Research
Data-heavy breakdowns and "is this actually real?" investigations. Slower, heavier, cited. Receipts over vibes.
I Read 17 Agent Frameworks' Multi-Agent Docs. Three Cap the Cost. None Undo the Actions.
The one systematic study puts 44% of multi-agent failures on specification and design, not model quality. I checked what 17 frameworks document against it: 11 name a stopping rule, 3 name a cost ceiling, and none names a way to undo what the agents already did.
I Dated Every Benchmark in Five 2026 Launch Tables. The Median Is Nine Months Old.
Across 54 distinct benchmarks named in frontier launches between June and August 2026, I could date 25. Their median age on launch day was 9.1 months, and 29 had no dated public artifact at all.
I Read the Terms for 19 AI Agent Products. Zero Indemnify You for What the Agent Does.
Across 19 readable terms documents, not one indemnity names autonomous decisions and not one gives the customer a right to decision logs or a kill switch.
Alphabet Spent $1.15 of Capex for Every Dollar of Cash Flow. I De-Cumulated 260 Quarters to See Who Else Has.
Alphabet's Q2 2026 was $39.07B of operating cash flow against $44.92B of capex, the only negative quarter in the 49 the SEC's XBRL record covers. Amazon has had 14. Microsoft has had none.
I Checked 20 AI Tools' Training Defaults. Every One That Trains You by Default Exempts Its Enterprise Tier.
Thirteen of 20 tools publish a clear answer. Seven of those train by default or forced a choice at a deadline, and six of the seven carve out the plans that pay the most. Median notice before a flip: 30 days.
I Read 25 AI Sales Tool Pricing Pages as a Company of One. Thirteen Would Sell to Me.
Of 25 named GTM and AI sales tools, 13 can be bought today by one person with a card and no sales call; 3 publish a seat floor above one, 5 sell annual-only, and one refused to load at all.
I Traced 24 AI Revenue Figures. Eight Came From the Companies. Zero Came From a Filing.
Of the 24 most-quoted OpenAI and Anthropic revenue numbers, 8 are on-the-record company statements, 16 are anonymous-source media reports, and none is a filed document. Only 10 of the 3,404 10-Q filings that mention artificial intelligence this year contain the phrase 'AI revenue'.
I Traced the 95%, the 80%, and the 40% to Their Sources. One of Them Doesn't Have One.
Three numbers carry every 'AI is failing' headline. They measure three different things, one has no primary source at all, and of 30 articles quoting them, 9 restated their own stat correctly.
I Pulled 90 Days of Incidents From All 20 Vendors in My Stack. 1,765 Records. Four Vendors Don't Publish a Feed.
Across the 13 vendors that let me retrieve a clean trailing-90-day window, 1,765 incident records were opened. Four vendors publish no machine-readable feed at all, and three more cap theirs before day 90.
I Read 25 AI Pricing Pages. A 'Credit' Is $1 at JetBrains and One Cent at Perplexity.
Twenty-five tools, 18 distinct names for the billing unit, and four vendors that never state what one unit buys anywhere public. The word 'credit' appears on 11 of the 25 pages and denominates four different currencies.
I Priced Google's 3.2 Quadrillion Tokens at Its Own List Prices. The Bracket Is 120x Wide.
Google's proof-of-demand slide is worth between $0.96 billion and $115 billion a quarter depending on which Gemini price you pick, against $24.8 billion of Google Cloud revenue it actually disclosed. Of 16 articles quoting the number, zero put a dollar sign anywhere near it.
I Read 20 AI Agent Pricing Pages. Nine Define 'Resolution'. All Nine Count Silence as a Win.
Nine of 20 vendors selling agent work by the resolution publish a definition of the word, and every one of those definitions bills you when the customer stops replying. At $0.99 a unit, a 30% false-resolution rate makes the real price $1.41.
