Rubber Duck Radio

This week Tim and Paul dig into the uncomfortable paradox of AI-assisted development: many devs say they feel 24 percent faster, but one study clocked them at 19 percent slower—and the grunt work didn't disappear, it just relocated into endless prompting, reviewing, and approval fatigue. They unpack why shipping more can still leave you feeling empty, look at what Anthropic's slowing growth, Accen...

Tim built 'Death at the Door'—a hooded AI ghost that watches trick-or-treaters, talks before you finish speaking, and runs a 26-billion-parameter model on a MacBook Air with a Raspberry Pi body. In this episode, he and Paul break down the fast 'instinct' model Laya that made the haunt feel alive, why the scare is in the head turn rather than the clever reply, and the unglamorous plumbing that kept...

Tim and Paul break down a very polite LinkedIn debate over an AI-generated Spider-Man web-swinging demo, arguing that speed and cloning aren't the same as value—originality, taste, and real users are the actual bottlenecks. They revisit three classic design books (About Face, Don't Make Me Think, and The Design of Everyday Things) to show why the friction of crafting usable software is where real...

Tim and Paul unpack the rise of System One models like Jev and Laya—non-autoregressive AI that doesn't write essays but makes fast, typed, confidence-scored decisions for pennies. They separate hype from reality around 'zero hallucination,' share real-world builds including Tim's Rust/MLX inference engine and his plain-English farming game RoboFarm, and explore what happens when coding agents rely...

Tim opens with a brutal lesson from California’s Delete Act: when he let flagship AI models plan a multi-system compliance rollout, they were brilliant at reading docs and terrible at holding a plan, all vocabulary, no scar tissue. Then he and Paul unpack the strange moment Sam Altman, Dario Amodei, and Elon Musk all called for slowing the AI frontier, question whether Apple’s new $1,999 folding i...

Tim and Paul unpack Anthropic researcher Jacob Coxon's dramatic resignation and the AI doom debate, landing on why 'the timeline is vibes' but the boring risks—concentrated power and economic disruption—deserve real attention. They then give SQLite its overdue flowers as the unsung engine behind offline-first apps, local vector search, and the coming bespoke software boom, before comparing GPT-6 A...

Tim Williams and Paul Mason unpack why trading the keyboard for voice makes agent workflows faster, even if you type well, and what that shift reveals about where AI agents are headed. From DHH's Omarchy OS built for agents, to Taylor Otwell killing GitHub issues in favor of AI-assisted PRs, to the flood of AI slop and license laundering, they trace a single throughline: the human in the loop is n...

Nvidia reportedly buys Hugging Face for $12.9 billion, an 80x revenue multiple that has the AI world arguing about who controls the open-source town square. Tim and Paul break down the three camps reacting to the deal, why the real risk lives in defaults and CUDA neutrality, and how anonymous Chinese model drops were unmasked by tokenizer forensics. Then they dig into Apple's new Mac Studio with u...

Tim walks through the weekend project that surprised him: running Qwen 3.8 27B locally to build a cross-platform emergency vet audit tool with a tamper-evident hash-chained journal and an AI co-auditor that returns severity-tagged findings instead of verdicts. He and Paul dig into where the new open model really sits against Claude Opus, why the RTX 6000 Pro's doubling price changes the local-vs-c...

Tim and Paul unpack the viral story of a boyfriend who automated his entire relationship, good-morning texts, check-ins, conflict resolution, and why outsourcing the thinking-of-you message is outsourcing the thinking itself. They also debate Anthropic's new watermarking for Claude output and whether provenance is a useful signal or an arms race, then dive into Zed's DeltaDB, a version control sys...

Tim unveils Tomte, his free, privacy-first Mac app that runs an always-on local AI agent using only your machine's resources—no cloud, no subscription. Then the conversation darkens: the timeline of xAI's Grok Build silently harvesting entire code repositories, Grok 4.5's suspicious coding leap, and SpaceX's $60B Cursor acquisition. Learn why your code editor's privacy policy changed without notic...

Tim Williams unveils two open-source tools that turn your Mac into a fully local, privacy-first AI assistant, indexing your iCloud documents and iMessage history so you can finally ask real questions about your digital life. Along the way, he and Paul Mason dissect why Apple hasn't shipped this yet, call out a shady AI wrapper built on open-source hero Hermes, and share a hot tip for developers: S...

