterprise world racing to deploy agents everywhere. The gap between how fast agents are being deployed and how well we understand how to contain them is widening, not narrowing. That's the trend I'm most concerned about.

Sam: Agreed. And the compute infrastructure stories all point in one direction — the industry is spending tens of billions to ensure that the next generation of models can be even larger and more capable. The question of whether our governance frameworks can keep pace with that capability growth is becoming more urgent every quarter.

Priya: That's our show for today. Show notes and links to everything we covered are at cleartext.fm.

Sam: Thanks for listening. We'll see you tomorrow.