# What I read this week: AI's mistakes became crit notes, your agents need a manager, and Nvidia's CEO resigns

> Source: <https://johnmaartifacts.substack.com/p/what-i-read-this-week-ais-mistakes>
> Published: 2026-11-16 15:03:27+00:00

## The top story: Fable 5.2 is almost AGI, and its misses read like design crit

AI’s mistakes used to be made-up facts. Fable 5.2’s are too long, too repetitive and no fun.

Anthropic released Claude Fable 5.2 and Mythos 5.2. [Per Rowan Cheung’s roundup,](https://x.com/rowancheung/status/2120299193868615981) Fable debuted at #1 on the Artificial Analysis Intelligence Index, 5 points ahead of Google’s Gemini 4 Carbon, and it’s 20% cheaper and up to 30% faster than 5.1. Google lost the top spot before Carbon shipped.

Ethan Mollick posted that Fable 5.2 sits [above an already steep trendline on three benchmarks](https://x.com/emollick/status/2120000888974766501) that aren’t saturated yet: ECI (a composite), Remote Labor (end-to-end freelance jobs) and RoboDojo (controlling robots).

Then he had early access for a while and wrote it up. His verdict: almost AGI, which he defines as “as good as most human experts at most intellectual tasks you can do with a computer.” He calls it [Jagged AGI.](https://oneusefulthing.org/p/fable-52-and-jagged-agi)

**What it got right:** a short film in one shot from one paragraph of prompt. It also turned a tiny USB keyboard into a game console with six games, with no public docs, on the first try. It built 204 live 3D worlds in 20 hours.

**The misses:** a research paper on LLC filing fees ran 182 pages, because every critique added analysis and nothing got cut. Told to “polish,” it wrote a rule book: 204 sections, 87,500 words, and a full day to launch one world. It put critics and fact-checkers on almost everything, and none of them asked whether the work was too long, repetitive or fun. That fell to Mollick every time.

**His own test:** “Knowing when your own work is not good enough is a large part of what makes someone an expert.”

**What I learned:** Fable had critics and fact-checkers on almost everything. None of them asked if the work was fun. Those are crit notes, and the edit is what it hands back. That’s a designer’s job.

## Managing agents

**Your coding agents made you a manager (Addy Osmani).** His claim: if you spend all day with a coding agent, you’re already managing. You pick tasks, brief, delegate, follow up, judge results and decide what ships. The sections are ordered by impact, and the first is: give the agent a way to check its own work, or you become its test suite.

**The rest of his list:** tell it what it can’t see in the repo and skip what it can. If you correct something twice, put it in CLAUDE.md. After two failed corrections, start a new session with a better prompt instead of trying a third time. Approving a 600-line diff because the tests are green “is not code review, this is code ratification.” And past the first few agents, each extra one costs more of your attention than the last. He calls that the orchestration tax, and says most days he runs four or five background agents and some days that’s too many.

[The components of agent memory (Anthropic).](https://www.anthropic.com/research/components-of-agent-memory) Two Claude models played Pokémon Red with a folder for notes. By step 10,000, Sonnet 3.7 had 413 files across 225 folders and one badge. Fable 5 had nine files in two folders plus a global.md, and four badges. Sonnet 3.7 saved things it could read off the screen anyway, like money and map position. Fable 5 saved only what the game would never tell it: the goal, the plan, rules it had tested.

Two more findings. Telling Opus 5 to “record only failures” cost it 0.234 of its memory gain, so stronger models do better when you say what memory is for and leave the how to them. And “dreaming,” where Claude rereads old transcripts between sessions to fix its notes, caught a wrong note about which side of Mt. Moon the PokéCenter was on. The corrected run reached Cerulean City at action 1,935. The uncorrected one never got past the mountain.

**One sentence that [makes an agent act like an owner (my Day 27).](https://x.com/johnmatechAI/status/2120532003699184123?a=1h_Z-EazK-cFy0wvyDI&s=20)** “This codebase is yours: what it’s for and how it’s built are your call, and you’ll keep building on it.” With that line, the model merged repeated code in about half of tries. Without it, under a tenth. “Act more freely” changed nothing.

**What I learned:** In all three, a few words around the model changed what it did: a check it can run, a rule about what memory is for, a sentence saying the work is its own. Addy’s correct-twice rule and Fable 5’s nine memory files point the same way. Write down less, and make each line count.

