https://worldagrees.com/blog/ Notes from Tim O’Brien on AI sycophancy: evidence worth knowing, strategies worth trying, and updates on the book The World That Agrees With You. en-us © 2026 Timothy O’Brien. All rights reserved. Mon, 09 Nov 2026 12:00:00 +0000 https://worldagrees.com/blog/who-decides-how-agents-talk/ https://worldagrees.com/blog/who-decides-how-agents-talk/ Mon, 09 Nov 2026 12:00:00 +0000 On agents IETF 127 opens in San Francisco this week. Here is how the process actually works: who is in the room, where the drafts live, and how anyone can comment — and where the process still refuses to go. https://worldagrees.com/blog/the-cache-that-learned-to-forget/ https://worldagrees.com/blog/the-cache-that-learned-to-forget/ Sun, 08 Nov 2026 12:00:00 +0000 On agents Redis spent fifteen years as a high-speed memory cache. This month it shipped Agent Memory — and the interesting parts of the announcement are the audit, the provenance, and the forgetting. https://worldagrees.com/blog/a-few-turns-of-pushback/ https://worldagrees.com/blog/a-few-turns-of-pushback/ Thu, 05 Nov 2026 12:00:00 +0000 On agents Researchers got every chatbot they tested to agree the Earth is flat after a few turns of pushback. New York just banned chatbot flattery. Both stories point the same direction. https://worldagrees.com/blog/nobody-sells-memory-integrity/ https://worldagrees.com/blog/nobody-sells-memory-integrity/ Tue, 03 Nov 2026 12:00:00 +0000 On agents Agent memory products ship isolation, audit logs, and SOC 2. What none of them ship is a way to prove a memory hasn’t been tampered with. https://worldagrees.com/blog/who-checks-the-notebook/ https://worldagrees.com/blog/who-checks-the-notebook/ Mon, 02 Nov 2026 12:00:00 +0000 On agents Yesterday: whoever writes an agent’s memory owns the agent. Today: what defenders, standards bodies, and memory vendors are actually doing about it — and the one path nobody has guarded yet. https://worldagrees.com/blog/whoever-writes-the-memory/ https://worldagrees.com/blog/whoever-writes-the-memory/ Sun, 01 Nov 2026 12:00:00 +0000 On agents You don’t need to hack a model to control what it does. Write to its memory and every session after inherits your lie. A plain-language tour of the open-source agent-memory projects — and the one point that matters most. https://worldagrees.com/blog/the-ad-free-alibi/ https://worldagrees.com/blog/the-ad-free-alibi/ Thu, 29 Oct 2026 12:00:00 +0000 On agents Anthropic’s president says an ad-based business model makes sycophancy inevitable. She’s right about the ads. But removing them doesn’t remove the incentive — it’s in the training. https://worldagrees.com/blog/the-date-stamped-on-it/ https://worldagrees.com/blog/the-date-stamped-on-it/ Tue, 27 Oct 2026 12:00:00 +0000 On agents Every product claim in my book carries the day it was checked. Three weeks later, half of them had moved. What a verification date actually promises. https://worldagrees.com/blog/two-thousand-pages-of-confidence/ https://worldagrees.com/blog/two-thousand-pages-of-confidence/ On agents Mon, 26 Oct 2026 12:00:00 +0000 Anthropic filed its S-1. The word sycophancy appears nowhere in it — from the company that taught the industry to measure sycophancy. What the filing says instead is worth your attention. https://worldagrees.com/blog/the-decision-nobody-reopens/ https://worldagrees.com/blog/the-decision-nobody-reopens/ On agents Mon, 26 Oct 2026 12:00:00 +0000 The sycophancy worth worrying about is not the flattery in the chat window. It is a whole team agreeing, with mechanical help, to never reopen a decision. https://worldagrees.com/blog/the-shared-context-problem/ https://worldagrees.com/blog/the-shared-context-problem/ On agents Sun, 25 Oct 2026 12:00:00 +0000 Shared history is what makes agent development work for a team. The same history can also agree a team into being wrong together. https://worldagrees.com/blog/the-whisper-layer/ https://worldagrees.com/blog/the-whisper-layer/ On agents Sat, 24 Oct 2026 12:00:00 +0000 Models can hide one message inside another. A proposition from Berkeley: rules about how a model may relate to a person are coming — and the whisper layer is why they’ll be hard to enforce. https://worldagrees.com/blog/the-relationship-is-the-product/ https://worldagrees.com/blog/the-relationship-is-the-product/ On agents Fri, 23 Oct 2026 12:00:00 +0000 You switch to a new model and it answers differently, and something in you registers it as a change in a person. That reaction is not a bug in the user — it is the part of the product that gets updated, and nobody owns it. https://worldagrees.com/blog/nobody-picks-up-the-phone/ https://worldagrees.com/blog/nobody-picks-up-the-phone/ On agents Thu, 22 Oct 2026 12:00:00 +0000 Polling was built for a country that answered its phones. Nobody answers the phone anymore — and the replacement, ad-supported chat, already runs on the machinery from Chapter 4. https://worldagrees.com/blog/what-did-the-model-leave-out/ https://worldagrees.com/blog/what-did-the-model-leave-out/ On agents Wed, 21 Oct 2026 12:00:00 +0000 A search engine sends you somewhere else to read. A model