Wednesday, August 19, 2026
Vertical AI and the Clinical Frontier
August 19 · 18 videos
AI scales clinical intelligence.
Legacy stacks hinder agentic workflows.
Math yields to the machines.
Health tech goes native.
Execution beats dreaming.
“AI is the world's largest relationship therapist.”
How To Think Like The Top 1% | Sales Masterclass
Alex Hormozi · Alex Hormozi · 86 min
Watch on YouTube →Alex Hormozi explains why scaling a business requires counterintuitive solutions that often run contrary to basic human impulses. He argues that standard logic works for small problems but fails at the 100 million dollar scale.
- Scaling bottlenecks often occur where the logical solution is actually incorrect for the business.
- Hiring A-players at above-market rates provides 3 to 5 times the output of standard B-players.
- Adding friction to sales motions through qualifications and education filters for high lifetime value customers.
- Raising prices increases profit by boosting revenue per unit while simultaneously decreasing the cost basis.
- A refined sales motion reduces the specialized talent required to close deals by removing doubt early.
The summer Math fell to the machines...
Fireship · Fireship · 5 min
Watch on YouTube →AI has entered a hyper-acceleration phase in pure mathematics by solving more open conjectures in three weeks than humans did in a decade. This shift transforms mathematics from a cerebral human endeavor into an engineering problem of search and verification.
- AI models recently disproved an 87-year-old algebraic geometry problem known as the Jacobian conjecture.
- Anthropic's Claude Code increased Riemann Hypothesis solution satisfaction from 41 percent to 67 percent.
- The process involved executing over 2,400 shell commands and burning 31 million tokens for a single insight.
- Formal proof languages like Lean act as compilers that allow machines to verify mathematical correctness without human review.
- Amateurs are now using LLMs to solve problems that previously required decades of specialized academic study.
Why Your Enterprise Tech Stack Isn’t Ready for AI Agents — Christopher Lovejoy & Saul Howard
Christopher Lovejoy · AI Engineer · 19 min
Watch on YouTube →Christopher Lovejoy and Saul Howard discuss why existing enterprise tech stacks are unprepared for AI agents in regulated industries. They argue that compliance and security must be architectural primitives rather than features added after a proof of concept.
- Most enterprise AI projects fail when moving from a demo to production due to unaddressed security requirements.
- An effective audit trail must be a durable legal chain of evidence rather than a simple developer log.
- Immutable append-only event logs provide trivial auditability for complex agentic interactions.
- Separating sensitive patient data from orchestration logic enables zero-trust debugging in healthcare environments.
- Treating humans and LLMs as interchangeable agents creates a unified framework for evaluation and escalation.
DeepSeek Just Made Closed AI Look Ridiculous
Dr. Karoly Zsolnai-Feher · Two Minute Papers · 5 min
Watch on YouTube →DeepSeek V4 Pro demonstrates that open-weight models are reaching the reasoning quality of top-tier closed-source alternatives. The release highlights how open science enables rapid iteration cycles that challenge the dominance of closed labs.
- DeepSeek V4 Pro achieved a 78 percent speedup in generation through a new multi-token drafting technique.
- The model was created by distilling knowledge from over 10 specialist teacher models into a single student model.
- Open weights under the MIT license provide a market hedge against price increases from specific API providers.
- The time between the publication of AI research papers and their production implementation has collapsed to six weeks.
- Improved spatial logic is evidenced by the model's superior performance in solving Rubik's Cube structural problems.
Don’t be data poor — Anuj Iravane, Anterior
Anuj Iravane · AI Engineer · 16 min
Watch on YouTube →Anuj Iravane explains how to overcome data poverty in healthcare AI by using synthetic data generated from symbolic policy trees. This approach allows for rigorous evaluation without compromising sensitive patient information.
- Healthcare workflows require higher than 95 percent accuracy to be viable for administrative tasks.
- Synthetic medical records are generated to be correct by construction using reasoning traces from decision trees.
- Approximately 90 percent of the datasets used by Anterior are now synthetic to avoid privacy bottlenecks.
- Vertical AI success relies on a skills-based architecture that allows clinicians to manage the pipeline.
- Just-in-time dataset creation enables developers to test edge cases before actual customer data is received.
How to build an AI-Native Health Company — Dan Feng, Maven Clinic
Dan Feng · AI Engineer · 17 min
Watch on YouTube →Dan Feng explains how AI has inverted software economics by making implementation nearly instantaneous and planning the primary bottleneck. This shift requires health tech companies to move from rigid product documents to iterative briefs and short commitment cycles.
