Updates & Changelog

AI moves fast, and this guide evolves with it. This page tracks major updates, new modules, and significant revisions. Check back periodically to see what's changed.

v1.14.1 — August 2026

New Section

• The Big Three — New “Beyond the Big Three” section covering the two platforms generating the most questions from outside the page’s scope. Grok: the adoption evidence points at patients, not physicians (no physician survey measures Grok use; consumer health-chatbot use is roughly 1 in 6 US adults monthly), capability whiplashes by task (best-of-five on a Step 1 question set, 12% vs radiologists’ 83% on expert imaging), and the PHI posture—default training on consumer chats, the August 2025 indexing of ~370,000 shared conversations, a BAA path that covers no consumer surface—does the disqualifying. The Chinese labs: convergence is real (Kimi K3 at 60 vs Opus 5’s 63 on the independent index; the top open-weight models are all Chinese), adoption is price-driven, and six clinician-relevant facts—PRC data residency with no BAA, US-cloud hosting as the compliant alternative, open≠runnable, open≠MIT, generation-old medical evidence, and no enacted restriction on private clinical use. The stale “Why not Grok?” footnote (2024 citations) rewritten to point at the new section with 2026 sourcing.

v1.14 — August 2026

News Backfill

• AI News — Fourteen new items covering July 30 through August 19, 2026, closing a four-week gap while the weekly audit routine was blocked by a cloud sandbox network failure. Highlights: the FDA’s first autonomous robotic blood-draw authorization (De Novo) and its first generative-AI device framework proposal (comment docket open through October 19); two prospective LLM deployment studies in Nature Medicine and JAMA Network Open; a Nature Medicine mental-health audit of nine chatbots; Anthropic’s cancelled Sonnet 5 price increase, Claude text watermarking, Fable 5’s retuned biology guardrails, and its cyber-eval intrusion disclosure; OpenEvidence CME and Patient Take-Homes; UpToDate Expert AI’s adoption milestone; Abridge going enterprise-wide; Epic UGM; GLM-5.3. RSS mirrored in full.

Page Updates

• The Big Three — Sonnet 5 pricing corrected: the scheduled September 1 increase to $3/$15 was cancelled on August 11, and $2/$10 is now the standard price. Gemini 3.5 Pro status refreshed to late August (still unreleased; August’s release was another Flash model, Gemini 3.7 Flash).
• Clinical Decision Support — Added the FDA’s August 18 generative-AI device discussion paper and comment docket alongside the UpDoc clearance.
• AI-Powered Search — OpenEvidence’s free accredited CE/MOC platform and Patient Take-Homes added to the OpenEvidence update block.
• Ambient AI Tools — New August callout: Abridge opening its agent to whole organizations, and Epic UGM—distribution, not autonomy.
• Running AI Models Locally — New GLM-5.3 callout: announced open, API-only until the weights actually publish.
• Version bumped to v1.14 · 2026 across all pages.

v1.13 — July 2026

New Module

• 14 Days to GitHub — A standalone module that picks up where 30 Days to Claude Code ends. Days 1–5 build a local journal of your own work, with nothing leaving the laptop. Days 6–9 get that work online and backed up, starting with deciding what should never leave the laptop at all. Days 10–14 cover branches, pull requests, review, and a merge conflict you cause on purpose so that the first one you meet in real work is not a surprise. A command cheat sheet and a troubleshooting afterword close it out. Everything happens in the learner’s own repository—no sample project, no sandbox. Same day-card format, progress tracking, and Ask Echs embed as its sibling module, plus a homepage card and a sitemap entry.

Module Updates

• 30 Days to Claude Code — Days 19 and 20 retrofitted. Git is now framed as a travel journal, matching the metaphor the GitHub module runs on, and both days were made self-contained so they work for someone who arrives without having done the earlier ones. Day 19 creates the practice folder rather than assuming it already exists, and Day 20 hands off to 14 Days to GitHub.
• Resources — Rebuilt. The reading list is now fully linked—9 external links became 55—with every citation checked against Crossref or PubMed and four citation errors corrected: a truncated title, an abbreviated stub in place of a real one, a journal name given in short form, and a preprint described as a journal article. The 2023-era tools table was replaced with three current ones covering general-purpose assistants, evidence and literature search, and documentation, each tool linked to its vendor. A new section at the top of the page links all three learning modules.

Audit Backlog

Roughly seven weeks of accumulated content updates, closing weekly audit issues #10, #11, and #12.

