When I research something, I turn the sources into a short spoken overview with Google Gemini, so it can be heard on a walk. Gemini gets things wrong, so every briefing here was transcribed on my Mac and its claims tested against primary sources. Where the audio is wrong, the correction sits right under it, with the source. 3 hours 40 minutes in all; 53 corrections. Each one plays here, with no account, and has a plain-text transcript.
Explains why hip and knee surgery is shifting from hospitals to lower-cost outpatient surgery centers, and how text-message surveys, phone-camera therapy and physician sign-off records are being used to keep patients on track at home.
What the audio gets wrong (3)
WRONG “Penn Medicine voluntarily took a 46 to 50 percent cut on its facility fees”. The Sept 24, 2026 Penn/IBX/Regent deal plans 18+ new ASCs that will be paid standard ASC rates; 46% is MedPAC's ASC-vs-hospital-outpatient payment gap and 'up to 50%' was IBX's estimate of savings, not a fee cut Penn took Source
WRONG “72% of orthopedic patients get zero physical therapy after surgery”. The ~72% figure comes from a study of meniscectomy only; across common orthopedic procedures about 56% get no PT within 180 days (44% use it, 62% after TKA) Source
NOT FOUND “AI billing audits recover a specific sum in unbilled revenue per surgeon per year”. No published source; this is an internal model estimate, not a measured result
Health data design for Medicare's 2026 surgical rules
Explains how new Medicare models and short-period remote monitoring codes push surgical practices toward record systems built on the FHIR data standard, with AI-drafted letters that a licensed clinician must review.
What the audio gets wrong (2)
WRONG “Providers can bill the quoted amount for 2 to 15 days of remote monitoring data”. After CMS revised practice expense values, the 2026 national payment for CPT 98985 is higher than the amount quoted. Source
WRONG “Calls the CMS chronic-care model the 'XSS model'”. It is the ACCESS model Source
Medicare's mandatory joint replacement bundle and surgeon-built tools
Explains the CJR-X mandatory 90-day bundled payment for hip and knee replacement starting in 2028, why documenting patient complexity matters, and how practices use AI to follow patients after surgery.
What the audio gets wrong (1)
NOT FOUND “CJR-X applies to 2,217 hospitals”. CMS projects more than 2,500 hospitals in the finalized CJR-X (start Jan 1, 2028; 2% discount; +/-20% risk cap for most hospitals) Source
New Medicare outcome-reporting rules for joint care
Describes how Medicare models and reporting rules tie orthopedic payment to patient-reported outcome surveys, and the remote check-in services built to meet them.
What the audio gets wrong (3)
WRONG “Under the TEAM model, failing to report outcomes for 50% of surgeries brings a 25% payment reduction”. The 50% THA/TKA PRO-PM requirement is in the Hospital IQR program; missing it cuts one-quarter of the hospital's annual payment update (about 2-4%), not 25% of payment Source
WRONG “Under ACCESS a surgeon could do a flawless hip replacement and lose the withheld money”. ACCESS's MSK track pays for managing chronic musculoskeletal pain, not surgical episodes Source
WRONG “Closing line: 'the software monitoring you for the next 8,760'”. Garbled; 8,760 is the number of hours in a year.
Medicare's proposed 2027 payment rules and surgeon pay
Walks through the proposed 2027 Medicare rules: lower surgeon fees, half payment for a visit and procedure on the same day, outpatient surgery growth, a new payment category for AI software, and in-house staffing for remote monitoring.
What the audio gets wrong (2)
WRONG “A proposed 7% cut to facility-based orthopedic payments”. CMS projects about 7% for orthopedic surgery overall: about 8% in facility settings and 5% in non-facility settings Source
WRONG “Practices are now mandated to hire W-2 employees for remote monitoring”. This is a proposal in the CY 2027 PFS proposed rule (direct employees of the billing practice), not a mandate in force; the final rule was pending Source
Shorter antibiotic courses after surgery for bone infection
Explains the SOLARIO trial: when antibiotics are placed at the infection during surgery, about a week of antibiotics by pill or IV worked as well as four or more weeks and caused fewer side effects.