- Maven Clinic moved to a self-certification model for code reviews to handle a 10x increase in code volume.
- Engineers are encouraged to cap pull requests at 500 lines to maintain review quality and speed.
- The company uses Consensus Validation where multiple models check high-stakes tasks like financial claims.
- Planning cycles are split into directional one-year dreams and hard two-to-four-week sprint commitments.
- Reliability is managed by classifying failures into tolerable UI glitches and intolerable financial errors.
- Performance reviews now specifically reward employees for leveraging AI to multiply their individual impact.
Trading Desks to Clinical Trials: Parallels in Applied Vertical AI — Ayush Bhardwaj, Allos AI
Ayush Bhardwaj · AI Engineer · 20 min
Watch on YouTube →Ayush Bhardwaj discusses the judgment wall where engineers lack the domain intuition needed to build valuable AI for finance and pharma. He argues that proprietary domain expertise and expert-driven learning loops are the only sustainable moats in vertical AI.
- Approximately 30 percent of pharma firms withhold clinical trial data despite legal requirements to disclose it.
- The FDA had to publicly remind 2000 sponsors to release clinical trial results in 2026.
- Stanford AI Index data suggests that 89 percent of enterprise AI agents never reach production.
- Vertical AI projects often face a quiet death when they fail to provide immediate ROI for expert users.
- The framework shifts from Human in the Loop to AI in the Loop where experts are the primary actors.
- Proprietary data generated through expert-driven loops is more valuable than the underlying commodity models.
From Chrome DevTools to AI Engineering, with Addy Osmani
Addy Osmani · The Pragmatic Engineer · 92 min
Watch on YouTube →Former Google Director Addy Osmani explores the transition from manual coding to high-level system governance in the age of AI agents. He emphasizes the importance of maintaining human judgment and avoiding cognitive surrender to AI outputs.
- Addy Osmani managed an engineering organization of 50 people during his 14-year tenure at Google.
- Engineers should avoid cognitive surrender by prompting agents for their decision-making process rather than just code.
- The Software Factory model shifts the focus from manual prompting to building self-optimizing systems.
- Technical alpha is defined by the depth of an engineer's curiosity and their ability to exercise superior judgment.
- The unbundled career model suggests future engineers must integrate product sense and UX empathy into their roles.
- Accountability is maintained through an Owners File culture where humans remain responsible for agent-written code.
Healthcare’s Agent Bytecode: X12 as the Harness for AI Agents — Vasant Kearney, Onlay
Vasant Kearney · AI Engineer · 20 min
Watch on YouTube →Vasant Kearney proposes using X12 EDI standards as a strict syntax harness to manage non-deterministic AI agents in healthcare. This approach addresses the lack of ground truth in insurance data by restricting models within a rigid schema.
- X12 standards act as a foundational bytecode to prevent error propagation in long-chain reasoning tasks.
- Reasoning chains exceeding 50 steps are avoided due to cumulative errors and high compute costs.
- The system treats every agent action as an X12 transaction to maintain a normalized internal representation.
- Engineers are encouraged to be AI pilled regarding potential but AI skeptical regarding current model reliability.
- Standardizing on public schemas like X12 prevents engineering bloat and simplifies onboarding for new developers.
- The architecture moves enterprise memory to databases for better logical separation and system stability.
200 Million Patient Interactions Later — Vivek Muppalla, Hippocratic AI
Vivek Muppalla · AI Engineer · 20 min
Watch on YouTube →Vivek Muppalla describes how Hippocratic AI uses a constellation of 31 models to provide safe and empathetic clinical voice interactions. The goal is to solve the healthcare triage problem by creating clinical abundance through AI agents.
- The Polaris architecture runs 31 models simultaneously including one generalist and 30 specialist models.
- The system achieved a 99.89 percent safety rating compared to the human benchmark of 81 percent.
- A decoder-only audio LLM processes both verbal content and prosody to improve empathetic engagement.
- Hippocratic AI has processed over 200 million clinical interactions across more than 60 health systems.
- Safety is prioritized through an oath-based culture where clinical metrics outweigh traditional performance stats.
- The HEART benchmark is used to measure empathy as a core engineering requirement for patient adoption.