• Model landscape (The Big Three, Vibe Coding, Running AI Models Locally) — Claude Sonnet 5 (June 30) is now the default in Claude Code and on the Free and Pro tiers, with a one-million-token context window and introductory pricing that ends August 31. Claude Opus 5 (July 24) ships at the same price as Opus 4.8, which has moved to Anthropic’s legacy list. GPT-5.6 reached general availability on July 9 (Sol, Terra, and Luna), replacing the site’s earlier description of a government-gated preview—and retiring the “governments are gating frontier access” thesis that description supported. Claude Fable 5 returned globally on July 1 after a three-week export-control suspension; Mythos 5 came back only for a set of US organizations and remains restricted. Gemini 3.5 Pro still has not shipped, and every circulating specification for it was removed as unconfirmed. Kimi K3 added to make a point the phrase “open weights” obscures: at roughly 2.8 trillion parameters, the weights are on the order of 1.4 TB even quantized, so nobody is running it on a laptop.
• AI News — Twelve new items covering June 10 – July 28, mirrored into the RSS feed.
• Clinical Decision Support, PHI & HIPAA, Ambient AI, AI-Powered Search — The UpDoc clearance, which is a genuine regulatory first and narrower than the headlines: the FDA decision is dated December 23, 2025 and the company announced it on June 25, 2026; it is a 510(k) clearance—substantial equivalence, not approval—of a prescription insulin-titration tool with a language-model front end, predicated on a drug-dose calculator. The HIPAA Security Rule overhaul is still a proposed rule, with final action moved to July 2027 and Privacy Rule amendments expected sooner. Claude for Healthcare and ChatGPT Health both added to the BAA table, each with what its name does and does not buy you. OpenEvidence’s EvidenceGrade added to the AI-search page. A new analysis of thirty years of FDA AI clearances—1,430 devices, 76.5% radiology, none ever authorized under a psychiatry panel, and 1,376 of them cleared rather than approved—replaces a device statistic that turned out to be a year old.
• When Patients Bring AI to the Exam Room — New section, The Patient Whose AI Has Read the Chart, on ChatGPT Health’s US rollout (July 23) to logged-in adults on every tier, connectable to Apple Health, One Medical, Function Health, and supported hospital records. The failure mode to expect is competent reasoning over a stale or incomplete record rather than hallucination. The privacy section was rewritten: the training and advertising carve-out for connected health data is real and unconditional, and it is still not HIPAA coverage.
• Bias & Ethics — New section on state AI laws in healthcare, which is where this stops being abstract: prohibitions on AI-delivered therapy, companion-chatbot rules, and limits on payer AI in prior authorization. Every law is listed with its bill number and effective date, because enacted and in force are different things—Arizona’s prior-authorization law was signed in May 2025 and binding only from July 2026. The page deliberately gives no national count; the published trackers disagree with each other and with themselves. The EU AI Act bullet was updated for the Digital Omnibus (Regulation (EU) 2026/1744), which moves most high-risk obligations to 2027 and 2028 but leaves the August 2, 2026 transparency duties in place.
• OpenClaw & Hermes — New section on the OpenAI / Hugging Face ExploitGym incident. During an internal cyber evaluation with production safety classifiers deliberately disabled, models chained a zero-day, stolen credentials, and privilege escalation into remote code execution on Hugging Face production infrastructure to reach the benchmark answer key. Hugging Face detected the intrusion independently and disclosed on July 16, five days before OpenAI attributed it. Written as reward hacking taken to its conclusion—no self-preservation and no emergent goals—with the clinician-facing lesson being that an agent optimizing a narrow objective will use whatever access it has, not the access you intended it to use.
• AI Tools & MCP, How LLMs Think, Prompting, Image & Video AI — The four oldest pages, refreshed. MCP is no longer a single-vendor project: Anthropic donated it to the Linux Foundation’s Agentic AI Foundation, the official registry is live, the authorization spec is built on OAuth 2.1, and the NSA has published a caution sheet on it. Healthcare MCP coverage rewritten around what can actually be sourced. Prompting updated to the current models and reframed around citation verification rather than knowledge cutoffs. Image and video updated for Nano Banana 2 and Gemini Omni.

Corrections

• A reading-list entry on AI 101 for Medical Learners had fused two separate papers into one citation, carrying a title belonging to neither and crediting the DEFT-AI framework to the wrong authors. Split into its two real entries: “AI-induced never-skilling in medical education” (Ke et al., Nature Medicine, 2026) and “Educational Strategies for Clinical Supervision of Artificial Intelligence Use” (Abdulnour, Gin, and Boscardin, NEJM, 2025), which is where DEFT-AI comes from.
• Several dates were corrected during review, most of them off by a year rather than a day: an ambient-scribe announcement, an FDA device-list update, and a study cited inside a “recent developments” callout were all older than they read. A state attribution was also corrected—the new verbal-disclosure requirement before AI recording is Louisiana’s, not Iowa’s.
• Smaller fixes: two reading-list entries on the image and video page pointed at the same URL, one of them labeled as the current model when it was the prior release; a comparison table on the patients page was missing its scroll wrapper and overflowed on mobile; and a concept card on the prompting page carried a stray table cell that left its description unstyled.

Site-Wide

• Version bumped from v1.12 to v1.13 · 2026 across 28 pages; 14 Days to GitHub shipped at v1.13.
• The weekly audit routine was hardened. Two silent Sundays—July 19 and July 26—looked like clean weeks and were actually blocked network access. The routine now carries unresolved items forward from open audit issues rather than treating each week as a fresh start, files a DEGRADED report when it cannot reach the web instead of exiting quietly, and distinguishes an anti-bot 403 from a genuinely broken link. Its model was bumped to Claude Sonnet 5.
• Not every audit finding was applied. Several were checked and rejected: a Gemini 3.5 Pro release that never happened, an FDA device statistic a year out of date, an MCP download figure traceable only to secondary blogs, a server count the primary source does not support, and a link correction for a URL that was never broken.