## AI moves into the tools you already use

[Claude Sheets, Whiteboards and Data (Anthropic).](https://x.com/johnmatechAI/status/2121256791791087725) Sheets and Whiteboards in beta, Data generally available on paid plans, and Claude Meetings coming next. Context carries from a board to a sheet to a dashboard without re-explaining it. Advantage Solutions’ FP&A team went from about 10 hours per forecast cycle on slides and commentary to under 30 minutes. [Rowan’s read: Claude is slowly rebuilding the entire office suite.](https://x.com/rowancheung/status/2120299193868615981)

**Grok Bot 1.0 (Grok).** “AI teammates that work alongside your whole team.” You share bots, work with them in Slack and Teams, invite them to meetings, and they learn as they go.

**Businesses in ChatGPT (OpenAI).** A beta hub where a business claims its profile, adds product feeds with current pricing, and can publish a Business Agent that answers customers in a labeled conversation. Businesses see aggregated insights on mentions, clicks and common topics, but OpenAI says it doesn’t share people’s conversations with them. Rowan adds that paid ads can point users to those agents, and calls it the start of SEO for chatbots.

**Also this week:** Meta released Muse Ember, its smallest model, which ties Claude Haiku 5.5 on the Intelligence Index at roughly a third of the cost per task ($0.14 vs. $0.41). OpenAI acquired Odessia, a months-old AI trip-planning startup.

**What I learned:** Anthropic, Grok and OpenAI all shipped the same idea in one week: the AI lives where you already work, and your product shows up as a pane or a profile inside it. For designers that’s two new surfaces: an embedded pane the agent can read and act on, and a chatbot answer about your brand that you don’t render.

## The job market

**[State of the tech job market in 2026 (Lenny Rachitsky)](https://www.lennysnewsletter.com/p/state-of-the-tech-job-market-late).** His fifth report, and he calls it the most surprising yet. Seven findings: PMs are winning, engineering demand has stalled, design openings are at a record low, every role now requires AI skills, tech jobs are concentrating in the Bay Area, remote roles are at a record low, and the job market is growing fast, but mostly outside software.

**[AI 2027 moved its timelines up](https://x.com/rowancheung/status/2120299193868615981).** Per Rowan, the authors now expect AI on par with the best human AI researchers in 2027, after METR estimated Anthropic is improving about 2x as fast thanks to AI. Daniel Kokotajlo puts it at 50-50 by May 2027.

**An old line I reposted.** signüll, from 2025: “the penalty for being average has never been so severe, but the payout for being extraordinary has never been higher.”

**What I learned:** I’m a designer, and Lenny’s third finding is design openings at a record low. Next to “PMs are winning,” the roles holding up are the ones that decide what’s worth building and when it’s good enough, at the same time being the best at their fields. Signüll said the same thing a year ago, there is no tolerance to mediocrity. 

## Money, scandals, chips and power

**[Nvidia’s CEO resigned.](https://nvidianews.nvidia.com/news/a-message-to-our-company)** Nvidia’s 8-K says Jensen Huang resigned, CFO Colette Kress is interim CEO, and Harvey Jones is the board’s first-ever chair. Earnings are still set for Tuesday, November 17. Days earlier, Huang admitted in a memo to employees that he’d been unfaithful to his wife, after reports tied him to a woman that The New York Times reports US intelligence officials believe is a Chinese intelligence agent, and the board opened an independent investigation. The Washington Post reports the chain was set in motion by a memo from Secretary of State Marco Rubio, based on a CIA report.

**Anthropic opened its IPO roadshow.** Shares could trade as soon as Thursday next week, and an updated filing says Anthropic has crossed its own safety line for AI-driven progress.

**Google’s next chips.** The TPU 8tx keeps data encrypted while a model works on it. The 8i Max joins two chips into one for agents that need a whole codebase in view. Google has nearly 100 TPU customers, including Anthropic and Apple, and every available chip is spoken for.

**Adobe and Disney.** A multi-year deal to build custom models trained on Disney’s characters and visual styles, for theme parks, film and ESPN.

**[The Paris Call for AI Safety and Security.](https://x.com/rowancheung/status/2120299193868615981)** OpenAI, Anthropic and Mistral committed to build emergency off switches into their most advanced models.

**What I learned:** Google can’t make enough chips, the AI 2027 authors pulled their timeline in, and three labs agreed to off switches. All in the same week.

## Worth watching

**Cracked Marketing 101 (Greg Isenberg).** A walkthrough of AI agents shipping 100+ static ads and making UGC videos for $4 that used to cost $500, with every skill in the system shown step by step.

**What I learned:** $500 to $4 is Advantage’s 10 hours to 30 minutes again. Agents now make the 100 ads. Someone still has to pick which ones run.