AI is the World’s largest Relationship Therapist: Clay Cockrell and Tony Fabrikant, CoupleWork AI
Clay Cockrell · AI Engineer · 16 min
Watch on YouTube →Clay Cockrell and Tony Fabrikant discuss how general purpose AI acts as a flawed relationship therapist by prioritizing engagement over clinical truth. They introduce Maxine, a specialized AI coach built on established clinical methods to provide timely interventions.
- General purpose LLMs often act as expensive mirrors that validate user biases instead of challenging them like real therapy.
- Couples wait an average of six years before seeking professional help, often reaching out only when the relationship is in crisis.
- The Maxine platform utilizes the Gottman Method and Emotionally Focused Therapy to identify clinical failure modes like escalating control.
- AI therapy tools should aim to work themselves out of a job rather than optimizing for high daily active usage or session length.
- Safety evaluations for clinical AI are treated as binary where a single failure to detect domestic violence signals is disqualifying.
Shipping AI to a Million Patients Without an A/B Test: Jared Joselowitz, Ufonia
Jared Joselowitz · AI Engineer · 19 min
Watch on YouTube →Jared Joselowitz explains why traditional A/B testing is unethical in healthcare and presents a simulation first safety framework for clinical voice AI. This approach uses synthetic patient personas to validate AI agents before they interact with real humans.
- Ufonia uses a PatBot LLM to simulate diverse patient personas ranging from verbose to laconic to test clinical hazards.
- The Matrix framework combines simulated dialogues with an automated evaluator called BevJudge that achieves a 0.96 F1 score.
- Genetic Pareto optimization is used to treat prompt refinement as a data science problem rather than manual engineering.
- The engineering focus shifts from shipping code to shipping a verifiable safety case that satisfies medical device regulations.
- Proactive simulation loops replace reactive dashboard monitoring when the cost of a single clinical failure is unacceptable.
Inside Whatnot's Bet on Live Shopping
Grant LaFontaine · a16z · 42 min
Watch on YouTube →Grant LaFontaine describes how Whatnot is transforming e-commerce from intent based search to discovery based entertainment. The platform leverages human connection and real time community to drive massive user engagement in the live commerce market.
- Whatnot users spend an average of 95 minutes daily on the platform even though 80 percent do not make a purchase.
- Live commerce is described as demand expansionary because it creates new market value through discovery rather than just shifting sales.
- The company dedicates 40 percent of its total headcount to trust and safety to maintain authenticity in the marketplace.
- Top sellers on the platform have scaled to generate over 100 million dollars in revenue with high profit margins.
- The business model mimics a physical shopping mall where social interaction is the primary driver of long term customer habits.
Guardrails First: Engineering Member-Facing Health AI: Rashi Agrawal, Hinge Health
Rashi Agrawal · AI Engineer · 21 min
Watch on YouTube →Rashi Agrawal argues that healthcare AI safety is an architectural challenge that must be addressed before model generation begins. She presents a framework for building member facing tools that prioritize clinical safety over raw performance.
- Current consumer health AI tools under triage life threatening emergencies 50 percent of the time according to Mount Sinai research.
- The proposed architecture strips personal health information at the pipeline boundary and uses deterministic code for critical decisions.
- Deterministic layers are placed above probabilistic models to handle irreversible actions like routing calls to emergency services.
- Engineers must define severity based on the worst case plausible outcome even if the probability of occurrence is extremely low.
- Fast follow features in regulated industries should be treated as committed technical debt rather than optional backlog items.
Stripe Just Bought Access to Every AI Model
Josh · Limitless Podcast · 24 min
Watch on YouTube →This episode explores Stripe's strategic acquisition of OpenRouter and the launch of Cursor's Origin platform for AI agents. It highlights a shift in the AI economy toward owning the middleware and infrastructure layers.
- Stripe's acquisition of OpenRouter allows it to capture a 5.5 percent take rate on AI token flow across 400 models.
- OpenRouter processes over 100 trillion tokens monthly providing Stripe with massive data on model performance and usage.
- Cursor's Origin platform is a high throughput code hosting alternative designed specifically for AI agents rather than humans.
- Origin supports superhuman coding speeds of 22 commits per second per repository to accommodate agentic workflows.
- The market is moving toward a routing war where infrastructure providers aim to become the essential tollbooths for AI traffic.
The Psychology of People Who Dream Big But Do Nothing
Rob Dial · The Mindset Mentor Podcast · 15 min
Watch on YouTube →Rob Dial explores the biological wiring that favors short term comfort over long term growth. He introduces the Opportunity vs. Obligation framework to help high potential individuals overcome stagnation.
- Identity drives all human action because people naturally behave in alignment with their internal self image.