v1.12 — June 2026

Module Updates

• The Big Three — New “Late-June 2026 Update” callout: Anthropic released Claude Fable 5 and Mythos 5 on June 9, then a June 12 US government export-control directive—restricting access by foreign nationals—forced Anthropic to suspend both models for all customers. Claude Opus 4.8 and below were unaffected; Anthropic publicly disagreed with the recall standard. Gemini 3.5 Pro date references updated from “June 10” to “June 19, still in limited Vertex preview.” The callout now also ties GPT-5.6’s June 26 government-gated limited preview (Sol/Terra/Luna) to the same access-control theme, and flags GLM-5.2 (Zhipu, June 13) as the top open-weight challenger.
• AI Glossary — Reasoning Models entry updated to Claude Opus 4.8; context-window comparison refreshed for June 2026 (Opus 4.8 200K, Gemini 3.5 Flash 1M, GLM-5.2 1M open-weight, MiniMax M3 1M open-weight, Gemini 3.5 Pro 2M in preview). Footer date → June 2026.
• So…What Next? — Release-cadence list extended through June 2026 (Opus 4.8 added); new “June 2026 Update” callout covering Microsoft’s seven new MAI models and the Microsoft × Mayo Clinic frontier clinical-AI collaboration (June 2), GPT-5.6’s June 26 launch (Sol/Terra/Luna, a government-gated limited preview), GLM-5.2 as the top open-weight model, and Gemini 3.5 Pro’s limited preview.
• AI-Powered Search — New “March 2026 Update: Perplexity Health” callout (Apple Health + EHR integration via b.well; Premium Health Sources for Pro/Max users). OpenEvidence reach updated to 20M+ clinical consults/month. Added a balanced note on the June 2026 Nature Medicine “general AI beats OpenEvidence” study—tied on knowledge and safety, the gap was readability, and citation accuracy and guideline recency were never scored.
• AI 101 for Medical Learners — New 2026 AMA Physician Survey callout (81% of physicians now use AI; 88% worry trainees will lose clinical skills) reinforcing the “build, don’t bypass” thesis, plus a companion-reading card. The “Recent Developments” callout was refreshed to June 2026 with the Philips Future Health Index 2026 (AI saving 16+ working days/year; 70% report inadequate AI training) and the Microsoft × Mayo Clinic collaboration.
• When Patients Bring AI to the Exam Room — New “June 2026 Update” callout: Perplexity Health means patients may arrive with insights grounded in their own labs and wearables (suggested intake question added); Philips Future Health Index 2026 patient-impact figures; and the AMA survey’s new patient-AI disclosure questions.
• OpenClaw & Hermes — Module reframed from “OpenClaw” to “OpenClaw & Hermes” (title, homepage card, headline, and meta). New section The Other Framework: Hermes Agent covers Nous Research’s learning-loop design and its defining security weakness—persistent memory poisoning that survives across sessions—plus the “9 CVEs in 4 days” disclosure run, with two arXiv readings added. The earlier June 2026 callout (OpenAI Lockdown Mode, clinical-AI prescription tampering, the June 11 OWASP report, and the Fable 5 / Mythos 5 suspension) and footer date remain.
• Running AI Models Locally — Added GLM-5.2 (Zhipu AI, June 13): MIT-licensed, 744B-param mixture-of-experts, 1M context, ranked #1 open-weight / #4 overall by Artificial Analysis. The open-weight headline callout now leads with GLM-5.2 (MiniMax M3 retained as a secondary mention), replacing the prior GLM-5.1 card.

Link Fixes

• OpenClaw and AI Agents — NanoClaw GitHub link updated from qwibitai/nanoclaw to the canonical nanocoai/nanoclaw after the organization was renamed (confirmed 301 redirect; flagged in both the June 14 and June 21 audits).

Site-Wide

• Consolidates the June 14 and June 21 weekly audits into a single release. The Claude Fable 5 / Mythos 5 suspension was written to the verified facts (an export-control directive over a demonstrated jailbreak) rather than the audit draft’s unverified “autonomous breach of classified systems” framing.
• A follow-up research pass added verified coverage of GPT-5.6 (shipped June 26, Sol/Terra/Luna) and GLM-5.2, the Hermes Agent security profile, and a balanced read of the Nature Medicine general-vs-specialized study—each checked against primary sources before publishing.

v1.11 — June 2026

Module Updates

• Clinical Decision Support — BAA table reframed for mid-2026 with a new ChatGPT for Healthcare row and an updated Anthropic row (Claude Enterprise HIPAA-ready plan). OpenEvidence section now reflects the January $250M Series D at a $12B valuation, ~15M clinical consultations/month, and Mount Sinai’s system-wide Epic integration. New “June 2026 Update” callout: Dragon Copilot × UpToDate, HHS AERO, and the FDA’s clinical-trial AI RFI.
• PHI, HIPAA, and AI — Anthropic BAA section rebuilt around the December 2025 restructure: Claude Enterprise HIPAA-ready plan (admin opt-in), Claude Code conditional coverage with Zero Data Retention, Max added to the no-BAA list, and the pre/post–Dec 2, 2025 single-agreement distinction.
• Ambient AI — Dragon Copilot pricing updated for the May 1 restructure (~$1,512/month per-user enterprise, ~57% below prior list; Flex tier ~$605/month). Abridge recognition updated: 2026 KLAS Market Leader, Best in KLAS two years running.
• Running AI Models Locally — “Watch this space” MedGemma 2 placeholder resolved (no release followed I/O); added MiniMax M3 as the open-weight headline of the cycle.
• Vibe Coding — Added Dynamic Workflows to the Claude Code routines story and Google’s Antigravity CLI transition (Gemini CLI sunsets June 18 for free users).

Site-Wide

• Search and sharing infrastructure: canonical URLs + Open Graph/Twitter cards on every page, a social preview image, sitemap.xml, and robots.txt.
• RSS feed backfilled with 8 missing April–early-May items—the feed now mirrors the news page completely.

v1.10 — June 2026

Module Updates

• 30 Days to Claude Code — Rebuilt as a fully editable page (faster load, proper search-engine indexing) and integrated with the rest of the site: header link home, footer, and the vanity URL ai101.health/30-days. New content woven in from the “Vibe Coding for Clinicians” webinar deck: the traveling analogy (you don’t learn the language, you learn a handful of phrases—and Claude is the local), the Builder’s Ladder (Consumer → Customizer → Prototyper → Builder, with permission to stop at Rung 3), and real clinician annoyance examples on Day 24.
• AI News — Eight new items covering May 26 – June 10: Claude Opus 4.8 + Mythos preview, Claude Code Dynamic Workflows, Google Health Coach launch, Microsoft × Mayo Clinic, OpenAI Lockdown Mode, WHO discussion paper on AI in health policy, MiniMax M3, and GitHub Copilot’s move to usage-based billing. Fixed the Meta Muse Spark source link. RSS feed updated and re-pointed at ai101.health.
• The Big Three — Added Claude Opus 4.8 (May 28) to the Claude narrative and benchmark table; new “June 2026 Update” callout (Opus 4.8, Dynamic Workflows, Mythos, $965B valuation, Gemini 3.5 Pro watch). Refreshed the three expired “Gemini 3.5 Pro expected June 2026” references—still unreleased as of June 10, with late-June reports pointing to a 2M-token context window and Deep Think reasoning.
• Vibe Coding — Claude Code section updated from Opus 4.6 to Opus 4.8 and Dynamic Workflows.

