I think this is an important paper, and I want to spotlight why. Recursive self-improvement is shifting from a hypothetical, to an emerging reality. Anthropic is having Claude help them build the next versions of Claude. These models aren't fully autonomous yet, but we are sta…
Who's Who of AI
Scott McGrath
Biomedical Informatics PhD • CITRIS Health @UC Berkeley • FAMIA • Focusing on Informatics and AI in medicine • Linfield U. Grad • Missoula MT
https://smcgrath.phd
What they're sharing
When AI builds itself anthropic.com
Is AI ruining our skills? Early results are in — and they’re not good nature.com
US Law Enforcement Warns of ‘Anti-Tech Extremism’ as AI Hatred Grows | WIRED wired.com
The Tokenpocalypse Is Here: Companies Are Scrambling To Stop Spending So Much on AI 404media.co
Introducing Claude Sonnet 5 \ Anthropic anthropic.com
Meta Contractors Posed as Teens to Prompt Rival Chatbots About Suicide, Sex, and Drugs | WIRED wired.com
Claude Fable 5 and Claude Mythos 5 anthropic.com
Microsoft Hacked to Deliver Malware to Claude and Gemini Users 404media.co
Amazon Shuts Down Internal AI Leaderboard After Employees Cheated 404media.co
Opinion | What A.I. Did to My College Class nytimes.com
AI costs how much? GitHub Copilot users react to new usage-based pricing system. arstechnica.com
Scientists have a bad case of AI FOMO, Nature poll reveals nature.com
Articles & links
AI deskilling is a real concern: We've seen physicians' polyp detection rates drop from 28.4% to 22.4% when an AI assistant was turned off. This clinical deskilling matches an Anthropic trial where AI-assisted programmers scored 50% vs 67% on coding quizzes. 🧵1/5 #Medsky
US intelligence agencies are shifting surveillance toward a broad new domestic threat category labeled "anti-tech violent extremism." Over 1,000 pages of unpublished reports from the FBI and DHS indicate a coordinated effort to monitor data center protests and digital advocacy.
AI Weekly's analysis
→
- Wired reviewed over 1,000 pages of unpublished DHS, FBI, and fusion-center reports coalescing around a new label called 'anti-tech extremism.'
- A Delaware Valley fusion-center bulletin flagged AI data centers as extremist targets while conceding it had no specific plots or suspects.
- The term 'anti-technology violent extremism' does not appear in any publicly available DHS or FBI extremism documents, per Wired's review.
Read full analysis →
Accenture data reveals that non-technical staff, not engineers, are driving an exponential surge in AI token consumption. A major culprit is non-specialized tasks like converting PDFs into markdown files.
Claude Sonnet 5 is out!
AI Weekly's analysis
→
- Anthropic released Claude Sonnet 5 on June 30, 2026, calling it 'the most agentic Sonnet model yet' and pitching it for autonomous browser and terminal use.
- Through August 31, 2026 Sonnet 5 costs $2 per million input tokens and $10 per million output, then steps to standard rates of $3 and $15.
- A new tokenizer means the same input can map to roughly 1.0 to 1.35 times more tokens than prior Anthropic models, partly offsetting the headline discount.
Read full analysis →
Just another reason I can't stand Meta...
AI Weekly's analysis
→
Read full analysis →
Anthropic announces the launch of Claude Fable 5, its next-gen Mythos-class AI. Early testers were able to convert a 50M-line Ruby codebase in a single day, something that usually takes 2 months.
Microsoft disabled 73 of its own GitHub repositories following a major data breach. Hackers pushed a malicious commit to Azure and AI agent tools to harvest developer credentials through platforms like Claude Code and Gemini CLI.
I think this is going to be looked back upon as an artifact of the peak of the current AI hype cycle. We are already seeing sticker shock of unsubsidized AI costs, and I think people will be incredulous when they find out there were internal competitions to see who could use t…
Interesting perspective on one's student's experience of spending their entire college experience in the AI era. #AcademicSky
GitHub Copilot is moving from flat-rate request tiers to usage-based billing, sparking widespread sticker shock among power users. Spot tests show complex prompts or large chat histories can burn through a monthly credit allotment in a single day.
Nearly 2 out of 3 researchers believe the risks of using LLMs for data analysis outweigh the benefits. Yet a poll of 1,900 scientists reveals 60% adopt them anyway out of fear of being left behind. The survey found models designed for specific scientific tasks remain more popu…
Their own posts
Recent commentary
An unanticipated danger of ambient AI: converting a the statement “female mail man” into a “Patient is a 26 year old biological male identifying as a female” #MedSky
Just finished recording my last lecture for an Introduction to AI for Clinical Students class that I’m teaching in two weeks. 30 lectures spread out over 4 weeks! Really interested in how it is received. #MedEd #MedSky
John Shepherd from U. Hawaii, taking about how the pairing of plain text and AI agents was a revolution waiting to arrive. #AIMAHEAD #AgenticAI
Editing times for pediatric admission notes plummeted from 48.5 to 10.8 minutes with ambient AI. Across 127k hospital notes, the tool slashed cognitive burden during initial ED & ward encounters, but it offered no time savings for heavily templated daily progress notes. #amplify2026 #medsky
Keynote talk from Dr. Lee, advancing healthspan with AI and Agentic AI. #amplify2026
Setting up for the Clinical Informatics Keynote: Advancing Healthspan with AI and Agentic AI: Transforming How We Care, Discover, and Share. Over 1118 people in attendance here in Denver! #amplify2026
Setting up in workshop #CI07 Building AI Agents for Healthcare: A Practical Introduction Using Microsoft Copilot Studio Workshop. Here are some nice visualizations of some medical AI agent use cases. #Amplify2026 #MedSky
Walking through how the speakers sort their risk categories for considering approval of Generative AI tools in a clinical setting. #Amplify2026 #Medsky
Standard IT evaluations fall apart for generative AI. There is a lack gold standards for subjective outputs like clinical notes, and background vendor updates cause model drift. Epic's AI discharge summary tool matches human quality but produces more errors. #Amplify2026 #MedSky
LLMs alone are blind to today's lab results and proprietary clinical protocols. RAG bridges this gap by chunking institutional data into searchable numeric vectors. It doesn't make the model smarter, but grounds it in specific, cited documents. #Amplify2026 #MedSky
Their network
In Scott McGrath's orbit
Center = Scott McGrath. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.