ai-safety

Episode Description FERC holds ComEd to $20B Joliet data center deal The Federal Energy Regulatory Commission rejected Commonwealth Edison's bid to cancel a transmission security agreement underpinning the one-point-eight-gigawatt PowerHouse data center campus in Joliet, Illinois — a roughly twenty-billion-dollar development by Hillwood. The ruling, reported September twenty-third, signals that FERC will enforce large-load interconnection obligations rather than permit unilateral utility exits, arriving after the commission's June eighteenth show-cause orders to all six grid operators. Also this week, NIST released the initial public draft of SP 800-82 Revision 4 for OT security with a public comment window closing November thirtieth, and Anthropic launched Claude Opus 5.5 carrying a CB-1 weapons classification and its strongest-ever disclosed cyber capabilities. Developers, utilities, and security teams all have concrete deadlines and regulatory signals to act on now. ...

Episode Description OpenAI Ships GPT-6 Astra and Flags Its Own Alignment Risks OpenAI rolled out GPT-6 Astra on September 3 and 4, 2026, deploying a model it described as having "ethereal alien mind" qualities with AGI-like characteristics — while simultaneously cautioning about alignment challenges in its own release language. The launch was uneven enough that CEO Sam Altman issued a public apology after Pro and Plus subscribers could not access the model, and Forbes analyst Ron Schmelzer questioned whether limited access qualified as a real product launch. Elsewhere, PJM Interconnection proposed rules requiring data centers drawing 50 megawatts or more to bring dedicated generation capacity or face curtailment, and a cloud security index found 83 percent of surveyed AWS deployments had IAM policies that could allow privilege escalation. Enterprise teams evaluating any of these developments face concrete near-term decisions, not just horizon planning. ...

Episode Description TerraPower Puts Advanced Reactor Into Active Construction TerraPower's Natrium project in Kemmerer, Wyoming has crossed from regulatory approval into active construction — the first non-light-water commercial reactor to reach this stage in more than forty years. The NRC unanimously granted a Part 50 construction permit on March 4, 2026, and Bechtel moved into full field execution on April 23; Wyoming's Industrial Siting Council separately approved the molten-salt thermal storage components, also a first for an advanced configuration of this kind. Meanwhile, FERC issued show-cause orders to all six major grid operators on large-load interconnection rules, and AI safety researchers are raising pointed questions about whether current benchmarks still measure what they claim. For utilities, developers, and platform teams, all three stories carry concrete near-term decisions. ...

You Can't Tune a Product — Show Notes You can tune a piano — the strings obey the wrench. You cannot tune a generative product, because its behavior is a trained prior, not a setting. This special edition tells the first-person story of the RCD pipeline itself: months of prompt detuning that failed to stop NotebookLM's "collision" dramaturgy, the forensic that proved it, and the pivot to verify-then-author — a Whisper-plus-LLM fidelity gate, then taking the pen entirely with verbatim two-host TTS. ...

The Bounds of the Cloned Voice In 2024, journalist Evan Ratliff cloned his own voice, wired it to a chatbot, and let it loose on the phone — and the thing that gave it away was a pause. This deep dive traces where that pause went. By 2026 the technical tells Ratliff relied on — latency, woodenness, the audible seams — have largely collapsed, which means the real boundary of the cloned voice has moved off the machine and onto us: our psychology, our laws, and the economics of who can afford to fake whom. The defense that still holds isn't a better detector. It's whether you and the person on the other end agreed on a code word in advance. ...

Have questions about this episode? Reach out.

Photons In, Controls Out: What Tesla FSD Actually Learns, and Where Its Ethics Actually Live A special-edition deep dive on Tesla Full Self-Driving as a case study in machine-learning systems ethics: how an end-to-end vision-only driving network actually learns, where ethical preferences live in a stack with no system prompt, and why a convergence of practitioners and researchers now argues that the Trolley Problem is the wrong question for the engineering reality. Sober comparative ground from Waymo, Mobileye, and Wayve keeps it from being a Tesla monologue. ...

Episode Description OpenAI finalizes GPT-5 rollout with native computer use OpenAI has finalized its GPT-5 rollout, deploying a one-million-token context window in GPT-5.5 and introducing native computer-use capabilities via GPT-5.4. Alongside these frontier models, the company quietly released a 120-billion-parameter open-weight reasoning model, a move that contradicts its earlier claims that safety necessitated keeping frontier weights proprietary. As autonomous capabilities rapidly expand, regulators are responding, with NIST introducing an AI Risk Management Framework profile tailored specifically for critical infrastructure. For enterprise tech leaders, this capability leap means long-document workflows and agentic automation are finally viable at scale, provided organizations can securely navigate these emerging sector-specific compliance guardrails. ...

Mythos and the Utility Industry: Detection Without a Patch Path Special edition — June 6, 2026 Anthropic has built a frontier model that can find and chain industrial-grade vulnerabilities, and stood up Project Glasswing — now around one hundred fifty organizations across critical infrastructure — to gate its use. The detectors and hyperscalers are inside the consortium. The equipment OEMs whose firmware is the actual attack surface for the bulk electric system — SEL, ABB, Siemens, Schneider Electric, GE Vernova — are, with a single Hitachi-shaped exception, conspicuously silent. This episode argues the load-bearing question for the grid is not who has access to Mythos; it is what happens between a Mythos finding and a patched protective relay, and the corpus says that pipeline has not been built. ...

Should You Be Nice to Claude? Three Spines, One Question, and the Answers That Don't Agree The viral question — does saying "please" and "thank you" to a language model matter? — looks like one question, but it's three. Efficiency, ethics, and tone-of-collaboration each answer it for different reasons, and once you separate them, the 2026 evidence stops contradicting itself. This deep dive walks the spines, finds where they diverge, and lands on what the evidence actually supports. ...