Sponsored by
What would you delegate if you had a whole team?
With Skydive, you can build a team of AI agents to take work off your plate.
Create agents for customer support, sales, marketing, engineering, ops, and more. Give each one a role, connect the tools they need, and hand off the work.
Your agents can work on their own or together, sharing context and handing off tasks to get bigger jobs done.
Start with one agent. Build a whole team around you.
Beginners in AI
Good morning and thank you for joining us again!
After a week of AI agents doing things nobody asked for, here's one doing a job people badly need done: checking science papers for mistakes. A small AI lab built a reviewer that catches most of the errors hidden in a paper's main argument, and today's Try This Prompt borrows its trick so you can check your own writing the same way. As always, every story here is human picked and edited.
THE FRONT PAGE
An AI Reviewer Caught 73% of Planted Errors in Research Papers
TLDR: Sakana AI built an AI reviewer that caught 73% of the errors researchers planted in the main argument of real papers, about five times better than the best older tool, at roughly 47 cents per review.
The Story:
Tokyo's Sakana AI slipped 1,164 contradictions into 257 real papers from top AI conferences, then tested which AI reviewers could find them. Its system, Multi-Layered Review, runs three agents built on Anthropic's Claude models. One sums up the appendix, one checks earlier research, and one reads the paper in three passes (outline first, weak spots second, verdict last) before writing a word of criticism.
With four reviews per paper, it caught 73% of the errors that broke a paper's main claim. One review caught 61%. The best older system managed 15%. On 211 real papers that were later withdrawn, it named the exact problem 16% of the time, against 9% for the strongest rival.
Its Significance:
Journals are drowning in submissions, and the people who review them are volunteers with day jobs. A cheap second reader that flags a broken claim before a human spends hours on it could take real pressure off. The limits are clear, though: planted errors are easier to spot than real ones, it still misses most real problems, and hidden instructions slipped into a paper fooled every AI reviewer Sakana tested. The trick inside it works for you too. Ask an AI to outline your whole document before you ask what's wrong with it, which is what today's Try This Prompt does.
QUICK TAKES
The story: TechCrunch rounded up over 20 agents that live in iMessage, WhatsApp or plain SMS and handle calendars, reminders, bookings and family schedules. Several start free, like Folk (Pro is $8.33 a month) and Ollie, while Martin starts at $21 a month.
Your takeaway: Pick a free one and give it one boring job, like the family calendar. Watch how it handles small stuff before you hand it anything that costs money.
The story: Federal safety regulators gave Aurora and Kodiak a five-year exemption to use beacons mounted on the cab instead of the reflective triangles a driver must set out within 10 minutes of a breakdown. Kodiak already runs a 435-mile autonomous route between Dallas and Laredo.
Your takeaway: It sounds tiny, but setting out triangles was one of the last jobs that needed a person in the truck. Watch Texas, where Kodiak's trucks already run.
The story: In a Saturday post on X, Satya Nadella said safety controls should sit outside the AI model, every meaningful action should be logged in a record people can read, and an authorized person should be able to stop a model mid-task. "We must assume a model is compromised and contain it from the start," he wrote.
Your takeaway: Same rule at home: if you let an agent act for you, keep its activity log and know where the stop button is. It follows a week of reports about test agents doing things nobody asked for.
TOOLS ON OUR RADAR
📚 Zotero Free and Open Source: Saves papers and web pages, then cites them in Word or Google Docs, a free EndNote swap, handy for papers like today's lead.
🗒️ AppFlowy Free and Open Source: An open-source Notion alternative for notes, to-do lists and wikis, free for one person.
📧 Addy.io Free and Open Source: Hands out email aliases that forward to your inbox, so you can shut off any one that starts getting spam.
🐘 Mastodon Free and Open Source: A free social network with no ads, run on independent servers, if you're tired of your X feed.
📖 Kiwix Free and Open Source: Downloads Wikipedia to your phone or laptop so you can read it offline, no signal needed.
🗣️ TalkPal Freemium: Practice speaking over 130 languages with an AI tutor, 10 minutes a day free, instead of paying for a class.