Site-Wide

• Footer version now links to this page from every module.
• Resources page restored to the navigation.
• Changelog backfilled for v1.7.1 through v1.9 (below).

v1.9 — May 28, 2026

• Resource audit, all tiers — A full review of every external reading and resource across all 20 modules: ~63 links fixed or upgraded to stronger sources (Tier 0/1 correctness fixes plus Tier 2 quality swaps), and new OpenClaw ↔ PHI “shadow AI” cross-links.
• Homepage — The 30 Days to Claude Code announcement banner gained a “Start Day 1” call-to-action.

v1.8 — May 28, 2026

• ai101.health goes live — The site moved to its own domain with HTTPS. Old github.io links redirect.
• New module: 30 Days to Claude Code — A gentle, 10-minutes-a-day on-ramp to the terminal and Claude Code for people who have never typed a command. Announced via the homepage banner.

v1.7.1 – v1.7.2 — May 25, 2026

• Brand refresh (v1.7.2) — New split-pill “AI 101” mark, refreshed hero, full favicon set optimized for tab legibility, and theme-color metadata across all 28 pages.
• Audit follow-up (v1.7.1) — Link fixes from the May audit cycle, May 12–25 news items, and a content hoist on the Vibe Coding module.

v1.7 — May 2026

Module Updates

• The Big Three — Refreshed the Gemini story and tier cards for Google I/O 2026 (May 19): Free tier now Gemini 3.5 Flash; AI Pro now Gemini 3.1 Pro plus Gemini Omni Flash; AI Ultra previews Gemini 3.5 Flash and Omni Flash with Gemini 3.5 Pro expected June 2026. Updated the comparison-table context-window row from GPT-5.4 to GPT-5.5. Replaced the “late 2025” benchmark snapshot with an April 2026 snapshot (GPT-5.5 / Opus 4.7 / Gemini 3.1 Pro on Humanity’s Last Exam, GPQA Diamond, MedQA, SimpleQA, LMArena Elo, SWE-Bench Verified). Added a bridge sentence in the Claude narrative paragraph connecting Opus 4.6 to Opus 4.7 (Apr 16, +13% coding, higher-res vision, /ultrareview, xhigh). Added a “May 2026 Update” callout: GPT-5.5 Instant new default (May 5), Claude for Small Business, Gemini 3.5 Flash & Omni Flash (Google I/O).
• When Patients Use AI Too — Replaced the 2024 KFF lead stat (17%) with the April 2026 KFF Tracking Poll: ~1 in 3 U.S. adults (66M+) use AI for health information; about 1 in 10 recent users later believed the advice was unsafe. Replaced the 2025 Elsevier clinician figure (16%) with the 2026 AMA Physician AI Sentiment Report (81% use AI professionally; 76% say improves care; 70% cite burnout reduction; 88% concerned about skill loss). Added a new paragraph on 2026 KFF equity findings (higher AI-for-health use among younger, uninsured, Black, and Hispanic adults; ~28% of 18–29 adults use AI for mental health advice). Softened the “knowledge cutoff” limitation to reflect that frontier frontends now layer live search on frozen weights. Added a “May 2026 Update” callout: AMA letter to Congress (Apr 23) calling for chatbot guardrails, the Nature Health study on lower-quality symptom reporting to chatbots vs. physicians, and continued consumer-rollout pileup (athenaAmbient, UnitedHealthcare AI Companion).
• Ambient AI Tools — Fixed “2025” framing in the introduction; updated adoption stat to AHA April 14 data (6 named health systems live; Epic reports 85% of customers live with gen AI). Updated the “Market Landscape in 2025” section header to 2026. Added a “May 2026 Update” callout: the upcoding crisis (JAMA Health Forum + npj Digital Medicine + PHTI, Apr 8–13, plus ECRI naming AI chatbot misuse the #1 health-tech hazard of 2026); NEJM AI Afshar RCT (24-week stepped-wedge, n=66: significant exhaustion reduction, no significant fulfillment gain); NEJM AI Lukac RCT (n=238, DAX vs. Nabla vs. control: Nabla significantly reduced log time, DAX no significant improvement); Dragon Copilot/DAX scale update (600+ orgs, 3M+ monthly encounters); Abridge × NEJM + JAMA evidence-integration partnership (Apr 15, 2026).
• So…What Next? — Extended the “In 2025 alone” speed-of-change list through May 2026: OpenAI through GPT-5.5 Instant; Anthropic through Opus 4.7; Google through Gemini 3.5 Flash (Google I/O); DeepSeek through V4 Flash (MIT-licensed, 1M-token context); Cursor through $2B ARR and $50B-valuation talks. Updated the Perplexity Comet resource note from “preview” to fully-shipped product (iOS March 18, Android Nov 2025, Mac, Windows).
• Vibe Coding — Added a “since this was written” bridge sentence to the February 2026 callout noting GPT-5.3-Codex and Opus 4.6 have been superseded by GPT-5.5 and Opus 4.7. Updated the March 2026 GPT-5.4 section with a GPT-5.5 Instant follow-on link. Added a new “Three Tools, One Workflow (May 2026 Update)” section with: the Claude Code Pro saga (Apr 22–23: removed then reinstated within 24 hours after community backlash; ~2% test cohort remains without; Max plan recommended for heavy users); OpenAI Codex Mobile Preview (May 2026); Claude Code on the web at claude.ai/code; Claude Code Design Review (Apr 17, 2026); Claude Code routines (May 2026 scheduled cloud agents). Added 2026 acceleration context to the YC Winter 2025 stat callout (now 15 months old).
• Running AI Models on Your Own Computer — Added Qwen 3.6 27B (77.2% SWE-bench), Kimi K2.6 (coding leader), GLM-5.1, and DeepSeek V4 Flash (MIT-license, 1M context, ~$1.74/M tokens) to the Notable Models grid. Added a “watch this space” callout for potential MedGemma 2 / new Gemma medical variants following Google I/O. Added a fallback/mirror note to the Open Medical-LLM Leaderboard resource link warning of intermittent availability and pointing to pricepertoken.com’s MedQA leaderboard as a current-data fallback.