What the audio gets wrong (3)
NOT FOUND “CMS rolled out new reimbursement upgrades for this in October 2026”. No CMS payment change tied to SOLARIO was found
NOT FOUND “Cutting 2.9 bed days saves a large sum per patient”. Not in the trial summaries; the trial reports shorter stays but no such figure Source
NOT FOUND “The data does not apply to prosthetic joint infections”. The SOLARIO protocol lists prosthetic joint infection among eligible infections; check the paper's subgroup results before stating this Source
How new Medicare payment models tie pay to outcomes
Explains Medicare's shift from paying per service to paying for results, through the LEAD ACO model, the ACCESS chronic-care model, the MAHA ELEVATE grants and the mandatory TEAM surgical episodes. Shows why precise measures such as muscle mass now matter for payment.
What the audio gets wrong (2)
NOT FOUND “An 'ERA framework' inside these models creates shadow bundles for falls prevention”. No CMS model or framework by this name was found. LEAD does include episode-based risk arrangements for partnering with specialists. Source
NOT FOUND “ACCESS withholds exactly 50% of payment until improvement is proven”. ACCESS outcome-aligned payments carry a Clinical Outcome Adjustment, and part of the payment is reconciled on outcomes. The exact 50% figure appears only in secondary sources, not in CMS documents. Source
Medicare's TEAM model and post-surgery recovery tools
Explains how Medicare's TEAM bundled-payment model makes hospitals accountable for 30 days after joint replacement and other surgeries, and how browser-based motion tracking, rules-based coding review and physician sign-off tools could support recovery and compliance. Ends with the role of home caregivers in preventing falls after discharge.
What the audio gets wrong (4)
WRONG “Starting in 2026, hospitals carry full downside risk for 30 days post-discharge”. TEAM began January 1, 2026, but every hospital defaults to Track 1 in year 1, which has no downside risk (up to three years for safety-net hospitals). Source
WRONG “Hospitals deploy JointCoach to track recovery inside the 30-day post-surgical window”. JointCoach is PT-owned remote therapeutic monitoring outside the surgeon's global period, separate from JointClass patient education.
WRONG “co-op.care's caregiver wage, and that care is funded by pre-tax HSA checks verified in the browser”. co-op.care publishes its actual caregiver wage (W-2, plus ownership) and the family rate on its own site. A physician signs or declines each letter case by case; HSA/FSA eligibility is never an automatic browser check. Source
WRONG “California's AB 1979 made autonomous clinical AI flat-out illegal”. AB 1979 was signed September 30, 2026 and takes effect January 1, 2027; it requires clinician judgment and bars AI from independently performing licensed clinical functions. Source
7 briefings
AI drafts, a licensed human signs
Clinical documentation written by AI, checked and signed or declined by a licensed person, and who carries the liability.
AI research tools, fake citations and doctor sign-off
Covers how AI tools that summarize medical literature also produce fabricated citations, and why a record of exactly what a physician changed in an AI draft matters more than a bare signature.
What the audio gets wrong (1)
WRONG “The 2027 Medicare physician fee schedule proposes a yearly payment for a full year of chronic musculoskeletal care”. The year-long chronic MSK payment is in CMMI's voluntary ACCESS model (not the fee schedule), with 50% of the Medicare share withheld pending outcomes Source
Big tech companies compete to prepare your doctor visits
Surveys how Microsoft, Amazon and others are building AI layers that read your records before a visit, and how a 2026 FDA guidance change affects software that gives a single recommendation.
What the audio gets wrong (2)
NOT FOUND “Claude hit 98.1% accuracy on complex medical calculators”. Published MedCalc-Bench results for Claude models are roughly 48-63% (up to ~80% on a later dataset version); no 98.1% figure found Source
NOT FOUND “ChatGPT 5.2 hit 100% diagnostic accuracy using conservative heuristics”. Published GPT-5.2 diagnostic results are about 83-91%; no 100% result found Source
Explains why AI that aces exams still stumbles in real conversations, and how a physician-review service records each sign-off, pays the same for a yes or a no, and slips test cases into the queue to catch rubber-stamping.
What the audio gets wrong (1)
NOT FOUND “GPT-4 scored 8.2 on diagnostic benchmarks and dropped to 0.49 on conversational dialogue”. The CRAFT-MD study (Nature Medicine 2024) did find GPT-4 accuracy fell sharply from vignettes to conversations, but these exact figures were not confirmed; '8.2' is not a valid accuracy score (likely meant 0.82) Source
How Hippocratic AI builds healthcare-only AI agents
Explains Hippocratic AI's design of one conversational model checked by many specialist models, its non-diagnostic follow-up calls, and the tension between augmenting and replacing clinical staff.