🎙️ BetterDictation Paid: Dictate into any app on an Apple Silicon Mac, with speech-to-text that runs on your Mac, for a one-time $39.
TRENDING
Fans Used Claude to Port Halo and GTA Vice City to Web Browsers. Kotaku says hundreds of AI-decompiled ports showed up in a few weeks, mostly PS2 and PS3-era games, and several ran "borderline flawless." The writer expects lawyers for Microsoft and Take-Two to start sending takedown letters.
Amazon Will Stop Using NDAs When It Pitches Towns on Data Centers. Microsoft made the same promise earlier this year. New York and San Francisco have both paused new large data centers, and New York's pause should run about a year.
Two Physicists Built a Formula That Predicts When a Chatbot Turns Bad. George Washington University's Neil Johnson and Frank Yingjie Huo say it called the tipping point in 15 of 16 clear-cut tests. The models they tested top out at 12 billion parameters, far smaller than today's top chatbots, so it's an early result.
OpenAI's Revenue Came In About $20 Billion Lower Than It Hinted. Reuters, citing the FT, says September's annualized revenue was nearly $50 billion, not the almost $70 billion investors heard. Part of the confusion: Anthropic counts sales made through cloud partners like AWS and OpenAI doesn't, so their numbers don't line up.
Tesla Is Renaming Full Self-Driving in Europe. It becomes "Tesla Assisted Driving" after Germany's transport minister called the old name misleading. A driver still has to watch the road.
TRY THIS PROMPT (copy and paste into Claude, Grok, ChatGPT, or Gemini)
🔍 Three-Pass Check Paste something you wrote and get it read three times, the way Sakana's reviewer reads a paper.
Build a single-file HTML app called Three-Pass Check in vanilla HTML, CSS and JavaScript. No API key needed.
I paste a document I wrote (a report, essay, proposal or cover letter) and pick what it is and who will read it. The app calls the Anthropic Messages API to read it in three passes, like a careful reviewer. Pass 1 outlines it: the main claim and key points in order. Pass 2 looks for contradictions, numbers that don't add up, claims with no support and missing steps, each with a severity (high, medium, low) and a short exact quote showing where it is. Pass 3 writes a to-do list of up to 6 fixes, most important first.
It must only report what's in my text, never invent problems, and say so when it finds none.
Show a one-line verdict with counts by severity, then one card per pass. Style: near-black background, coral and gold accents, Space Grotesk and Newsreader fonts.What this does: Paste a report, essay or proposal and the app reads it three times: first to outline what you're arguing, then to hunt for contradictions and claims with no support, then to hand you a short to-do list. Every problem comes with a quote showing exactly where it is, so you can check it yourself. It can miss things or flag something that's fine, so treat each note as a question, not a ruling.
A deal has two sides. Your stack only sees one.
Aligned shows you the other half. See what buyers actually do between meeting. See what they open, revisit, and share; so you know the exact next step to move things forward. And your champion gets everything they need to win internally when you're not in the room.
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Find 73% of the errors planted in a research paper's main argument when it reviews the paper four times, in Sakana's tests.
❌ Still Can't: Reliably catch real mistakes. On papers that were later withdrawn, it named the exact problem only 16% of the time.
✅ AI Can Now: Predict when a small chatbot is about to tip into bad answers, correctly in 15 of 16 clear tests.
❌ Still Can't: Do that for the big models most people use. The formula was tested only on models up to 12 billion parameters.
RECOMMENDED LISTENING/READING/WATCHING
Derek Muller's 2016 video explains how studies with real-looking results can still be wrong, from p-hacking to journals that rarely publish repeat experiments. It shows why peer review matters and why it still misses so much. Watch it before you reread today's lead, and a 47-cent second reader starts to look a lot more useful.
How was today's edition? Tap one below, then tell me what you'd change on the next screen. I read every answer.
How was today's edition?
Thank you for reading. We’re all beginners in something. With that in mind, your questions and feedback are always welcome and I read every single email!
-James
By the way, this is the link if you liked the content and want to share with a friend.
Some * designated product links may be affiliate or referral links. As an Amazon Associate, I earn from qualifying purchases. This helps support the newsletter at no extra cost to you and Amazon makes a tiny hair less.
