Taking photos for a blog post seems harmless until a QR code in the background links to your internal dashboard. While text redaction is practically solved, images remain a challenge. This episode explores the open-source tools and cloud APIs available for automated image obfuscation — from pyzbar and OpenCV for QR detection to OCR pipelines that find PII in photos. We break down the two-stage detection problem and what actually works today.

On Tisha B'Av, a day of mourning and reflection, Daniel built MK Claude—a multi-agent AI system that researches current law, stress-tests claims, and synthesizes coherent policy positions. This episode explores how the pipeline works, what it says about Israeli democracy, and what happens when a citizen decides to build the political platform nobody is offering them. From rental law reform to forkable manifestos, we examine whether AI can be a better policy ideator than the politicians we've got.

When your to-do list is longer than your waking hours and working around the clock isn't closing the gap, you've hit a scale mismatch — not a productivity problem. This episode explores the neuroscience of overwhelm, why your brain registers constant effort as failure, and what to do about it. We cover the dopamine trap of finite-project thinking, the one percent rule for all-or-nothing brains, and why scheduled breaks aren't surrender — they're part of the system. If you've ever felt like you're bailing the ocean with a bucket, this one's for you.

Three days after Anthropic launched Claude Fable 5 — their first publicly available Mythos-class model — the US Department of Commerce suspended it, blocking even the company's own foreign-national employees from accessing it. The emergency action wasn't about what the model might say, but what it could do autonomously. This episode unpacks the confusion around the Fable-Mythos relationship, the novel fallback safeguard that silently routes sensitive queries to a less capable model, and what it means when capability class becomes a regulatory category. We trace the three-week scramble from suspension to clearance, and explore why this moment marks the first time export controls have been applied to a general-purpose AI model post-launch — not to hardware or chips, but to the model itself.

When AI coding agents work in parallel, they can silently undo each other's work — a problem rooted not in bad tooling, but in fundamental amnesia. This episode unpacks three approaches to solving it: vector-backed memory servers, built-in tool memory like CLAUDE.md files, and agent-maintained documentation in the repo. We cover the failure patterns of each approach — stale embeddings, junk-drawer configs, and logs agents write but never read — plus practical fixes like read-before-write enforcement and separating decisions logs from changelogs. If you're running multiple AI coding agents on the same codebase, this episode will save you days of archaeological debugging.

Traditional software testing assumes determinism — run the same test, get the same result. Agentic AI shatters that assumption. This episode maps the emerging QA landscape for probabilistic systems: from golden datasets and LLM-as-judge to trajectory evaluation and adversarial prompting suites like Garak. We explore what carries over from traditional testing, what requires entirely new methodologies, and what a sane minimum testing stack looks like for teams shipping agentic systems.

The terminal is the purest interface we have — and also the blindest. Without visual scaffolding, our spatial memory for file paths decays faster than we can maintain it. This episode explores the full landscape of terminal navigation tools: from frecency-based jumpers like zoxide, autojump, and fasd, to shell-native options like CDPATH and pushd, to visual terminal explorers like broot and yazi. But the real conversation is about what's missing: a favourites-and-recents manager that accounts for agentic workflows where tools like Claude Code are deeply tied to the working directory. If you've ever forgotten an alias you made three months ago, this one's for you.

That "enhance prompt" button hides a dangerous secret — especially when it's running in an agentic pipeline with no human oversight. We break down the two failure modes Daniel is worried about: content loss (the enhancer drops a constraint you needed) and content invention (it hallucinates requirements you never asked for). Then we get concrete: what a well-constructed enhancer system prompt actually looks like, why distillation beats expansion, and when the best enhancement is no enhancement at all.

Most people think the cold chain is just a refrigerator on a truck. It's not. It's a system where a single degree of temperature change for thirty minutes can destroy ten thousand doses — and nobody knows until the vial reaches the clinic. This episode breaks down the chemistry, containers, and logistics that made Israel's Pfizer vaccine rollout a global case study. From dry ice sublimation and phase-change materials to GPS loggers that record every minute, we explore what actually happens when a vaccine has to stay colder than Antarctica through airports, trucks, and loading docks — to a clinic that doesn't even own a -70 freezer.

Before Levi Strauss and Jacob Davis filed their patent in 1873, working people had terrible options: wool that rotted when wet, canvas that unraveled from a single snag, linen that chafed like sandpaper, or leather that turned into a stiff, cracked mess. This episode traces the forgotten universe of pre-denim workwear, why every fabric before denim failed at the job, and how a tailor in Reno solved a mechanical problem with a copper rivet. We also explore why jeans stayed locked in the workwear ghetto for sixty years, and how they finally escaped through the unlikely gateway of rich Easterners vacationing at dude ranches.