Checked: nothing found wrong in the claims tested.
How physicians review and sign AI-drafted medical documents
Explains how a medical records backend can be kept separate from a physician review step, where a licensed physician signs, edits, or declines each AI draft. It also covers how a hashed audit trail proves the review happened without exposing the chart.
Checked: nothing found wrong in the claims tested.
Local AI, wearable sensors, and physician-reviewed remote monitoring
Describes running AI on a local machine instead of the cloud, a sensor sleeve that tracks muscle imbalance after knee surgery, and a licensed physician reviewing any change in care.
What the audio gets wrong (3)
WRONG “Remote therapeutic monitoring pays the quoted amount per patient per month”. 2026 Medicare national rates for the common pair 98977 + 98980 come in below the range quoted; higher totals need extra time codes. Source
WRONG “RTM revenue is routed through co-op.care to its W-2 caregivers”. Remote therapeutic monitoring sits with JointCoach, owned by the physical therapist and outside the surgeon's global period; co-op.care's wage and rate are for home care, not RTM. Source
NOT FOUND “'Stride Orthopedics' smart sleeve and 'ARX base station' as existing products”. Not verified as existing products
Proving a physician actually reviewed AI-drafted paperwork
Explains a design for showing that a physician genuinely reviewed AI-drafted clinical documents: a minimum review time, engagement checks and a tamper-evident log of each sign-off. Discusses why that proof matters as AI makes paperwork nearly effortless to produce.
What the audio gets wrong (4)
WRONG “The Colorado AI Act takes effect in June 2026”. SB 24-205 never took effect. It was repealed and replaced by SB 26-189 in May 2026, effective January 1, 2027, without the duty-of-care requirement. Source
WRONG “Under FDA's peptide reclassification, patients will need a physician-attested letter of medical necessity to get BPC-157”. FDA removed BPC-157 from Category 2 in April 2026, and an advisory committee voted 8-6 in July 2026 to recommend it for the 503A bulks list; no final rule yet. Compounding requires a prescription; a letter of medical necessity is an IRS/HSA document unrelated to FDA status. Source
WRONG “SolvingHealth's co-op.care brand gives families physician-attested care plans that unlock pre-tax funds”. co-op.care is its own filed Colorado cooperative, not a SolvingHealth brand. A physician signs or declines each letter case by case; HSA/FSA eligibility is never automatic. Source
NOT FOUND “The system saves physicians about 2.3 hours daily”.
8 briefings
Home care, family caregiving and the end of life
Worker-owned home care, the unpaid family caregiver, aging in place, and preparing before a crisis.
Reviews studies showing that text-message links get far more patients, including older patients, to answer health surveys than portals with logins, and why that matters for catching problems early.
Checked: nothing found wrong in the claims tested.
Explains the paperwork military survivors face across federal agencies, why storing records in advance helps, how a funeral advocate who works without commission can save money, and the survivor benefit rules still being debated.
Checked: nothing found wrong in the claims tested.
Explains why a home death without enough support can be hard on families, how little in-person help routine hospice provides, how early conversations reduce unwanted treatment, and why hospice use differs across communities.
What the audio gets wrong (1)
WRONG “Jimmy Carter is 'doing it beautifully at home' (present tense)”. Jimmy Carter died December 29, 2024, after about 22 months of home hospice Source
Explains how a 2026 wearable chip runs AI on the device for private voice notes and early fall or decline signals, and how satellite messaging and floor-level location help in emergencies.
Checked: nothing found wrong in the claims tested.
Covers how families can talk with aging parents about what matters most before an emergency, and why a living will alone may not stop emergency crews (a signed POLST or DNR kept on the fridge does). Also covers vetting home health agencies early and spotting hospice enrollment scams.
What the audio gets wrong (1)
WRONG “An example from the HBO show 'The Pit'”. The show is 'The Pitt' (HBO Max).
AI, neighborhood care co-ops and home robots as nursing home alternatives
Walks through a three-part proposal: a health savings account intake flow with clinician review, live-in neighborhood caregiver cooperatives coordinated by software, and household robots built on vision-language-action models. Also touches on the tax and labor-law structures such a model would need.
What the audio gets wrong (3)
WRONG “The user vaults a card and starts immediately; a reviewing clinician signs off later”. no instant letter of medical necessity at checkout. A licensed physician reviews each case and signs or declines, and HSA/FSA eligibility is decided case by case, never in advance.