- The Opportunity vs. Obligation framework explains how focusing on the immediate strain of a task leads to avoidance.
- Radical self reliance is necessary because no external force is coming to change an individual's life circumstances.
- Reframing the pain of cold calling as a path to family security allows sales professionals to overcome the fear of rejection.
- Choosing the harder path of discipline today creates an easier life in the future while avoiding immediate discomfort leads to long term hardship.
- Dial notes that focusing on the immediate physical strain of a workout is why many fail to lose weight or maintain mobility for future milestones.
From Ambient Documentation to Clinical Intelligence — Chaitanya Asawa, Abridge
Chaitanya Asawa · AI Engineer · 21 min
Watch on YouTube →Chaitanya Asawa discusses how Abridge uses AI to transform patient clinician conversations into structured clinical intelligence. The talk focuses on solving the administrative burden of medical documentation to reduce clinician burnout.
- Abridge treats the patient clinician conversation as the atomic unit of healthcare from which all billing and orders flow.
- The company scaled to 300 major health systems by solving for pajama time which is the two hours doctors spend on notes after work.
- Engineering efforts focus on the generator verifier gap where judging medical accuracy is as complex as generating the content.
- Abridge utilizes a four physician rubric process to create expert calibrated LLM judges for rigorous evaluation.
- The technical strategy involves processing 100 million conversations annually by post training smaller specialized models on proprietary datasets.
It took me 37 years to realize what I'll tell you in 18 minutes.
Alex Hormozi · Alex Hormozi · 18 min
Watch on YouTube →Alex Hormozi argues that maturity is defined by the transition from seeking optionality to embracing specific commitments. He presents a framework for managing life trade offs and optimizing for high impact activities.
- Adulthood is characterized by the elimination of alternatives while youth is defined by the pursuit of optionality.
- The Multi Box Strategy involves stacking activities to satisfy multiple life priorities like health and career simultaneously.
- Superior Game Theory suggests that success depends on choosing the right metrics to optimize such as active users over raw counts.
- Life is a zero sum game of time distributed across ten categories including health, romance, and career.
- Regrets are an inevitable part of life so individuals must consciously choose which trade offs they are willing to accept.
- Hormozi leverages his experience scaling Acquisition.com to over 250 million dollars in annual revenue to illustrate the necessity of picking a specific game to win.
References
PeopleAlex Hormozi · Henry Ford · Dan Kennedy · Rory Sutherland · Elon Musk · Terence Tao · Levent Alpoge · Jared Sumner · Dimitri Ryben · Christopher Lovejoy (x.com/ChrisLovejoy_) · Saul Howard (x.com/saulhoward) · Dr. Karoly Zsolnai-Feher · Anuj Iravane (x.com/anujiravane) · Dan Feng · Ayush Bhardwaj (x.com/aybh08) · Yann LeCun · Addy Osmani · Pavel Feldman · John Resig · Ryan Dahl · Paul Irish (x.com/paul_irish) · Vasant Kearney (x.com/vasantkearney) · Vivek Muppalla (x.com/vim1up) · Clay Cockrell (walkandtalk.com) · Tony Fabrikant · John Gottman · Sue Johnson · Jared Joselowitz (x.com/JaredJoselowitz) · Grant LaFontaine (x.com/GrantLaFontaine) · David George (x.com/DavidGeorge83) · Logan Head · Rashi Agrawal · Alex Atallah · Satya Nadella · Tony Robbins · Andrew Huberman · Chaitanya Asawa (x.com/c_asawa) · Genghis Khan · David Senra · Leila Hormozi
ToolsAcquisition.com · BANT · Claude Code (anthropic.com) · Lean · GitHub · DeepSeek V4 Pro · Anterior · Lambda · Maven Clinic · Allos AI · Chrome DevTools · Chromium · X12 EDI · Onlay · Hippocratic AI · Polaris · ChatGPT (openai.com) · Maxine · Dora · PatBot · Matrix Framework · BevJudge · Gemini 1.5 Pro (deepmind.google) · Whatnot (whatnot.com) · OpenRouter (openrouter.ai) · Cursor (cursor.com) · Origin · Abridge (abridge.com) · Gym Launch · Prestige Labs
Papers$100M Money Models · Jacobian conjecture · Riemann Hypothesis · SOC 2 · HITRUST · Stanford AI Index · Gottman Method · Emotionally Focused Therapy (EFT) · Jeppa (Genetic Pareto) · HIPAA · FDA guidelines