Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Is Anthropic a Cult? AI Beats ER Doctors & Recursive Self-Improvement | This Week in AI E12
TL;DR
Healthcare AI is already outperforming doctors in narrow diagnostic tasks — The hosts cite a Harvard/Beth Israel-style ER study where OpenAI’s o1 scored about 67% on 76 real emergency cases versus roughly 50–55% for physicians, with a 98% clinical reasoning score on one task, framing AI copilots as the obvious next step.
The real bottleneck for scaling AI may be energy, not model ideas — Naveen Rao says Unconventional AI is rebuilding compute from first principles because token demand has exploded in the last six months, and lowering energy cost per token by multiple orders of magnitude is now existential for continued scaling.
Recursive self-improvement no longer sounds sci-fi to frontier builders — Jack Clark reportedly put a 60% probability on fully automated AI R&D by the end of 2028, and Naveen says he’d take the under because the gap from today’s AI-assisted research loops to models training their successors feels like roughly a one-year engineering problem.
Software development is already drifting toward an AI-majority workflow — Trey says his 70-person engineering team is seeing 70–80%+ of software work touched by AI now, making Dario Amodei’s claim that 90% of software development could be done by AI by 2027 feel aggressive but no longer crazy.
The biggest software opportunity in healthcare may be rebuilding records for AI, not humans — Both guests argue Epic’s EMR was designed for human consumption and insurance workflows, while the next winner could be an AI-native layer that prioritizes relevant patient context, ambient notes, and dynamic retrieval instead of brittle tabs and forms.
Anthropic’s culture became the episode’s wildest debate — Jason’s closing take is that Anthropic’s safety posture is not just PR but a sincere worldview, describing the company as a kind of AI priesthood that genuinely believes it is protecting humanity from superintelligence.
The Breakdown
AI as the “final boss,” plus two very different builders
Jason opens in full-apocalypse mode: AI may be “the last technological innovation,” the culmination of semiconductors, the internet, and decades of computing. He brings on Naveen Rao of Unconventional AI, who says the real wall for AI isn’t imagination but energy, and Trey Holterman of Tennr, who’s building the healthcare plumbing that gets patients from diagnosis to actual treatment.
Why tokens, not demos, are the new constraint
The conversation quickly zooms from ChatGPT hype to what Jason calls the current “race for tokens.” Naveen says his whole mission is driving down compute cost by orders of magnitude because AI demand is outrunning supply, and the trio frames the stack as a multi-front war: better models, faster transport, and radically cheaper hardware all compounding at once.
Healthcare’s black hole, and the new AI front door
Trey explains Tennr’s blunt thesis: half of patients fall into a “US medical black hole” and never make it to specialists, therapies, or devices because the system is so broken. He says Tennr now touches over 90% of US counties, 300,000 providers, and more than 20% of US healthcare, while also admitting the front door is changing fast as patients increasingly arrive armed with Perplexity, ChatGPT, and Gemini.
AI beats ER doctors — but only if it gets the right context
Jason brings up a study using 76 real ER cases where OpenAI’s o1 outperformed physicians on diagnosis and reasoning, then asks the obvious question: why isn’t every ER doctor already using a copilot? Trey and Naveen agree the upside is huge, but Trey adds the key caveat — these models shine when they get complete context, while real ER decisions often happen with messy, missing information and subtle human signals like body language.
Recording the visit, ambient scribes, and who owns your medical story
Jason goes into full quantified-self mode, talking about Whoop, Eight Sleep, Function Health, and even carrying a Plaud Pin recorder so patients can reclaim their own medical data trail. Trey notes doctors are already using ambient tools like Abridge, Suki, and Microsoft to turn conversations into EMR notes, and the bigger shift is obvious: patients and clinicians are both becoming data producers while the old records system struggles to keep up.
Why Epic is hated, and how an AI-native records layer could beat it
On Epic, Trey carefully says the company absorbs lots of frustration because it sits at the center of a broken healthcare system, while the uglier incentives may live deeper in insurers, PBMs, and specialty pharmacy. Naveen’s sharper point is that EMRs were built for human reading and payer workflows, not for LLMs — so the disruption path is to ingest all that messy history, prioritize what’s relevant, and turn it into AI-ready context rather than forcing doctors to click through ancient tabs.
Recursive self-improvement stops sounding theoretical
The show turns philosophical fast when Jason cites Anthropic co-founder Jack Clark saying AI systems are about to start building themselves, with a 60% chance of fully automated AI R&D by 2028. Naveen says he’d actually bet sooner, because today’s workflow already looks like “tell the model in English what I want, let it gather data and run loops,” and the missing steps to self-improving systems feel surprisingly small.
Capex, hyperscalers, and the Anthropic “cult” argument
They close on Brian Nowak’s forecast of hyperscaler capex rising from $805 billion in 2026 to $1.1 trillion in 2027, with Naveen arguing this is bigger than ROI math — it may be a reshaping of the whole economy, where machine-made utility goes nearly free and humans increasingly sell taste, empathy, and vibe. Then Jason unloads his spiciest theory: Anthropic’s safety framing isn’t just savvy messaging but a genuine belief system, with employees cast as AI monks, Night’s Watch defenders, and protectors against “AI Jesus,” while Naveen mostly agrees that, yes, many of them really do seem to believe it.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