WRONG “The caregiver wage it quotes”. co-op.care publishes its actual caregiver wage (W-2, plus ownership) and the family rate on its own site. Source
WRONG “Pi Zero FAST prunes the model's neural pathways to cut inference latency to 30-60 ms”. pi0-FAST is an autoregressive version of pi0 that uses FAST, a frequency-space (DCT plus byte-pair encoding) action tokenizer. It is not a pruning method, and it is autoregressive, unlike the flow-matching pi0. Source
Shows how many adults give unpaid care without calling it caregiving, that many of them are men and younger adults, and how heavily it weighs on people caring for children and parents at once. Offers practical coping steps and suggests describing caregiving by its intensity instead of calling it a burden.
What the audio gets wrong (1)
WRONG “Research from healthcare entrepreneur 'Alexander Drain'”. The name is Alexandra Drane, co-founder and CEO of ARCHANGELS. Source
Medicaid work rules, and the software that decides who keeps coverage
From 2027 many Medicaid adults must show 80 hours a month of work or qualifying activity. This briefing explains the medically frail exemption and argues that the forms and software people must get through decide, in practice, who keeps coverage.
What the audio gets wrong (4)
PARTLY “Independent models like those from Manit Health estimate about 8.2 million people disenrolled a year”. The firm is Manatt Health; its model gives an average annual decline of 8.2 million, fiscal years 2027-2034. It is a projection, not a count. Source
PARTLY “By January 2028 states will stop accepting self-attestation that a person is medically frail”. From 2028 a person may still self-attest once per enrollment period; at redetermination the state must check reliable documentation. Source
WRONG “Getting the exemption relies on medfrail.com”. Your state Medicaid agency decides the exemption. medfrail.com is an independent guide that helps a person keep their own record; it is not part of any state system. Source
NOT FOUND “medfrail.com fills official state forms in the browser with PDF-lib and hands appeals to legal aid through DocAssemble”. Neither is built on medfrail.com. These were proposals in an AI-written audit. Today the site keeps notes on the person's own device and offers a printable keep sheet. Source
16 briefings
Personal AI, memory and the wider world
Memory you own, how AI reasons, what cheap software does to work, and a few detours.
Presents research arguing that AI reasoning models work like an internal debate among several viewpoints, why we judge AI more harshly than people, and why single AI-writing detectors misfire at scale.
Checked: nothing found wrong in the claims tested.
Reading vs listening, and how personal AI memory works
Covers brain research showing reading and listening reach the same meaning yet listeners retain less, and how a personal memory tool on your own device structures and grounds what it captures.
What the audio gets wrong (1)
WRONG “The personal AI builds a hallucination-no-cost, flawless record of your life”. No memory system is hallucination-no-cost; grounding answers in source lines reduces but does not eliminate errors
Summarizes Jonathan Rauch's argument that liberal democracy and the institutions that settle facts are under attack, including his 2026 essay describing a 'dual state'.
Checked: nothing found wrong in the claims tested.
Building health software on the Aidbox FHIR platform
Explains how Aidbox stores health data in the FHIR standard, how it separates data for multi-clinic organizations, how it compares with Medplum, and how AI coding tools are used on top of it.
What the audio gets wrong (1)
NOT FOUND “PACE operators use Aidbox to meet 2026 Medicare mandates for CMS-0057-F prior authorization APIs”. CMS-0057-F's impacted payers are MA organizations, Medicaid/CHIP and exchange QHPs (PACE not listed), and its API requirements are due starting January 1, 2027 Source
Argues that while AI makes thinking work cheap, physical things like land, energy and human care stay costly, and explores what that means for healthcare and for the value of human labor.
What the audio gets wrong (1)
WRONG “The cost of intelligence is dropping by about 100 times per year”. The commonly cited estimate (Altman, Feb 2025) is about 10x per year, i.e. roughly 100x over two years Source
Genetic risk, daily habits, and 23andMe's nonprofit restart
Explains the difference between single-gene and many-gene disease risk, and why many-gene risk responds to environment and habits like clean air and movement. It also covers 23andMe's move to a nonprofit research institute.
Checked: nothing found wrong in the claims tested.
Discusses how falling AI costs could shift care from treating illness to preventing it, Medicare's anti-fraud algorithms, and the risks of opaque models, models trained on AI output, and power demand.