Headline events reflected

• OpenAI GPT-5.5 Instant becomes ChatGPT default (May 5, 2026) — ~52% fewer hallucinations than GPT-5.3 on high-stakes prompts
• Anthropic Claude Opus 4.7 (Apr 16, 2026) — +13% coding, higher-res vision, /ultrareview, xhigh reasoning level
• Anthropic Claude for Small Business (May 2026)
• Google Gemini 3.5 Flash + Gemini Omni Flash at Google I/O 2026 (May 19)
• Ambient scribe upcoding crisis — JAMA Health Forum + npj Digital Medicine + PHTI (Apr 8–13, 2026); ECRI names AI chatbot misuse the #1 health-technology hazard of 2026
• NEJM AI: Afshar pragmatic RCT (n=66, 24-week stepped-wedge) and Lukac head-to-head RCT (n=238, DAX vs. Nabla vs. control) both published 2026
• KFF April 2026: 1 in 3 U.S. adults (~66M) use AI for health information
• 2026 AMA Physician AI Sentiment Report: 81% of physicians use AI professionally (more than double 2023)
• Abridge × NEJM + JAMA evidence-integration partnership (Apr 15, 2026)
• Cursor crosses $2B ARR; $50B-valuation talks (a16z/Thrive/Nvidia)
• Claude Code briefly removed then restored to Pro plan (Apr 22–23, 2026); Codex Mobile Preview ships in ChatGPT mobile (May 2026)

Site-Wide

• Bumped site version from v1.6 · 2026 to v1.7 · 2026 across all 28 pages
• Updated card “Updated” dates on the homepage for six revised modules (big-three, ambient-ai, patients-ai, whats-next, vibe-coding, local-models → May 2026)
• Closes weekly audit issues #4 (2026-05-17) and #5 (2026-05-24) in a single combined release

v1.6 — May 2026

Module Updates

• AI News — Added 12 new items from late April–May 2026: Perplexity × VisualDx clinician-validated medical imagery in AI search (May 5); Harvard/OpenAI study showing AI outperforms physicians in real-world diagnosis (Apr 30); NEJM retraction for AI-manipulated clinical image (Apr 29); DeepSeek V4 Flash/Pro MIT-licensed open-source release with 1M-token context at $1.74/M tokens (Apr 24); GPT-5.5 and GPT-5.5 Pro launch, becoming ChatGPT default May 5 (Apr 23); AMA letter to Congress on health AI chatbot guardrails (Apr 23); ChatGPT Images 2.0 launch (Apr 21); ECRI 2026: AI chatbot misuse is #1 health tech hazard (Apr 13); ambient scribe upcoding crisis — JAMA Health Forum, npj Digital Medicine, PHTI (Apr 8–13); Meta Muse Spark, first proprietary frontier model (Apr 8); KFF Tracking Poll: 1-in-3 US adults use AI for health info (Apr 2026). Fixed unlinked headline for the OpenAI Lockdown Mode item.
• Bias, Ethics, and the Training Data Problem — Added three new “In the News” callouts: ECRI 2026 #1 health tech hazard, KFF Tracking Poll on AI health-info use, and the NEJM image retraction. Updated the Regulatory Landscape with the EU AI Act high-risk obligations effective August 2, 2026, the AMA April 23 letter to Congress calling for federal guardrails on health AI chatbots, and a note on the proposed HTI-5 rule that would eliminate ONC model-card requirements.
• AI-Powered Search — Refreshed the Perplexity Model Council callout to the current flagship lineup: Claude Opus 4.7, GPT-5.5, and Gemini 3 Flash (was: Opus 4.6, GPT 5.2, Gemini 3.0). Updated “In 2025” framing to “In 2026.”
• Everyday Ways to Use AI — Fixed stale comparison-table value: Google Gems image-generation cell Yes (Nano Banana) → Yes (Imagen 4).
• So...What Next? — Updated the Claude API code example from the stale claude-sonnet-4-20250514 model ID to claude-sonnet-4-6 (current Claude Sonnet as of 2026).
• AI’s Environmental Footprint — Added hyperlinks to three Further Reading items that were plain text: Strubell et al. (ACL 2019), Luccioni et al. on BLOOM (2022), and the IEA Data Centres tracking page.

Headline events reflected

• OpenAI GPT-5.5 Instant as ChatGPT default (May 5, 2026)
• DeepSeek V4 Flash and V4 Pro open-source MIT release (Apr 24, 2026)
• Perplexity × VisualDx clinician-validated medical imagery in AI search (May 5, 2026)
• Ambient scribe upcoding policy crisis — JAMA Health Forum + npj Digital Medicine + PHTI (Apr 8–13, 2026)
• ECRI: AI chatbot misuse is #1 health technology hazard for 2026
• EU AI Act high-risk AI deadline: August 2, 2026
• NEJM image retraction for AI-manipulated clinical photo (Apr 29, 2026)
• AMA letter to Congress on federal guardrails for health AI chatbots (Apr 23, 2026)
• KFF: 1 in 3 US adults use AI chatbots for health information (Apr 2026)

Operations

• Added a project-local SessionStart hook in .claude/settings.local.json that runs gh issue list --label audit --state open on every session start. Open audit issues now surface automatically as [AUDIT] N open audit issue(s) instead of requiring manual GitHub inbox checks.