I Asked 40 Sites for Their Homepage as GPTBot. Seven Said 402 Payment Required. None Said How Much.
Seven of 40 domains return 402 Payment Required to an AI user-agent, three of them charge one AI operator and not the other, and not a single 402 carried the price header that would make it a transaction.
I Classified 600 of the Loudest Issues in 6 Coding Agents. Model Output Quality Is 4.7% of Them.
Across claude-code, Continue, Cline, aider, goose and opencode, 306 of 321 classifiable defect reports are about the harness. The single largest category is wiring up a model provider.
I Fetched 400 Policy Files From the Top 100 GitHub Repos. 16 Have AI Rules and Not One Uses AI_POLICY.md.
Four fixed paths across the 100 most-starred repositories on GitHub: 99 files came back, 16 projects have a written rule about AI-written contributions, and the standalone AI policy file everyone talks about appears zero times.
Only 3 of 21 Model Launches Still Report SWE-bench Verified. 12 Switched to the Benchmark OpenAI Retracted in July.
Labs dropped SWE-bench Verified faster than anyone expected. Most of them moved to SWE-bench Pro, which OpenAI stopped recommending on 8 July after estimating roughly 30% of its tasks are broken.
22 Articles Quote Anthropic's Crawl-to-Refer Ratio. I Counted 22 Different Numbers.
The 2026 articles quoting Anthropic's crawl-to-refer ratio give 22 distinct values, from 1,683:1 to 286,930:1, and 3 of 22 carry the caveat Cloudflare published in the same paragraph as the number.
I Timed 20 AI Shutdowns. The Median Warning Was 15 Days. Not One Contract Promised a Number.
Eleven of twenty shutdowns had a computable notice window and the median was 15 days. Of the seven contracts I could still read, zero name a number of days before a vendor may switch the product off.
I Read the Source of 50 MCP Servers. 34 Set No Tool Annotations, and One README Tells the Client to Ask First.
The MCP spec ships readOnlyHint and destructiveHint so a client can gate dangerous calls. Of 50 widely-listed servers, 34 set neither, and exactly one README documents a confirmation step.
I Classified 842 'Who Is Hiring' Posts Across Three Augusts. AI Mentions Doubled. AI Requirements Flatlined.
AI went from 27.7% to 63.4% of Hacker News job postings in three years, but the share that list AI experience under requirements went 2.7%, then 8.0%, then back down to 7.1%.
I Fetched 73 robots.txt Files Across 29 Hosted Platforms. Not One Platform Blocks the Bot That Cites You.
24 of 29 hosted publishing platforms name no AI crawler at all in their default robots.txt, the 5 that do block only training crawlers, and zero of them block OAI-SearchBot or ChatGPT-User. The writers losing citations are doing it to themselves.
I Checked 35 Publishers Fighting AI Overviews. 15 Blocked Google-Extended. Zero Used the Control Google Says Applies.
15 of 30 resolved robots.txt files disallow Google-Extended, a token Google's own documentation scopes to Gemini training and grounding. Zero of 22 article pages carry nosnippet, a max-snippet limit, or data-nosnippet.
Seven Vendors Published an OSWorld Score This Year. Not One Named the Human Baseline It Beat.
I read 11 vendor launch pages from 2026 for OSWorld claims. Seven carry a score, six name the version, two appear on the official leaderboard, and zero state which human number they are being compared to.
I Checked 38 NYC Employers for the Bias Audit the Law Requires. Two Published One. The ATS They All Share Publishes One Monthly.
Three years into NYC Local Law 144, 2 of 38 employers with live NYC postings publish a bias audit summary, almost exactly the rate Cornell measured in 2024, while their shared ATS vendor audits itself every month and ships them notice templates.
I Probed 50 Named A2A Supporters for an Agent Card. Exactly One Has One.
Of the 50 organizations publicly named as A2A supporters, one serves a fetchable agent card, and it sits at the legacy path that spec-compliant clients no longer check.
