Artificial Developer Intelligence
Three engineer friends argue about AI so you don't have to. Model releases, research papers, tools, and the occasional existential crisis — no LinkedIn cringe required.
Latest Episodes
Ep 28 Jun 2, 2026Claude Opus 4.8, Undocumented Claude Code Features, Eval Harness for AI Skills, Pope on AI
Anthropic shipped Claude Opus 4.8 just 41 days after 4.7, and its new dynamic-workflow tool — high thinking plus a massive fan-out of coordinated parallel agents, or 'Gas Town, by Anthropic' — points at the default Shimin names in the cold open: trading compute for human labor, spending compute to pre-attack your own code from four angles so the human only does the final review. Rahul's out, so Shimin and Dan also cover Pope Leo XIV's AI encyclical 'Magnifica Humanitas' (models are grown not developed, AI isn't neutral, automation must protect workers — with Anthropic's Chris Olah in the room), the undocumented Claude Code features from a Building Better source-code read, a Pinterest harness for testing whether your AI skills actually fire (Codex 73%→95%, Claude 62%→73% and worse when combined), Jamie Hurst's 'Is This Sustainable?' on senior engineering and perishable AI depth, Owen McGrann's 'Dead Economy Theory,' and a Two Minutes to Midnight on SpaceX's $26.5T-TAM S-1 and Anthropic overtaking OpenAI on a $65B Series H — clock moves to 5:30.
Ep 27 May 26, 2026OpenAI Beats Musk, Gemini 3.5 Flash & AI Burnout Mitigation
Sam Altman beat Elon Musk in court — and a long New Yorker piece argues we all lost, because the real question is whether any one person should own AI safety. Rahul's out this week, so Shimin and Dan cover the OpenAI–Musk verdict (won on a statute-of-limitations technicality, with Musk's attorney arguing 'we could all die' while funding xAI, and the courtroom 'butt pillow' as a complacency metaphor), Gemini 3.5 Flash shipping Flash-only straight into AI Overviews with the viral 'can't search disregard' harness bug, a researcher's $48K six-GPU RTX 6000 Ada home server that breaks even near 80% utilization, AI-assisted-engineer burnout from Evil Martians and Siddhant Khare plus Microsoft cancelling Claude Code subs after costs topped human devs, Borretti's 'Human Bottlenecks' and the myth of the $100M startup in your laptop, Dan's rant on the 'slop grenade,' and a Two Minutes to Midnight on the SpaceX/OpenAI/Anthropic IPO index squeeze that could funnel ~10% of the major indexes into three AI firms — clock eases back to 6:15.
Ep 26 May 19, 2026LLM Neural Anatomy with David Noel Ng, Forward Deployed Everybody, Running LLMs at Home
Why is Mira Murati's first product from Thinking Machines a switchboard model rather than a frontier one? Why do beam-searched cross-layer combinations on Qwen 3.5 lose to a simple repeated middle block? And what does it mean that a data-driven AI matches your behavior better than an AI you painstakingly prompted? Shimin, Dan, and Rahul cover Thinking Machines' interaction models, Meta employees flyering against the keystroke-and-mouse surveillance program, Scott Werner's case for Palantir's forward-deployed-engineer title eating every department, an interview with Dr. David Noel Ng on LLM neural anatomy (layer duplication, multi-token prediction as wave-function collapse, and the Grace Hopper module he bought in a Bavarian pig forest), Dan running DeepSeek-V4 Flash at ~10 tok/s on a 128 GB Ryzen 395 Max plus a vibe-coded ESP32 dashboard in C, Ellis & Huang's 'Should I State or Should I Show?' paper on stated vs revealed preferences for AI alignment, and a Cerebras IPO popping 108% in the bubble check.
Ep 25 May 12, 2026Elon vs OpenAI Trial Drama, Billion Token Context Race, Multi-Agent Patterns 2026
Why is Anthropic now running on Elon Musk's GPUs — right after Elon sued OpenAI for the same kind of behavior? Why does breaking the context-window bandwidth wall require entirely new GPU architecture? And what does it mean when even Simon Willison admits he runs Claude Code with --dangerously-skip-permissions by default? Shimin, Dan, and Rahul cover Brockman's leaked deposition journal and the Tesla painting saga, the Anthropic-SpaceX/XAI Colossus One compute deal, NVIDIA's Rubin CPX disaggregation architecture on the road to a billion-token context, Phil Schmid's four sub-agent patterns for 2026, Jack Clark's 60%-by-2028 prediction for automated AI research, Simon Willison's reluctant embrace of vibe coding, Dexter Horthy publicly recanting on dark-factory agents, James Shore on why productivity gains evaporate without inverse maintenance gains, and Grok's collapse from 20M to 8M downloads.
Ep 24 May 8, 2026OpenAI's Goblin Problem, 10 Lessons When Code Is Cheap, AI Addiction Loop
Why does OpenAI's leaked Codex prompt forbid goblins, gremlins, and pigeons? Why is OpenAI gating GPT-5.5 Cyber after dissing Anthropic for gating Mythos? And what does it mean that Dan tried to write code without Claude and physically couldn't? Shimin, Dan, and Rahul cover the Codex CLI system prompt leak and RLHF post-mortem, Addy Osmani's long-running-agent patterns, Jesse Vincent's adversarial-review prompt, Drew Brunig's 10 lessons for agentic coding, Ivan Turkovic's history of programmer-elimination tools, Nilay Patel's software-brain thesis, a Nature paper on warm-model sycophancy, and Dan's three-month AI addiction loop.