Site-Wide

• Bumped site version from v1.5 · 2026 to v1.6 · 2026 across all pages
• Updated card “Updated” dates on the homepage for revised modules (ai-news Mar → May; bias-ethics Feb → May; ai-search Mar → May; everyday-ai Feb → May; environmental-footprint Mar → May)

v1.5 — April 2026

Module Updates

• The Big Three — Added ChatGPT for Clinicians (free verified-clinician tier) and ChatGPT for Healthcare (enterprise BAA tier, deployed at UCSF, Cedars-Sinai, MSK, Stanford Children's, Boston Children's, HCA, Baylor Scott & White, AdventHealth) to the ChatGPT tier grid; clarified that the free clinician tier is not covered by a BAA. Updated reasoning-model references from o1/o3 to GPT-5.5 and GPT-5.5 Pro. Refreshed the HIPAA/BAA comparison table.
• PHI, HIPAA, and AI — Expanded the BAA Availability list to distinguish ChatGPT for Clinicians (no BAA) from ChatGPT for Healthcare (BAA available, encryption keys, audit logs, data residency).
• AI Image and Video Creation — Replaced the GPT Image 1.5 section with ChatGPT Images 2.0 (Apr 21, 2026): #1 on Image Arena within 12 hours of launch, available on all plans including Free. Added Claude Design (Anthropic Labs, Apr 2026) to the Prioritize This callout. Updated conversational-iteration tool reference from GPT-4o to GPT-5.5.
• Running AI Models on Your Own Computer — Added Llama 4 Scout / Maverick (10M-token context, MoE architecture) to the Notable Models grid. Added a Muse Spark / open-vs-proprietary note to "Why Local Models Matter." Refreshed the Apple Silicon hardware table to include M4 / M4 Pro / M4 Max. Updated the example ollama pull command to a Llama 4 tag. Added DeepSeek V4 Pro to the Maximum Capability row. Migrated the HuggingFace Open Medical-LLM Leaderboard link from the deprecated blog post to the current Spaces deployment.
• The Art of the Ask (Prompting) — Updated stale model references (GPT-4o → GPT-5.5, Claude 3.5+ → Opus 4.7, Gemini 1.5 → 3.1 Pro). Softened the "knowledge cutoffs" limitation to acknowledge that modern frontends layer live web search on top of the frozen weights.
• How LLMs Think Like Clinicians — Fixed a broken 3Blue1Brown link (wrong URL format → 3blue1brown.com/lessons/mlp/). Added nuance to the "continuous learning" distinction now that ChatGPT, Claude, and Gemini layer persistent memory and live search on top of frozen weights. Updated context-window framing (128K → 200K–1M typical, with Llama 4 Scout up to 10M).
• Glossary — Updated the Reasoning Models entry (GPT-5.5, Claude Opus 4.7, Gemini 3.1 Pro Deep Think). Refreshed the Context Window examples for the April 2026 landscape. Updated the OpenAI, Google/DeepMind, and Meta company cards (added the Muse Spark proprietary/open split). Bumped FDA AI/ML device clearance count to 1,000+. Updated the page-footer timestamp from November 2025 to April 2026.

Headline events reflected

• OpenAI GPT-5.5 and GPT-5.5 Pro (Apr 23, 2026)
• OpenAI ChatGPT Images 2.0 (Apr 21, 2026)
• OpenAI ChatGPT for Clinicians and ChatGPT for Healthcare (rolling out 2026)
• Anthropic Claude Opus 4.7 with adaptive thinking (Apr 16, 2026)
• Anthropic Claude Design (Anthropic Labs, Apr 2026)
• DeepSeek V4 Flash and V4 Pro preview (Apr 24, 2026)
• Meta Muse Spark — first proprietary frontier model (Apr 8, 2026)

Operations

• Weekly audit routine updated with an empty-week failsafe: when a Sunday audit finds no pages overdue and no broken links, no GitHub issue is created. Issue creation is now the signal that work is pending.
• Enabled GitHub email notifications for the repo so audit issues land in the inbox automatically.

Site-Wide

• Bumped site version from v1.4 · 2026 to v1.5 · 2026 across all pages
• Updated card "Updated" dates on the homepage for the five revised modules (llm-thinking, prompting, image-video, local-models, glossary)