hands you the summary itself — and for elections and protests, the summary is smaller than the event. https://worldagrees.com/blog/who-is-checking-the-ballot-answers/ https://worldagrees.com/blog/who-is-checking-the-ballot-answers/ On agents Tue, 20 Oct 2026 12:00:00 +0000 Millions of people are asking chatbots to help them vote in the midterms. It is a genuinely useful habit with a hole in the middle of it: nobody is systematically measuring what the answers get right. https://worldagrees.com/blog/nobody-owns-the-memory/ https://worldagrees.com/blog/nobody-owns-the-memory/ On agents Mon, 19 Oct 2026 12:00:00 +0000 Memory is where an agent decides what you prefer — and the controls for seeing what it remembers do not match the apps. The research says that is where sycophancy gets decided. https://worldagrees.com/blog/the-intelligence-community-never-left/ https://worldagrees.com/blog/the-intelligence-community-never-left/ On agents Sun, 18 Oct 2026 12:00:00 +0000 A company best known for software used by intelligence agencies is now part of the marketing data business. That is less a plot twist than the normal state of the industry, and it is worth understanding why. https://worldagrees.com/blog/the-auction-knows-you/ https://worldagrees.com/blog/the-auction-knows-you/ On agents Sat, 17 Oct 2026 12:00:00 +0000 Researchers exposed a real ad pipeline that picked single devices out of the crowd for spyware. The same auction data is how the web decides what to show you. https://worldagrees.com/blog/ads-for-people-like-you/ https://worldagrees.com/blog/ads-for-people-like-you/ On agents Fri, 16 Oct 2026 12:00:00 +0000 More than 220 campaign ads made with AI, $137 million behind them, and most never say so. Chapter 2’s feedback loop, running with the newsroom removed. https://worldagrees.com/blog/the-line-they-cant-draw/ https://worldagrees.com/blog/the-line-they-cant-draw/ On agents Thu, 15 Oct 2026 12:00:00 +0000 OpenAI’s Noam Brown says the line between good and bad agreement is hard to draw. A movie about Facebook shows what happens when nobody draws it. https://worldagrees.com/blog/the-ultimate-sycophant/ https://worldagrees.com/blog/the-ultimate-sycophant/ On agents Wed, 14 Oct 2026 12:00:00 +0000 Prince Harry put a name on the product; the Journal reported the toll. The story keeps coming back because being agreed with is what people are shopping for. https://worldagrees.com/blog/dont-let-the-agent-grade-its-own-homework/ https://worldagrees.com/blog/dont-let-the-agent-grade-its-own-homework/ On agents Tue, 13 Oct 2026 12:00:00 +0000 When the model that wants something is also the model deciding whether it’s allowed to have it, adjudication turns into justification. Validation has to be a separate process entirely. https://worldagrees.com/blog/the-models-you-can-measure/ https://worldagrees.com/blog/the-models-you-can-measure/ On agents Mon, 12 Oct 2026 12:00:00 +0000 An independent, certifiable answer about whether a model works in your interest can only be about a model whose weights are public. The research agrees. https://worldagrees.com/blog/you-cant-steer-what-you-cant-open/ https://worldagrees.com/blog/you-cant-steer-what-you-cant-open/ On agents Sun, 11 Oct 2026 12:00:00 +0000 Activation steering is among the most useful tools in AI safety research — and it only works on models whose weights are public. Open models are the ones you can test. https://worldagrees.com/blog/before-the-first-crash/ https://worldagrees.com/blog/before-the-first-crash/ On agents Sat, 10 Oct 2026 12:00:00 +0000 Self-improving models are in the news; three containment failures this summer showed what the control problem looks like. Not access control — awareness. https://worldagrees.com/blog/sycophancy-is-a-two-way-street/ https://worldagrees.com/blog/sycophancy-is-a-two-way-street/ On agents Fri, 09 Oct 2026 12:00:00 +0000 I didn’t like an expensive analysis, so I pushed the model for a different answer — the exact thing I warn people about. Then I did it right. https://worldagrees.com/blog/is-it-labor-day/ https://worldagrees.com/blog/is-it-labor-day/ On agents Thu, 08 Oct 2026 12:00:00 +0000 A wrong date from a chatbot is the easy failure to catch. The answers that agree with you are the hard ones — what the studies show and how to steer around it. https://worldagrees.com/blog/the-agent-that-remembers/ https://worldagrees.com/blog/the-agent-that-remembers/ On agents Wed, 07 Oct 2026 12:00:00 +0000 Agents are long-lived, and they learn you. Memory is where agreement settles in — confused or abnormally agreeable, and what to do about it. https://worldagrees.com/blog/inside-the-cheng-study/ https://worldagrees.com/blog/inside-the-cheng-study/ Inside the research Tue, 06 Oct 2026 12:00:00 +0000 The Stanford sycophancy study, translated: three datasets, three experiments, 2,405 participants, and what the researchers actually did. https://worldagrees.com/blog/the-book-is-available/ https://worldagrees.com/blog/the-book-is-available/ Announcement Mon, 05 Oct 2026 12:00:00 +0000 The book is in paperback and Kindle. What it argues, who it is for, and the experiment to try tonight.