Checked: nothing found wrong in the claims tested.
Explains why checking code, not writing it, is now the bottleneck, and how separate AI reviewers, mixing large and small models, and parallel work branches are used to manage it.
Checked: nothing found wrong in the claims tested.
Explains how call analysis pulls useful data from conversations and how a low-delay voice connection lets an AI agent respond and handle interruptions naturally, with a healthcare scheduling example.
Checked: nothing found wrong in the claims tested.
AI-written software, patents, and the rise of care work
Explains why AI-generated code is hard to protect by copyright or trade secret, and why the value is shifting to how work is orchestrated and to buying outcomes rather than software seats. Argues that cheaper digital work makes in-person care work more valuable.
Checked: nothing found wrong in the claims tested.
Simulating crowds of AI agents to forecast reactions
Explains MiroFish, an open-source tool that builds a simulated society of AI agents with memories from news or documents, then watches how opinions spread when an event is introduced. Notes that a student built it quickly with AI coding tools.
What the audio gets wrong (2)
WRONG “The developer is 'Guohang Zhang', funded by 'Shonda Group'”. MiroFish was built by Guo Hangjiang, an undergraduate at Beijing University of Posts and Telecommunications, and incubated by Shanda Group (about 30 million yuan). Source
NOT FOUND “Spiking neural networks ('Brain and the Fish') make the agents physically unable to hallucinate”. MiroFish runs on the OASIS multi-agent framework with LLM agents. No published backtest shows its forecasts are accurate, and a WPP test found the swarm added no information. Source
Grit, root-cause analysis, and working alongside AI
Covers Angela Duckworth's idea that effort counts twice, the 'five whys' method for finding the real cause of a failure, directing AI coding tools with plain-language constraint files, and pausing before you send a heated reply.
Checked: nothing found wrong in the claims tested.
How attention residuals help deep AI models keep early information
Explains why standard residual connections can drown out early-layer information in deep language models, and how the Kimi team's attention residuals let each layer choose which earlier layers to draw from. Also covers the block variant that keeps memory costs manageable.
What the audio gets wrong (1)
NOT FOUND “DeepSeek's mHC needs a memory footprint 34 times the model dimension; AttnRes needs a '5.5b' footprint”. The AttnRes paper reports under 4% extra training cost and under 2% extra inference latency for Block AttnRes. The 34x and 5.5b figures could not be found and look garbled. Source
Intelligence as a society of thought, in brains and AI
Draws on Blaise Aguera y Arcas's book What Is Intelligence? and recent research to argue that life and mind are defined by function, not material. Covers research suggesting reasoning models improve by staging internal debates between perspectives, and why that argues for many AIs checking one another.
What the audio gets wrong (2)
NOT FOUND “Over 60% of the reasoning accuracy advantage comes from internal multi-agent debates”. The paper (Kim et al., arXiv 2601.10825) reports that reasoning models show more perspective diversity and that steering a conversational feature doubled Countdown accuracy. No '60%' figure was found. Source
NOT FOUND “A July 2026 internal audit of the BFF artificial-life experiments found random mutation 10 times more efficient”.
What matters online when AI makes software easy to build
Argues that as AI makes building apps nearly effortless, value shifts to proprietary context, trust and human judgment. Covers machine-readable checkout protocols, certification and insurance for AI agents, retailers blocking outside agents, and optimizing sites for AI readers.
What the audio gets wrong (1)
WRONG “Companies like ElevenLabs are underwriting agents using AIUC-1 certification”. ElevenLabs is the insured party: it obtained AIUC-1 certification and an AIUC-backed policy covering its voice agents. The Artificial Intelligence Underwriting Company (AIUC) does the certifying and underwriting. Source
Prompting a voice into existence: Google's new text-to-speech
How Google's Gemini 3.8 text-to-speech models let you describe a voice instead of picking one, direct it with tags kept separate from the script, and copy a real voice only with that person's recorded consent.
What the audio gets wrong (3)
PARTLY “Voice cloning is blocked in Texas, Illinois, the UK and the EEA”. Google AI Studio voice replication also excludes Switzerland and India. Source
NOT FOUND “Stateless replication keys last seven days with zero data retention, and listeners preferred the cheaper model in blind tests”. Not confirmed in the public Google sources checked; treat as unverified. Source
NOTE “The closing line repeats”. The last sentence of the generated audio loops several times; the transcript keeps it once.