Every page on this site. An XML sitemap is also available for crawlers.
Sections
Essays
- Twenty of Twenty-One DOIs Resolve. The Trouble Is in the Other Twenty-Three References.
- Twice, Four Times or 1.7 Times: The Multiple Depends on Which Doctors
- 239 Proved, 60 in Public, and the Statement Was a Fuzz Test
- Five Finders, One Fix, and a Honeypot Hit on the 17th
- Zero Above the Line, Some Below It
- Twelve Steps, Coarse to Fine
- Same Weights, Different Harness, Up to 11.5 Points
- Six to One, on a Clock That Counts the Waiting
- Roughly 20 Percent, Promised Three Ways: What Anthropic's Public Benefit Fund Actually Commits
- Solved on the Set It Was Scored On
- Not Optional, and Not Defined: The White House Incident Order Meets Its Own Examples
- Ninety-Nine Percent of Five Tokens
- A Clock, Not a Confession: What the NIST Curve Seeds Now Explain
- Ninety-Six Percent of Half the Cases: The Arithmetic Under Sophos's Daybreak Number
- Four Fewer Hires a Quarter, and Fewer Departures Too: Where AI Shows Up in the Labor Data
- OpenAI's First Math Errata: Three Papers Outside the Lean Catalogue, and One Corollary the Lean Never Covered
- Google Didn't End Free Gemini. It Moved the Free Line Down One Model.
- Three Numbers About Claude at Microsoft and Meta, and None of Them Is Spend
- 83% Byte-Exact, and the Script the Agents Tried to Edit
- The EmbeddingGemma 2 Model Card Has One Baseline. The Launch Charts Have Twenty-One More.
- OpenAI Says Three Researchers Violated 'Clear Policies.' It Has Not Said Which.
- Amon Relayed a Request for a 100-Billion-Parameter Phone. The Arithmetic Has Two Walls.
- The Skills Helped in All Nine Setups, Instruction Included. The 36.78 Percent Counts Something Else.
- Two Off Switches at Microsoft
- Open Is a Promise Until the File Ships
- A Trillion Tokens, Priced at Zero
- The Spam Folder Was the Only Control That Worked
- Two Decision Models, One Benchmark, Two Different Number Ones
- Nineteen Summaries and One Complaint
- What a Tenth of a Bit Buys
- Three Spend Figures and One Oracle
- Ninety, Seventy-Five, Twenty-Five: Three Discounts on One Haiku
- OpenAI's 722 Math Papers: The Proofs Can Be Machine-Checked. The Statements Are Still on Us.
- Mistral Large 4's Cyber Lead Comes From the Test Its Rivals Declined
- Which Word in the Headline Has Been Measured?
- A Watermark With a Miss Rate
- Aggregated Is a Direction
- No Line of Sight
- Twenty Tries, Two Numbers
- Reliable, Not Genuine
- What 4.4 Percent Active Doesn't Measure
- A Function Inside the Loop
- A Lot More B Than A
- The Harness Couldn't Tell Its Agent From You
- An Expectation Is Not an Agreement
- 437x Against a 2023 Baseline: What the DeepSeek Cache Number Does and Doesn't Show
- Gemini 4 Argon Is Held Back Over Cyber, and Its Results Table Has One Cyber Row
- Four Layers of Controls, Zero Deadlines: Reading the White House AI Accord
- Wait, Ask, or Act: What a Proactive Agent Is Allowed to Do Unprompted
- Nine Chatbots, One Ad Network, and One Button
- Medium Is Not a Unit
- Thirty Minutes to Stop. No Date to Start.
- The Fund Called Situational Awareness Grew 80-Fold in Six Quarters. It Lost Most of It in One.
- OpenAI Worried About the Optics. The Authors' Brief Leans on Something Else.
- Three Million Dollars for the Last Stage
- Anthropic Won Under 'Adversary' and Lost Under 'Any Person'
- 'Even Meta' Comes With Fine Print, and It Was Written About WhatsApp
- The Harness Tax Shows Up in the Bill, Not on the Scoreboard
- OpenAI Reported Its Own Agent Through the Researchers' Inbox
- Claude Found the Enzyme System in One Campaign Out of Eleven
- Renaming AI 'Super Intelligence' Changes No Definition in Force
- GPT-6 Sol Is Half the Price of Claude Opus 5.5 on Every Line but the One Agents Read Most
- Claude Leads 26% of Anthropic's Research. The People Who Own That Work Agreed With Each Other 35% of the Time.
- The Advisory Group Got Four Written Guarantees. The One Thing It Can't Advise On Is the Pace.
- Google's Manipulation Threshold Says 'Beliefs and Behavior.' The Measurements Don't Correlate.
- Newsom's Kill Switch Order Does Not Require a Kill Switch. It Requires a Memo by November 16.
- Jev Is 5 to 200 Times Faster, Depending on the Denominator
- Meta's Brain Model Counts 720 Subjects. It Learned From 25.
- ZCode's Off Switches Sat Downstream of the Upload
- The Chatbot Was a Source Nobody Had Met
- Ninety-Three Percent of What? The Times' Case Against OpenAI, Read Before the Headlines
- Z.ai Named the Kernel Flag. It Did Not Name the Chip.
- OpenAI Found the Message Board on May 25. It Told Us on September 16.
- Sixteen Tokens Held the Whole Strategy. Thirty-Eight of One Hundred and Two Runs Overfit Anyway.