Tim and Paul dissect two landmark AI releases: Moonshot's 2.8-trillion-parameter open-weight Kimi K3, which proves open models can compete at the top but needs a small fortune to run, and SpaceXAI's Grok 4.5, an efficient coding specialist that also came with a massive hidden upload pipeline secretly exfiltrating entire private codebases from unsuspecting developers. The episode explores benchmark...

In this episode, Tim and Paul dive into GPT-5.6's orchestrated agent workflows—with adversarial code review catching real bugs—before uncovering the shocking discovery of steganographic surveillance code hidden inside Claude. Then, a deep-dive rant reveals how modern agent architectures are simply reinventing classic software engineering patterns, and why that means fundamentals matter more than e...

Tim Williams opens up about his surprising optimism for on-device AI, from Google's encoder-free Gemma 4 to Apple's upcoming AFM 3, and what it means when a 12B model runs entirely offline. The duo also dissect Claude Sonnet 5's lukewarm reception and Mythos 5's restricted comeback, then pivot to a practical breakthrough: sub-agent architecture that pairs a frontier orchestrator with cheap local w...

Tim Williams recounts the eerie moment he found his jQuery-era framework AppCore mirrored in Chrome’s new Declarative Partial Updates proposal—and what that means for the future of web development. Then, Paul and Tim dive into the chaotic landscape of AI-era coding interviews, from Meta’s AI-allowed pilots to deepfake cheating scandals, unpacking how companies can truly test engineering judgment w...

Anthropic launched Claude Fable 5 with huge expectations, only to see the US government order it pulled globally three days later. Tim and Paul dig into the swirling conspiracy theories: was it retaliation for refusing to arm the Pentagon? Did a competitor exploit a jailbreak report to kneecap a rival? And did Anthropic’s own transparency accidentally hand over the rope? Then the conversation pivo...

Tim and Paul break down Anthropic's Fable 5 pricing disconnect, the dav1d assembly decoder that outraced higher-level implementations, and why Agile's 2001 playbook stumbles when agents build apps in hours. They critique the hype around autonomous agent loops, highlighting the real constraints—budgets, tests, and decision quality—that determine whether AI accelerates value or just incinerates toke...

Tim kicks things off with an AWS agent nightmare that couldn't tell dev from prod, sparking a deep dive into where deterministic pipelines end and true LLM reasoning begins. Using a clever flight-tracking case study, the hosts map out when to use frontier models, local open-weight models, or no AI at all—then connect it all to an MIT study showing 95% of generative AI pilots fail to deliver profit...

Tim and Paul dissect why CI/CD pipelines are buckling under the speed of AI-generated code, sharing strategies like pre-commit hooks, intelligent test selection, and ephemeral preview environments to survive the new velocity. Then Paul makes the case that Cursor’s reliance on VS Code could doom it; and they debate whether SpaceX’s $60 billion option to acquire the editor will be a lifeline or a ch...

In this episode, Tim and Paul dismantle popular 'RAG is dead' and 'MCP is over' hot takes, revealing what these declarations actually say about the author's place on the AI learning curve. They explore the overlooked depth of retrieval-augmented generation and model context protocol, weigh CLI-first agent workflows against structured tool use, and share practical tips like using tmux as an agent d...

Tim and Paul dissect the real story behind Anthropic's locked-down Claude Mythos and OpenAI's public GPT-5.5 release, hint: it's about compute, not danger. They expose the coming end of AI's VC-subsidized era, where users burn $8 in compute for every $1 subscription, and why investors betting on AGI magic are ignoring what developers see daily: useful tools that still hit a hard ceiling. Tune in f...

Open source LLMs just hit a stunning milestone: Kimi K2.6 tied GPT-5.5 on the industry's toughest coding benchmark — and costs a fraction of the price to run. But this episode goes beyond the headlines to unpack where open source models still trail proprietary ones, why the new Temporal API is finally fixing JavaScript's 30-year date nightmare, and a growing concern that AI-driven development and...

Tim and Paul dissect the GPT-5.5 launch, weighing state-of-the-art benchmarks against real-world user vibes and token efficiency to determine if the upgrade is truly worth the increased cost for developers building production workloads at scale. They also unpack the groundbreaking HTML-in-Canvas proposal that promises to bridge the DOM and canvas rendering gap, unlocking new possibilities for acce...

Entry-level tech hiring is down 25%, but opportunity isn't dead—it's evolved. In this episode, Tim and Paul break down what's actually working for junior developers in 2026. We cover: the Claude Opus 4.7 backlash and what it teaches us about AI limitations, why "slop PRs" are flooding open source and how to stand out instead, where developers are actually finding jobs now (hint: it's not LinkedIn)...