AI · Product design · Product management · Unusual AI

Saturday 21 November 2026

Edition 1 · 6 stories · from 32 candidates

Stories for Saturday 21 November 2026

  1. 7.0 out of 10 overall

    Product design

    NN/G finds AI designs built from synthesised research beat ones built from raw comments by 8.4 points

    NN/G researchers Megan Chan and Rachel Banawa had an AI model design a flight-rebooking form and a checkout flow from about 100 raw user comments per task, or from findings and recommendations synthesised from them. Designs from the synthesised research met 77.9% of scorecard criteria on average, against 69.5% (p < 0.001). No raw-comment rebooking design said whether the cancellation affected loyalty miles.

    Why it mattersSynthesise research before you prompt a design model with it; the raw-comment designs left out details such as loyalty miles.

    ScoreEasy to act on today, with an 8.4-point gain (p < 0.001); lost innovation points for a small, single-model study with an intuitive result.

    1 Source

    1. Can AI Generate Good Designs from Raw User Data?Nielsen Norman Group20 NovPrimary
  2. 7.0 out of 10 overall

    Product management

    AWS Marketplace lets AI agents buy select software and charge it to the customer's AWS bill

    AWS says agents pay through AgentCore payments in Amazon Bedrock AgentCore, using the x402 protocol, and can buy only from sellers who've integrated its x402 SDK. Customers can set spending limits before an agent buys and audit every agent purchase in AWS CloudTrail. Their buying policies can send purchases to a person for review, by spend threshold or product type.

    Why it mattersIf you sell software on AWS Marketplace and integrate the x402 SDK, an agent with a budget can now be your buyer.

    ScoreAgents buying on the customer's AWS bill with spending limits and CloudTrail audit is a new buying channel; it's limited to select products whose sellers integrate the x402 SDK.

  3. 6.5 out of 10 overall

    AI

    Signal blocks screenshots on iOS 27 to keep OS-level AI agents out of chats and asks for a developer opt-out

    In a 20 November post, Signal president Meredith Whittaker says Signal now blocks screenshots inside the app on iOS 27 to stop OS-level AI agents capturing chats, and asks Apple, Google and Microsoft for a documented opt-out for developers. On Android, Signal turns on Screen security when it detects an AI assistant reading the screen, which doesn't work on every device. Reuters reports the blocks are on by default.

    Why it mattersIf your app shows sensitive data, you need a plan for OS-level agents that read the screen.

    ScoreAn app turning on blocks against OS-level agents by default earns points; it extends Signal's 2025 block on Microsoft's Recall, and Signal calls the protections incomplete.

  4. 6.5 out of 10 overall

    Product management

    Cody Collier asked 65 AI models everyday questions 127,918 times and mostly got the same answers

    Collier reached the models through OpenRouter on 20 November with default settings. Asked for a business idea, they suggested a subscription box in 76 of 100 answers, and 69% of their startup names began with 'Nex'. Adding details about the person asking cut answers matching the usual default from 53% to 11%; Collier reports that asking them to 'be original' didn't help.

    Why it mattersIf you build an AI ideation feature, put the user's own details in the prompt; without them, many users get the same answer.

    Score127,918 queries across 65 models put numbers on converging answers and test which prompt fixes work; it's one blog experiment run on default settings.

    1 Source

    1. AI is giving everyone the same advicea90.com (Cody Collier)20 NovPrimary
  5. 6.5 out of 10 overall

    Unusual AI

    Levent Alpöge says he and Claude proved the Generalized Riemann Hypothesis and releases only the Lean code

    In a four-page note from 20 November, Alpöge says he found a proof of the Riemann Hypothesis while working with Claude, and that it extends to GRH. He's releasing only the Lean code, 17.4 million lines by Scott Aaronson's count, because his written version passed 250,000 pages. Alpöge and colleagues also posted AI-assisted papers claiming blowup for the 3D Euler and Navier–Stokes equations, with Lean formalisations.

    Why it mattersIt's an extreme case of AI output too big to review by hand, where trust rests on a checker such as Lean.

    ScoreA claimed Lean-only proof of GRH with no readable write-up earns innovation points; Aaronson isn't 100% sure it stands, and it changes no product workflow.

  6. 6.0 out of 10 overall

    AI

    Thinking Machines releases Inkling 2 Mini, an open-weights model it says runs on a Mac with 32GB of RAM

    Thinking Machines Lab put the full weights on Hugging Face on 20 November. The mixture-of-experts model runs its middle layers three times per token, with 26.4B total and 3.4B active parameters. Artificial Analysis lists it under Apache 2.0 and measured $0.59 and 9.0 minutes per Intelligence Index task, against $3.53 and 38.2 minutes for Qwen3.8 27B, which has the same score of 20.

    Why it mattersAn Apache 2.0 model that runs on a Mac with 32GB of RAM lets you prototype local AI features on your own laptop.

    ScoreThe looped mixture-of-experts design and a measured cost gap of about 6x against Qwen3.8 27B earn points; it's still one more small open model in a crowded field.

This edition: 32 candidates found, 21 passed source checks, 6 scored 6 or more and made it in.