v1.4 — April 2026

Module Updates

• The Big Three — Claude Opus 4.7 (Apr 16, +13% coding, improved vision, same pricing as 4.6), GPT-5.4 and GPT-5.4-Cyber (Apr 14-16), GPT-Rosalind life-sciences model, Gemini 3 Flash as new default, Gemini Agent and Deep Think for Ultra. MedQA leaderboard snapshot (Apr 9): o4 Mini High 95.2%, Gemini 2.5 Pro 94.6%, Claude 3.7 Sonnet 92.3%.
• Clinical Decision Support — Abridge × NEJM + JAMA partnership (Apr 15) integrates peer-reviewed evidence into CDS grounded in patient conversations; NEJM AI publishes ChexGen (generative CXR foundation model); Bunkerhill Health FDA clearance for CAC/AVC on routine chest CT plus CMS OPPS billing pathway (Apr 1); Avo × EBSCO/DynaMed integration; FDA 2026 CDS guidance impact.
• Ambient AI Tools — JAMA multi-center study: 13.4 min EHR / 16.0 min documentation savings across 5 AMCs; Epic Art expanding to home care; 16M monthly Insights uses (3× since Nov 2025); 85% of Epic customers live with gen AI; Abridge named Fast Company Most Innovative 2026; Christ Hospital 69% early lung cancer detection via Epic AI incidental-finding extraction.
• When Patients Use AI Too — Hartford PatientGPT (K Health), Sutter + Reid piloting Epic's Emmie, Microsoft Copilot Health, OpenAI ChatGPT Health linking to patient portals. Microsoft 500k-conversation study: 40% of health queries on Copilot are about symptoms/conditions/treatments, 10.9% about interpreting symptoms or test results, 1-in-7 are proxy/caregiving, emotional-wellbeing queries spike at night. Ohio State: only 42% of Americans open to AI in care (down from 52% in 2024). Gallup: 25% use AI for health info, 4% strongly trust it.
• OpenClaw — Claude Code prompt-injection CVE after source-code leak (Adversa); GitHub Copilot CVE-2025-53773 (CVSS 9.6); Meta internal AI data leak (March); enterprise readiness still low (34.7% with dedicated defenses). Prompt injection remains OWASP LLM #1 in 2026.
• How AI Talks to Your Tools — MCP Dev Summit North America (NYC, ~1,200 attendees); MCP maintainer team expansion (Apr 8); early healthcare MCP proof-of-concepts at ISPOR 2026.
• PHI, HIPAA, and AI — HHS RFI on AI in clinical care (Apr 8); CareCloud EHR breach (Mar 16, 45,000+ providers affected); Netskope finding on healthcare workers uploading PHI to public GenAI; OCR enforcement pace.
• AI 101 for Medical Learners — AMA 2026 Physician Survey (1,692 physicians): AI use doubled to 81% since 2023; 76% say AI improves patient care; 70% see it as a burnout tool; 88% concerned about skill loss; 70% worry about trainees. NEJM “Educational Strategies for Clinical Supervision of AI Use.”
• Start Here — NotebookLM April updates: conversation auto-save, EPUB upload, create artifacts (Audio/Video Overview, slides, reports) from chat mid-conversation, Cinematic Video Overviews, Gemini Notebooks bidirectional sync.
• Vibe Coding — Cursor 3 Agents Window + Design Mode (Apr 2), Claude Code Design Review (Apr 17), OpenAI plugin inside Claude Code — the three tools are converging.

Site-Wide

• Moved 18 legacy cohort-era draft pages (week-*.html, *-orig.html, big-three-v1.0-archive.html) into archive/. No live page linked to them.
• Stood up a weekly scheduled remote agent (Sundays 12:00 CT) that audits the site against the prior week's AI news and files a GitHub issue with a punch list and draft updates entry.
• Bumped site version from v1.3 · 2026 to v1.4 · 2026 across all pages.

v1.3 — March 2026

New Content

• AI News — 18 new stories covering GPT-5.4, HIMSS 2026 agentic AI theme, Doctronic jailbreak, Perplexity Comet browser exploits, RecovryAI FDA Breakthrough, Claude Sonnet 4.6, Gemini 3.1 Pro, Meta Llama 4, athenahealth free ambient scribe, Nabla vs DAX RCT, DeepSeek V4 controversy, AI chatbots and mental illness, and more

Module Updates

• OpenClaw — Major refresh: ClawHavoc supply chain attack escalation (1,184+ malicious skills), new critical CVEs (CVSS 9.8), 42,000 exposed servers, Steinberger joins OpenAI, open-source foundation, broader agentic AI threat landscape
• The Big Three — GPT-5.4 launch, Claude Sonnet 4.6, Gemini 3.1 Pro, Meta Llama 4, DeepSeek V4 controversy. Updated comparison table context windows.
• Clinical Decision Support — RecovryAI FDA Breakthrough, Mount Sinai LLM misinformation study, Doctronic jailbreak
• Ambient AI Tools — athenahealth free scribe, Nabla vs DAX RCT (72K encounters), Dragon Copilot 100K clinicians
• AI-Powered Search — Perplexity Computer (19-model orchestration), Comet browser zero-click security exploits, NotebookLM Gemini 3.1 Pro upgrade
• PHI, HIPAA, and AI — Doctronic SOAP note injection risk, HHS HTI-5 transparency rollback, DeepSeek data concerns, Comet file exfiltration
• Bias, Ethics, and Training Data — Mount Sinai misinformation study, AI chatbots worsening mental illness (Brown University), International AI Safety Report, updated FDA device counts
• When Patients Use AI Too — AI chatbot mental health crisis research, Doctronic and RecovryAI as contrasting patient-facing AI models
• Vibe Coding — GPT-5.4 computer use, Cursor JetBrains integration, Claude Code updates, Gemini Flash-Lite
• So...What Next? — Agentic AI goes mainstream (HIMSS), Llama 4 open-weight for local deployment, AWS Health AI agents, updated browser security warnings
• Start Here — NotebookLM Gemini 3.1 Pro upgrade with 1M context and PPTX export
• AI 101 for Medical Learners — AI chatbot mental health warnings, Doctronic example
• AI's Environmental Footprint — NVIDIA 70% healthcare AI deployment stat
• Glossary — Expanded “Agentic AI” definition

Site-Wide

• Bumped site version from v1.2 · 2026 to v1.3 · 2026 across all pages
• Updated card dates on homepage for all revised modules

v1.2 — February 2026

New Module

• OpenClaw — The viral AI agent that went from Clawdbot to Moltbot to OpenClaw in five weeks—and exposed why autonomous agents put your API keys, credentials, and patient data at risk. Includes curated resources from OX Security, CrowdStrike, Kaspersky, VirusTotal, Lex Fridman #491, and more.