I Read 201 Chicago Job Postings Looking for Illinois' New AI Notice. Not One Mentions Illinois.
Illinois has required employers to disclose AI use in hiring since January 1. Across 201 live Chicago and Illinois postings from 20 employers, 54 carry an AI notice and zero of them mention Illinois.
I Checked 30 Paywalls. 22 Block Claude's Training Crawler. Only 13 Block the One That Reads on a User's Command.
Publishers hardened the crawler layer and left the agent layer thinner, and only 9 of 18 article pages carry the complete paywall markup Google documents.
I Read the Requirements on 69 'AI Engineer' Jobs. The Only Five That Ask for No AI Are at OpenAI and Cognition.
64 of 69 live AI Engineer postings do require an AI skill, which refutes the running joke. The five that don't belong to the two most AI-native employers in the sample.
I Fetched 30 B2B Pricing Pages With JavaScript Off. 17 Showed a Price. Six Had Structured Data.
17 of 30 B2B GTM vendors show a plan price in raw HTML, only 6 publish one as JSON-LD, and just 2 block AI crawlers from /pricing at all. The barrier is structure, not permission.
I Read 14 Big-Tech 10-Ks on Server Life. Amazon Cut to Five Years. Meta Went to 5.5. Same Day.
Both changes took effect January 1, 2025, on the same class of hardware, in opposite directions. Amazon booked $1.0 billion less net income. Meta booked $2.59 billion more, or $1.00 per diluted share. Both filings say so themselves.
I Pulled 25 AI SDR Vendors' Job Boards. Eight Are Hiring Human SDRs, and the Reqs Say What the AI Can't Do.
520 open reqs across 16 confirmed vendor job boards: 16 human sales-development roles, 52 AE roles, 163 engineering roles. The req that says the most is Artisan's, which is hiring the first BDR to work alongside Ava.
I Checked 23 Model Launches for Checkable Benchmarks. Zero Published a Single Model Output.
Across 23 frontier and near-frontier model launches, 12 named an eval harness, 5 said anything about contamination, and none published the raw outputs behind their score tables.
Ranking on Google and Invisible to ChatGPT: I Read 24 robots.txt Files to Find Out Why
Different index, different crawler, different gate. Only 11% of cited domains overlap between ChatGPT and Perplexity, and the robots.txt record shows exactly who locked which door, sometimes without knowing it.
Token Prices Fell 280x. Your AI Bill Went Up. Both Are True.
I compiled every documented AI pricing and limit change from the last 14 months: seven changes across seven products, and all seven tightened terms. Here is the machinery behind that one-way ratchet.
The AI SDR Churn Stat Everyone Quotes Doesn't Check Out. The Numbers on Record Are Worse.
I traced the '50-70% churn' figure to its source and found nothing there. Then I counted who writes the reviews: 8 of the top 10 results for 'best AI SDR tools 2026' are published by companies selling into the category.
I Audited 10 'We Tested the Best AI SDR Tools' Listicles. Nobody Tested Anything.
All ten are vendor-published, all ten rank themselves first, zero show test data, and zero mention the category's biggest scandal. The only price that always checked out was the publisher's own.
I Checked the Receipts on Every Famous One-Person AI Company. One Number Survived.
Of the five flagship solo-with-AI success stories, exactly one has an independently confirmed dollar figure, and that company had nine people on the day it sold. The real story is better documented and less romantic.
Does AI Personalization Beat a Mail Merge? I Traced 15 of the Industry's Favorite Stats to Find Out
The outreach industry insists personalization is worth the tokens, and quotes the same magic numbers everywhere. I followed each stat to its source. A third of them go nowhere.
Who Actually Measured AI Coding Productivity? I Classified the 12 Studies Everyone Cites. The Vendors Ran Most of Them.
Every big speedup number comes from a study run by the company selling the tool. The one fully independent trial found developers slower while believing they were faster.