AI Podcast Roundup: Week of February 3-9, 2026

Three episodes worth your time from the past week.

Latent Space - "The First Mechanistic Interpretability Frontier Lab"

Released: February 5, 2026 | Guests: Myra Deng & Mark Bissell of Goodfire AI

Goodfire just closed a $150M Series B at a $1.25B valuation, which makes this episode well-timed. They're positioning themselves as the company that takes mechanistic interpretability from research demo to production infrastructure.

The core thesis is provocative: the AI lifecycle is fundamentally broken. The only reliable control we have over models is data, and we post-train, RLHF, and fine-tune by "slurping supervision through a straw," hoping the model picks up the right behaviors while quietly absorbing the wrong ones. Goodfire's answer is to build a bi-directional interface between humans and models - read what's happening inside, edit it surgically, and eventually use interpretability during training so customization isn't just brute-force guesswork.

What makes this episode valuable is the production detail. Mark and Myra walk through what interpretability looks like when you stop treating it like a lab demo: lightweight probes that add near-zero latency, token-level safety filters running at inference time, and workflows that survive messy real-world constraints.

The Rakuten case study is particularly interesting - they deployed interpretability-based PII detection at inference time to prevent routing private data to downstream providers. The constraints were gnarly: no training on real customer PII, synthetic-to-real transfer, English plus Japanese support, and tokenization quirks across languages. Their probes ended up being operationally cheaper than LLM-judge guardrails because they don't require hosting a second large model.

The demo segment is worth watching: they steer Kimi K2 (roughly a trillion parameters) live by finding features via SAE pipelines, auto-labeling via LLMs, and toggling a "Gen-Z slang" feature across multiple layers without breaking tool use. The emerging view is that activation steering and in-context learning are more closely connected than people think.

They also discuss interpretability for science - using the same tooling across genomics, medical imaging, and materials to debug spurious correlations and extract new knowledge. The north star is moving from "data in, weights out" to intentional model design where experts can impart goals and constraints directly.

Links: Goodfire | Myra Deng | Mark Bissell

Dwarkesh Podcast - "In 36 months, the cheapest place to put AI will be space"

Released: February 5, 2026 | Guest: Elon Musk (with John Collison)

Nearly three hours with Elon, and they actually use the time well.

The headline claim is that orbital data centers will be the most economically compelling place to run AI within 30-36 months. The logic chain: electrical output outside of China is essentially flat, chip output is growing exponentially, so where does the power come from? "Magical electricity fairies?"

Solar panels get about 5x the effectiveness in space versus ground - no day-night cycle, no seasonality, no clouds, no atmosphere (which alone causes 30% energy loss). You also don't need batteries. The solar cells themselves are actually cheaper for space because they don't need heavy glass or framing to survive weather events. Once your cost of access to space becomes low, generating tokens in orbit becomes an order of magnitude easier to scale.

The hardware bottleneck discussion is where it gets interesting for anyone who hasn't tried to actually build infrastructure. Elon walks through the xAI Colossus build: they had to gang together turbines, hit permit issues in Tennessee, cross the border to Mississippi, run high-power lines a few miles. The real constraint isn't turbines - it's the vanes and blades inside them. There are only three casting companies in the world that make these, and they're backlogged through 2030. SpaceX and Tesla will probably have to make turbine blades internally.

He also breaks down why naive GPU power calculations are wrong: you need to power networking hardware, CPU and storage, size for peak cooling on the worst hour of the worst day of the year (40% overhead in Memphis heat), and have margin for taking generators offline for service (another 20-25%). His estimate: 110,000 GB300s with all support infrastructure needs roughly 300 megawatts at the generation level. A gigawatt services about 330,000 GB300s.

Other topics covered: xAI's business model and alignment plans, what it would take to manufacture humanoids at high volume in America, why China wins by default if the US can't build, lessons from running SpaceX, and DOGE.

Opening exchange sets the tone: "Are there really three hours of questions? Are you fucking serious?" "You don't think there's a lot to talk about, Elon?"

Links: Full transcript | YouTube

Cognitive Revolution - "AGI-Pilled Cyber Defense"

Released: February 8, 2026 | Guest: Alexis Carlier, CEO of Asymmetric Security

Asymmetric Security just came out of stealth, and Logan Graham (who leads Anthropic's Red Team) described Alexis as "one of the most AGI-pilled founders in the space." Nathan Labenz wanted to know what that means for cybersecurity.

The framing: if we assume AGI is coming and represents a near-infinite supply of intelligent labor, how should we redesign cyber defenses? Alexis's answer is to shift from reactive, emergency triage to proactive, continuous digital forensics.

The threat landscape breakdown is useful context. There's a spectrum: "spray and pray" tactics from financially motivated criminals, more sophisticated ransomware attacks from cybercrime gangs, and then patient, high-stakes IP theft from nation-states like China. Alexis shares details on the "North Korean remote worker" phenomenon - state-backed actors infiltrating Western tech companies not just to steal secrets but to earn salaries that fund the regime.

Asymmetric is building AI agents that can perform deep investigative work previously only available from expensive human experts. Off-the-shelf models already achieve 90% accuracy on many investigative tasks, but they're going to market with a services-first model focused on business email compromises. The reasoning: they need to deliver consistently for customers while building the proprietary dataset to close the final 10% accuracy gap.

The most interesting argument is about differential acceleration. While most AI technology is dual-use, Alexis makes the case that digital forensics specialists don't tend to become outstanding hackers. If true, this suggests forensics could be a domain where we can intentionally accelerate defensive capabilities without equally accelerating offense. Nathan calls this out as potentially the most important contribution - the strategy of shaping the capability frontier - and asks for more people to apply this thinking to other domains like biosecurity and mental health.

Links: Asymmetric Security | Episode

February 9, 2026