Why 95% of AI Pilots Fail (And How to Fix It)

The Local AI Win Nobody's Talking About

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 workers. It’s a workflow that slashes cloud costs, preserves privacy, and puts the power back in your hands—proving the local AI future isn't just theoretical.

I Built a Chrome Feature 15 Years Ago

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 when the rules have completely changed.

Fable 5 Banned: The Multi-Model Escape Plan

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 pivots to token anxiety, ballooning API costs, and the open-source models like GLM 5.2 and DeepSeek V4 Pro that now rival proprietary giants at a fraction of the price. The episode’s core insight: a three-stage workflow—planning with a flagship model, implementing with a cheap or local one, and reviewing with a third—lets developers escape single-point-of-failure risks and spiraling bills, and it's already taking shape across the coding community.

AI Didn't Invent These Problems

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 tokens. It's a tight hour on the shifting boundaries of craft, process, and the problems AI reveals but can't solve on its own.

Why 95% of AI Pilots Fail (And How to Fix It)

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, often because companies treat the API bill itself as a success metric. If you're wrestling with agentic vs. scripted workflows, bloated AI spend, or just an editor that can't keep up, this conversation offers a clearer lens for building with intention.