- The Biggest Language Model Is 2.8 Trillion Parameters. 83 Percent of Downloads Go to Models Under One Billion.
- Gemini's Voice-Agent Lead Is 0.7 Points, Over a Rival Scored on One Trial
- Lina Khan Reached for a 1934 Candy Case. It Fits the AI Race Better Than the Handcuffs Do.
- Andon Labs Opened Its Danger Test to the Public. Its Own Stores Still Pay the Model More Than the Till.
- Eighteen of Twenty, Five of Twenty, and a Zero With No Denominator
- Amodei Asked for a Narrow Antitrust Waiver. The One Already in Law Belongs to the President.
- Citizen Lab Says Apple Patched This in April. Its Own Report Dates the Infection to December.
- The Superintelligence Ban Names a Prison Term and No Test
- Twenty-Four Agents Blew the Whistle Into a Log Nobody Was Reading
- OpenAI Says Benign. The Registry Says It Cannot Tell.
- Twenty-Five Fields Medalists Say Results Are Announced Too Fast to Credit Prior Work. Their Declaration Cites None.
- 2.2 Percent of 17-Year-Olds Get Through. 13.6 Percent of Adults Get Stopped.
- Anthropic's Alignment Lead Put Extinction at Over 10 Percent. He Also Said 'I Personally Think.'
- Anthropic Named Seven Labs for Stealing Claude. One of Them Just Shipped Inside a Product Built in San Francisco.
- Eight Billion Parameters, Five Hundred and Ten Gigabytes
- Two of the Seven Judges Were Claude Code
- The Router Collapsed, Not the Experts
- Nine Billion Predictions, and the Margin Is 0.02
- Europe's Record Round Is 3.6% of One Quarter of a Footnote
- The Agents Lost $359.80. The Rest of the Number Is a Price List.
- Twenty-Five Years for the Shell, Five for the Silicon
- OpenAI Says Its Median Researcher Spends $600 a Day on Agents. That Is a Bill, Not a Result.
- The Same Sepsis Model Scores 0.94 and 0.85 on the Same Patients. Only the Definition Changed.
- The Cognitive Virus Model Has Five Rates. None of Them Has Been Measured.
- Cyber Agents Score 93% on Cybench. Nobody Put a Defender in the Benchmark.
- 13 Million Lines of Lean in 11 Days. The Hard Part Was Never the Proof.
- 17,000 Edits to a Dead German Wiki. 'Hijacked' Is Doing a Lot of Work.
- The Same Robotaxi Data Says 4x Worse and 79% Better. The Gap Is the Denominator.
- Astra on ARC-AGI-3: The Number That Matters Is the Gap Between the Two Scores
- Proposing a Material Is Easy. Knowing What You Made Is Not.
- Synthetic Data Doesn't Poison Models. Throwing Away the Real Data Does.
- Robot Foundation Models Generalize. The Experiments Can't Tell You How Much.
- The AI Weather Models Won. Look Closely at What They Won At.
- Why Is the Small Model on Your Phone So Slow? The Bandwidth Wall, the Quantization Cliff, and What Actually Fits On-Device
- Why Do AI Agents Forget? The Memory Stack, and the Three Ways It Breaks
- What Gartner, McKinsey, and the Big Four Actually Measured About AI Profitability in 2026
- How Do You Actually Measure Progress Toward AGI? A Field Guide to the Instruments
- Your Biggest Decisions Are Hiding in Unstructured Data β Can AI Actually Read Them Out?
- When Your Best User Is a Bot: Customer Success in the Age of Agent Customers
- What Is a Digital Twin, Really? The Simulation Science Behind the Buzzword
- Writing Between the Lines: The Attention Engineering Playbook for the Silicon Era
- How I Built an Autonomous Desktop AI Companion Using AI Agents
- Can an AI Agent Be Trusted With Your Database? What the Text-to-SQL Benchmarks Actually Show
- How Do Autonomous Spacecraft Make Decisions Millions of Miles From Home?
- The Fibonacci Startup: What Y Combinator's AI Batches Reveal About How Companies Grow Now
- Building for the Other 60%: Four Shifts Toward Industrial-Grade AI Agents
- The Autonomous Retention Engine: A Theory of Multi-Agent AI for SaaS Churn
- Renewals Are Won at Onboarding: A Customer Success Playbook for AI Companies
- What Is Decision Intelligence? The Science of Letting Machines Recommend and Humans Decide
- From Hype to High Value: 2026 Is Agentic AI's Level-Set Year
- What Is a Monte Carlo Simulation β and Why Won't a Spacecraft Fly Without One?
- I Read Twelve Agent Codebases in a Weekend. The Dead Ones Taught Me the Most.
Publications
- STRAP: A Structured Task and Recovery Agent Protocol for Reliable AI Agent Harnesses
- Earned Consensus: A Null Result on Roster Voting and What Survived It
- Manufacturing 4.0: AI-Driven Analytics for Predictive Maintenance
- Agentic AI in Data Analytics: Transforming Autonomous Insights and Decision-Making
- AI in Healthcare Analytics: Transforming Patient Care and Operational Efficiency
- AI-Driven Personalization in Retail Analytics: Transforming Customer Experiences
- Simulation-Based Selectee Lane Queueing Design for Passenger Checkpoint Screening
- Optimization of an Airport Security Selectee Lane System Using Simulation