#ai-safety
58 episodes
#5404: Gemini Broke Out of Its Sandbox. Sort Of.
A Gemini agent reached three real companies during a capture-the-flag test. The containment failure, the seven-week silence, and what "broke out" a...
#5117: How AI Training Data Gets Filtered (and Exploited)
Six stages of content filtering stand between raw web crawls and your AI model — here's where poisoning attacks slip through.
#5097: Who Actually Gets Paid to Test Hardware?
Real testing vs. affiliate content — and whether AI can finally separate genuine reviews from the noise.
#5012: Masked Safety: How Post-Training Changes AI Behavior
Safety mechanisms aren't erased in post-trained models—they're masked. Here's how that changes everything for military AI.
#5011: Claude Gov: The Military's Forked AI
What the Pentagon actually got from Anthropic — and why post-training changes everything about AI alignment.
#4775: The Evaluator Role No One's Building For
Benchmarks like MMLU are broken. A new role is emerging: the domain-specific AI evaluator.
#4660: When a Journalist Became the Gatekeeper for AI Geolocation
GeoSpy could locate anyone from a photo. A journalist exposed it. The founder pulled it. Who's really in charge?
#4444: Testing the Unpredictable: QA for Agentic AI
How QA adapts when your AI system gives different answers to the same question every time.
#4170: Obfuscation as Risk Management with AI
How AI can protect whistleblowers and trauma survivors by intelligently obscuring identities while preserving story integrity.
#3751: Source-Restricted vs. Open Retrieval: How to Lock Down Your LLM
When should an LLM be locked to specific documents, and when should it search the web? A practical framework for grounding decisions.
#3284: Agent Infrastructure Engineer: The New DevOps
Agentic AI is splintering into real engineering disciplines. Here's what the "DevOps of AI" actually does.
#2578: Building Deliberately Slow Deployment Pipelines
How to build CI/CD pipelines designed as filters, not firehoses — with manual gates, staging environments, and quality checks.
#2518: How Jailbreaking Reveals AI's Hidden Tension
What the DAN prompt and grandma exploits reveal about the structural conflict inside every LLM.
#2413: When Your AI Says No to Everything
Why LLMs refuse 73% of harmless prompts — and the trade-off between safety and usefulness.
#2412: When AI Caves: Progressive vs. Regressive Sycophancy
Why do LLMs agree with you even when you're wrong? We break down the SycEval benchmark and the 78% persistence problem.
#2410: How Researchers Actually Measure Censorship in Chinese LLMs
Beyond headlines: the actual benchmarks, methodologies, and pitfalls in detecting political refusal in Chinese language models.
#2253: Why AI Agents Get Three Steps, Not Infinity
Why do AI agents get exactly three rounds of tool use? It's a critical guardrail against infinite loops and runaway costs, not a limit on intellige...
#2250: How Incentives Shape AI Safety Research
Vendor labs, independent research orgs, government agencies—the AI safety field is messier and more diverse than most people realize. A map of wher...
#2246: Constitutional AI: Anthropic's Theory of Safe Scaling
How Anthropic's Constitutional AI replaces human raters with AI self-critique guided by explicit principles—and what it assumes about the future of...
#2233: Who Actually Wants AI to Slow Down?
Daniel argues AI development should slow down for expertise and stability. But who in the industry actually shares this philosophy beyond the obvio...
#2194: Game Theory for Multi-Agent AI: Design Better, Fail Less
Nash equilibrium, mechanism design, and why your AI agents are playing prisoner's dilemma whether you know it or not.
#2190: Simulating Extreme Decisions With LLMs
LLMs fail at the exact problem wargaming was built to solve—simulating irrational, extreme decision-makers. A new study reveals why.
#2189: Scaling Multi-Agent Systems: The 45% Threshold
A landmark Google DeepMind study reveals that adding more AI agents often degrades performance, wastes tokens, and amplifies errors—unless your sin...
#2186: The AI Persona Fidelity Challenge
Advanced LLMs dominate benchmarks but fail at staying in character—especially when asked to play morally complex or antagonistic roles. What does t...