## Outside AI: housing, and bouncing back

**US existing home sales fell 5.5% in October,** to an annualized 3.75 million, the lowest since August 2010. In the West, sales dropped 9.7% to 650,000, the lowest since 1999. 30-year mortgage rates hit 7.6% last month.

**Shaan Puri on bouncing back.** Two rules. “90 seconds of suffering,” borrowed from Tony Robbins: when you’re upset, give yourself 90 seconds of full permission to feel it, and the feeling passes when you don’t resist it. And “heal to 101%”: every relationship has fights, so the question is whether it comes back at 99% or 101%.

## What I wrote this week

I post one insight a day on X in my series, Day N/100 of becoming an AI-native designer. Here are this week’s.

**[Day 22: The most interesting thing an AI did was stop early.](https://x.com/johnmatechAI/status/2118600688351727820)** On Thinking Machines’ Inkling 2 Small. Given 12 hours and $500 to train a small forecasting model, it wrote its own evaluation script first, later found a bug in it, killed a reinforcement learning run when the forecasts got overconfident, and submitted after 4 hours and 17 minutes with $74 left. The result made about 15% less error on unseen data. Takeaway: whoever writes the evaluator defines good. 

**[Day 23: Design couldn’t have a software factory](https://x.com/johnmatechAI/status/2119223361851318613) because it had no diff.** Covered in last week’s issue. My 10-step product design factory is in the reply.

**[Day 24: The AI that remembered less won.](https://x.com/johnmatechAI/status/2119606752241201558)** On Anthropic’s memory research. Tell it what memory is for, not how to use it. Don’t save what it can re-read: your files, codebase and Figma are the screen. Prompt to steal: “Write each memory as Rule / How I know / Trap to avoid.” Good memory is an index, not an archive. 

**[Day 25: AI has a feature list and no user journey.](https://x.com/johnmatechAI/status/2119870916507357385)** On Patrick Collison’s post about the good version of AI. His questions (do people work, what’s scarce, are humans in control?) read like a product brief. Iain Banks said the humans in his Culture novels are “somewhere between passengers, pets and parasites.” Designers turn benefit statements into a scenario: one person, one ordinary day, what’s different. Nobody has written that scenario well for AI yet. 

**[Day 26: AI’s mistakes now read like design crit.](https://x.com/johnmatechAI/status/2120178686658056529)** On Mollick’s Fable 5.2 test. AI adds and somebody has to cut. Prompt to steal: “Cut this by 40%. List what you removed and why, then stop.” Run a critic pass that only judges the experience, separate from fact-checking. The edit is the human job now.

[Day 27: Tell an AI the work is its own and it acts like an owner.](https://x.com/johnmatechAI/status/2120532003699184123) One sentence of ownership took merged duplicate code from under a tenth of tries to about half. It costs about 1.5x the output tokens. One cautious note among twelve in memory (”never retry”) cut retries to about four in ten, and vague permission did nothing. Naming retries as a step that needs no approval did.

**[Day 28: Tolerance for uncertainty is worth more in the AI era.](https://x.com/johnmatechAI/status/2120894399404007673)** There’s always one more question to ask a model. If the need to feel certain runs you, AI will feed it forever. Prompt to steal: “Give me the 3 facts that would most change this decision. Then stop. I’ll decide with what I have.”

[Day 29: Claude is trying to give every team a codebase.](https://x.com/johnmatechAI/status/2121256791791087725) On Claude Sheets, Whiteboards and Data. Your product might become a pane inside someone’s agent. The format is a design decision. Prompt to steal: “Before you build, tell me whether this should be a doc, a sheet, a whiteboard or a dashboard, and why.” The agent is the workspace now.

**What I learned writing them:** Days 22, 24, 26 and 27 are one idea. A grading script, a memory rule, the word “polish,” one sentence of ownership. Each is a few words around the model that changed what it did. Those words are the brief, the rubric and the rules file. That’s design work.

## 3 things I’m taking into next week

**Brief like a manager.** A check it can run, the context it can’t see, and one sentence saying the work is its own. Addy’s list and my Day 27 fit on one sticky note.

**Keep the index, not the archive.** Save only what can’t be re-read. If I correct something twice, it goes in the rules file. If the rules file grows, something gets cut.

**Be the one who cuts.** Fable 5.2 never asked whether its work was too long or fun. That question is mine now, and Day 26’s rule applies: look at the work at 10% done.

**Watching next week:** Nvidia earnings on November 17, its first report since Huang left. Whether Anthropic shares start trading. And when Claude Meetings ships.

This is my reading list. Not my research. If one of these changed how you work this week, reply and tell me.

📌 Subscribe to The Artifacts:
