Neural Mastery / Field notes
Clear, practical essays on AI releases, agent systems, and the ideas that make modern models work.
SERIES: How US Healthcare Works (Dedicated Course Hub ↗)
Who the players in US healthcare are, how many of each there are, and how money and information move between them when someone sees a doctor.
Latest essay · 7 min read
Why learning agentic engineering backwards breaks in production, what 2026 hiring data asks for, and the four disciplines to study in order.
Read the essayJeva is an open 4-billion-parameter model that returns a choice, a probability or a score instead of text. Here is the line of classifiers it grew out of, how its decision head works, and what its results do and don't show.
How a prior authorization request moves from provider to payer and back, the decision timelines in federal rules, what the 2024 CMS rule changes in 2026 and 2027, the electronic standards, and where reform and AI fit.
How payers try to keep members healthy and out of hospital: risk stratification, case and disease management, transitions after discharge, quality measures like HEDIS and Star Ratings, and the evidence on whether it saves money.
How a provider turns care into a claim: registration, documentation, charge capture and the chargemaster, coding, claim scrubbing, clearinghouses, acknowledgements and rejections, with one office visit followed end to end.
What happens inside a payer's system between receiving a claim and paying it: the adjudication pipeline step by step, coding edits, pricing, denials and their reason codes, pends, payment timing and corrected claims.
The checks a claim examiner works through on a pended claim, in order, from eligibility and provider status through duplicates, authorization, coding, coordination of benefits and pricing, to a documented decision, with four worked cases and the appeal paths that follow.
How payers find and recover wrong payments before and after they're made: improper payment rates, prepayment and post-payment review, clinical and coding validation, COB and subrogation recovery, fraud, waste and abuse, and risk adjustment audits.
The payment methods behind US healthcare: Medicare's physician fee schedule and RVUs, DRGs and APCs, commercial rates as a multiple of Medicare, capitation, value-based contracts and bundles, and how CMS pays Medicare Advantage plans.
How the drug benefit runs on its own rails: pharmacy benefit managers, formularies and tiers, real-time pharmacy claims, list prices and rebates, specialty drugs, the medical-versus-pharmacy split, and Medicare Part D's recent changes.
The transactions, privacy rules and interoperability standards that every part of US healthcare runs on, the data problems that cause most failures, and where AI helps, where it's risky, and what rules now govern it.
Premiums, deductibles, copays, coinsurance and the out-of-pocket maximum, worked through one family's year of claims, plus networks, surprise bills and HSAs.
ICD-10 diagnoses, CPT and HCPCS procedures, modifiers, DRGs, revenue codes, NDCs and the identifiers that turn a doctor's visit into data a payer can process.
How a plan decides whether a service is covered at all: Medicare's national and local coverage determinations, Medicare Advantage rules, commercial medical policies, and the criteria vendors behind them.
How members get into a payer's system, how a provider checks coverage before a visit, what goes wrong with retroactive changes, and how two plans decide which pays first.
How payers build provider networks: contracting and rates, credentialing and primary source verification, the provider data behind directories and claims, and the rules on network adequacy.
How a health plan's benefits, written in documents for people, become configuration a claim system can run: service categories, cost-share rules, limits and authorization flags, and where it goes wrong.
How payers review whether care is medically necessary before, during and after it happens: prospective, concurrent and retrospective review, inpatient versus observation, who makes the decision, and where review goes wrong.
Kimi K3.1 Pro is a 2.8-trillion-parameter open model in which three of every four attention layers keep a fixed-size memory instead of a growing KV cache. Here is how attention got there, step by step.
A login screen is safe after tapping Sign in and suspicious after tapping View order. A new paper checks each step of a phone or browser agent by predicting the next screen in embedding space, and catches hijacks a screen-only monitor provably cannot.
Agent Skills became a vendor-neutral standard on 22 October 2026. Here is what a skill is, how agents load them without filling their context, and where they fit next to MCP.
MemoryTax found that a strong model searching its raw history beat memory frameworks, and TurnCost found that an agent's bill grows with its turns, not with the size of tool output.
Who the players in US healthcare are, how many of each there are, and how money and information move between them when someone sees a doctor.
Medicare and its four parts, Medicaid and CHIP, employer plans, the ACA marketplaces and the military and veterans' systems: who each one covers, who pays for it, and how it reaches providers.
How vision models went from needing a label on every image to world models that plan by predicting in their own representation space, and what the new H-JEPA paper adds.