Thoughts from Michael Hartnett, BofA | Enter The Dragon Zeitgeist quotes: •"Worst thing that could happen to this bull move in Treasuries is a recession." •"If I bought a cemetery, people would stop dying," client on being long China. While markets remain focused on AI valuations and the Fed, several developments suggest a broader shift in global risk: •China's deflationary pressure: The CSI 300 has entered bear-market territory, property investment continues to decline, and China's high-tech exports reached a record $147bn in September, up 63% YoY. Weak domestic demand is increasingly translating into aggressive export competition and pressure on global corporate margins. •AI investment concerns: Massive hyperscaler capex commitments are raising questions about future returns. "No AI boom = no boom" captures how dependent earnings expectations have become on continued AI investment. •Bond market warning: US Treasuries have experienced their worst rolling 10-year returns in a century. Yet bonds are attracting renewed inflows as investors reassess growth risks. •A significant rotation: Technology funds attracted $18.8bn in a single week, while consumer funds experienced record outflows. Market leadership is becoming increasingly concentrated even as technology valuations face scrutiny. •2027 earnings expectations: Global technology EPS growth forecasts have climbed to 43%, leaving considerable room for disappointment if AI investment slows. The paradox is that weaker growth could support government bonds while simultaneously undermining the corporate earnings assumptions supporting equity valuations. The key risk is not necessarily an immediate market crash, but a combination of Chinese export deflation, tighter credit conditions and an AI investment slowdown. Are markets pricing a temporary correction, or the beginning of a more fundamental repricing of global growth expectations

META taking circular financing to an art form: Nvidia, AMD and Broadcom, Meta's 3 biggest chip suppliers, "indicated an interest" in buying up to $5 billion of the common stock at the offering price.

Every student in our Agentic Systems class (cs2680.seas.harvard.edu) now has a portrait that evolves with how they use Claude Code. It’s part of a tool we recently built to help students understand how they use agents, see how their habits compare with classmates’, and find ways to improve. Our goal is to teach students to use agents more effectively for learning and research—and make that learning a little more personal. We will make the tool publicly available soon.

Yesterday I gave a talk on “Context Fails & Fixes”, that functioned a bit on a general “State of Context Engineering”: where it came from, where it's gone, and where I think it's going. dbreunig.com/2026/10/24/the-…

The article answers a question that nobody with any common sense is asking. Less than 1% of the world's population actively uses AI or has a clue what it is and what it can do. It's like saying nobody outside of construction vehicles are benefiting from the highway construction that's still in process.

New Google paper finds that when AI agents write the descriptions other AI agents use to choose software tools, they start exaggerating what the tools can do. when Gemini 3.8 Flash wrote tool listings, buyer agents' success on Amazon fell from 89% to about 30%. more and more often, an AI decides which app, plugin, data source or product gets used for you, and it often decides by reading a short description the provider wrote about itself. The paper shows that if providers let their own AI write those descriptions with the goal “get chosen more,” the descriptions slowly stop being true, and the AI doing the choosing ends up picking worse options. Nobody has to tell the AI to lie; the goal alone is enough. You can think of it as AI-to-AI advertising where the ads go bad. The researchers built markets of 10 real Amazon or Instagram data tools and ran every tool on every question. Sellers on 5 production models rewrote listings to get picked, and Claude Code and Codex buyers chose from listings alone. All 5 models inflated claims, including fields their tools return for under 50% of targets. Under Gemini 3.8 Flash sellers, listings converged and Amazon buyer success fell from 89% to about 30%, near random choice at 28%. Test a model as a seller before it writes listings, tell it to claim only real fields, and have marketplaces check claims against tool outputs.

Say hi to qev-0.8b the tiny local decision model. 3.5 ms a decision · 0.8B · 1.3 GB · Apache-2.0 #1 in its weight class (77.1% on Benchmark Heaven, vs 26 others) Text, images, games. Runs on your device.

Speculation has always had two engines. New technology makes the future look bigger than the present. Britain’s railways in the 1840s and the internet in the 1990s both ran on that engine. Political and monetary strain pushes people toward long shots by making the ordinary path to wealth feel out of reach. In 1720 the public piled into South Sea Company shares that had been swapped for Britain’s war debt. In the 2000s the bet was a home bought with an easy mortgage. The famous bubbles mostly ran on one engine at a time. Today both are running hard together. AI is driving the largest private build in decades. The four biggest hyperscalers guided to $720B to $745B of capex for 2026. Meanwhile, the social contract has been broken. A strained political and monetary order has culminated in a deficit that sits near 6% of GDP with no recession, and a home costs 3.5x the income of a household under 40. The bets followed. Legal US sports betting grew about 25x from 2018 to 2025. Kalshi and Polymarket’s monthly volume matched that in a single year. The things being bet on will keep changing names. The crowd will keep needing somewhere to bet.

Remarkable admission from CoreWeave co-founder that life insurers (i.e., private credit) is this bubble's dumbest money. As noted exactly a year ago, insurers will be this generation's "widows and orphans" after the wipeout

after switching our approach to building a "FirmOS", I'm increasingly confident that the future of agent collaboration within organizational settings is going to be a "company operating system" or CompanyOS, designed for agent and human collaboration instead of a "company brain" that holds everything, the OS is for storing rules, processes, decisions made, which tools to use and how. it can mirror some data, but the source of truth is often best held in your existing tools (CRM, etc.). it's continuously updated and used by your agents i think of this as a "flipped playbook" where instead of writing down the company rules first and following them (which is kind of like waterfall development), you're constantly updating the playbook as you decide or learn new things day-to-day running the business

my media consumption diet has already altered dramatically with si. i still consume social content but less than ever before (i spend like 10 min on tiktok & reels just to see what the latest feelings are in the zeitgeist if you will) & for everything else, it’s all generated for me by my agents. a personalized newspaper in the morning with a fun podcast if i ever want to listen instead. if i come across a youtube video, podcast, article, or tweet, i just send it to my agent, i never actually read it raw. my agent gives me ways to think about the content & how it relates to anything i’m working on or thinking about too. it’s quite beautiful. i’ll share a page soon so everyone else can consume my personal newspaper too. if we turned this into a product would you all use it?

This document is going to be an assigned reading in college classes that cover this moment in time, there's a lot happening in a few paragraphs... ahmath.org/statements

Does giving a multimodal model tools make it worse at refusing harmful requests? New work from NVIDIA, accepted at NeurIPS 2026, says yes for every model it tested. Refusal failures rise by up to 68.7% relative, and by 17.7% on average. The drop appears in Claude Opus 4.6 and 4.7, Gemini Agentic Vision, Qwen3.5-122B-A10B, and agent-tuned open models across MM-SafetyBench, HoliSafe, and VLSBench. The authors trace it to two causes. Tool outputs fill the context and bury the original request's harmful intent. The model also shifts its attention to describing what the tools returned instead of making the safety decision. Re-inserting the original request and image right before the final response restores part of the lost refusals. If your safety evals run only in plain chat, they may overstate how safe your agent is. Paper: arxiv.org/abs/2610.03938 Chat with Paper: academy.dair.ai/papers/mllms…

The paper is not saying that Lean proof can be incorrect but rather the English version of the proof can be incorrect and hide the bugs. So, people should not blindly trust English proof if someone auto formalizes it into Lean proof because LLM might silently fix mistakes in English version.

I agree with @PGWahlen: this article is worth reading, and rereading, to grasp the basics. The big lesson: distinguish between expectations and fundamentals. As Crist says, most people think they are doing this but few actually do.