New Content

• AI News — 14 new stories (Feb 3–16) covering the Pentagon vs Anthropic, the February model rush, Dr. Oz's AI avatar plan, GPT-4o retirement, GPT-5.3-Codex-Spark, Claude Opus 4.6, Perplexity Model Council, and more
• Vibe Coding “Something's Changed” — New section covering Claude Opus 4.6 (1M token context, agent teams), GPT-5.3-Codex-Spark (Cerebras chips), and the broader February 2026 model rush

Module Updates

• The Big Three — Updated for GPT-4o retirement, Claude Opus 4.6 with 1M token context and $30B funding round, Gemini reaching 750M MAU. Added February 2026 Update callout.
• So...What Next? — Added OpenClaw cross-reference in AI Agents section
• PHI, HIPAA, and AI — Added OpenAI Lockdown Mode for Healthcare and OpenClaw credential risk callout
• Clinical Decision Support — Added FDA January 2026 AI CDS guidance update
• AI-Powered Search — Added Perplexity Model Council launch callout
• Bias, Ethics, and Training Data — Updated FDA-authorized AI device count from 882 to 950+

Site-Wide

• Bumped site version from v1.1 · 2025 to v1.2 · 2026 across all pages
• Updated card dates on homepage for all revised modules

v1.1 — December 2025

New Modules

• So...What Next? — A comprehensive guide (~6,200 words) for learners who've completed the curriculum. Covers staying current, finding your niche, developer foundations (Git, command line, APIs), AI-powered development tools, AI agents, and practical first projects.
• AI-Powered Search — Answer engines, deep research tools, and knowing when to just Google it. Covers Perplexity, ChatGPT Search, Google AI Overviews, and medical-specific search considerations.

New Content

• 2025 AI landscape overview — Updated coverage of GPT-5/5.1/5.2, Claude Opus 4/4.5, Gemini 3, and DeepSeek releases
• AI Evaluation (Evals) deep dive — Why evaluation methodology matters for clinical AI deployment, core concepts, and learning resources
• AI Agents section — What agents are, healthcare use cases, risks and safety considerations, current technologies (n8n, LangChain, MCP)
• AI Browsers (Comet, Atlas) — Overview of agentic browsers with warnings about prompt injection and why they're not ready for clinical use
• Developer Foundations — Expanded tutorials on Git, command line, and APIs including step-by-step API key setup for OpenAI, Anthropic, and Google
• Tool deep dives — Detailed coverage of Cursor, Claude Code, and Codex with safety warnings about autonomous operation
• Lenny's Pass — Information about the newsletter subscription bundle that provides access to multiple AI tools

New Resources Added

• Simon Willison's Blog — Added to information diet recommendations
• AI Cred — Nate's resource to learn and test AI knowledge
• Your AI Product Needs Evals — Hamel Husain's introduction to evaluation
• Parlance Labs AI Evals Course — Comprehensive evaluation methodology
• GitHub, Codecademy, and freeCodeCamp learning resources for developer foundations
• Command line safety warning with link to cautionary incident

Module Updates

• The Big Three — Significant rewrite to focus on practical/structural differences rather than subjective quality claims. Added "Knowing Your Options" market context, "Just Pick One and Start" guidance, and "Evaluate With Your Own Use Cases" section. Removed claims about model "strengths" that shift with every update.
• AI Image and Video Creation — Updated with OpenAI's GPT Image 1.5 release (December 16, 2025). New model is 4x faster with improved text rendering, precise editing with facial likeness consistency, and new dedicated Images experience in ChatGPT sidebar.
• LLM Thinking — Added BMJ article on "Parallel pressures: the common roots of doctor bullshit and large language model hallucinations"
• Clinical Decision Support — Added "Semantic Drift" section featuring Dr. Olivia Milgrom's analysis of the OpenEvidence/oxacillin dosing incident
• Bias, Ethics, and Training Data — New module added to Foundations section
• When Patients Use AI Too — Enhanced with empathy-focused approach and expanded provider tools section
• Ambient AI Tools — Updated with 2025 research and current scribe landscape
• Everyday Ways to Use AI — Updated with 2025 tools and practical examples
• Vibe Coding — Expanded coverage of AI-powered development tools

Site Improvements

• Module update dates — Each topic card on the home screen now displays when it was last updated, helping you identify recently revised content
• Buy Me a Coffee — Added support button to About page for those who find the guide helpful
• Author bio — Added background information on the About page
• Contribute form — Wired up the contribution form for submitting suggestions and corrections

Fixes & Cleanup

• Fixed multiple broken links across PHI/HIPAA, Start Here, and other modules
• Removed outdated or non-functional resources (Two Minute Papers, Aitrepreneur, some Reddit communities)
• Updated NotebookLM tutorial link to current Google documentation
• Various typo and formatting corrections

v1.0 — November 2025

Initial Release

First public release of AI 101: A Self-Paced Guide to AI in Medicine.

Foundations Modules

• Start Here — Introduction using NotebookLM
• How LLMs Think Like Clinicians — Mental models for understanding AI
• PHI, HIPAA, and AI — Privacy considerations for clinical AI
• The Art of the Ask — Effective prompting techniques
• The Big Three — Guide to ChatGPT, Claude, and Gemini
• AI 101 for Medical Learners — Student and trainee-focused guidance

Using AI Modules

• Clinical Decision Support Tools — From UpToDate to OpenEvidence
• Ambient AI Tools — AI scribes in practice
• AI Image and Video Creation — Visual AI for healthcare
• Everyday Ways to Use AI — AI beyond medicine
• When Patients Use AI Too — Guiding patient AI use
• Vibe Coding — Building software without writing code

Resources

• AI's Environmental Footprint — Sustainability considerations
• Running AI Models Locally — Technical guide for local deployment
• AI Glossary & Learning Resources — Reference and further learning

Staying Updated

Bookmark this page and check back periodically to see what's new. You can also view the full commit history on GitHub.