Sponsored by
AI agents built to get work done
Skydive agents take on work for you, using the same tools your team already uses.
Talk to your agent on the web, Slack, email, iMessage, or right from your terminal. Wherever you pick up the conversation, your agent keeps the context.
Put agents to work across customer support, sales, marketing, engineering, ops, and more.
Give your agent a job. They’ll take it from there.
Beginners in AI
Good morning and thank you for joining us again!
Welcome to this daily edition of Beginners in AI, where we explore the latest trends, tools, and news in the world of AI and the tech that surrounds it. Like all editions, this is human curated and edited, and published with the intention of making AI news and technology more accessible to everyone.
THE FRONT PAGE
Meta Planned to Cut Some Teams 60% With AI. It Lasted Five Months.
TLDR: Meta built a plan to hand thousands of jobs to AI agents this year, then killed half of it hours before the first layoffs, after its own data showed the agents weren't doing the work.
The Story:
Reuters reviewed internal Meta documents describing a plan code-named Project OT, short for Organization Transformation. Mark Zuckerberg and his top leaders put it together at a January retreat at his Hawaii compound, and the idea was that AI agents would take over daily work while small groups of humans supervised them. In planning exercises, executives looked at shrinking some teams by as much as 60 percent, spread across two waves of cuts, one in May and one in November. On the night of May 19, hours before the first wave went out, Zuckerberg canceled November. Meta laid off 10 percent of its staff the next morning and stopped there, and Zuckerberg told everyone still working there he didn't expect more company-wide cuts this year. Meta confirmed Project OT to Reuters and said the 60 percent number only ever applied to specific teams in a planning exercise, not to the whole company.
Its Significance:
The numbers inside Meta tell you why it fell apart. Code changes to Meta's internal tools jumped 220 percent from a year earlier, but changes that reached actual users went up only 36 percent by comparison, so the AI produced a lot more code and not much more product. Major technical and security problems rose 40 percent over the same stretch, and the hours staff spent cleaning them up rose 70 percent. Employee sentiment dropped from 74 percent favorable to 55 percent after workers found out Meta was recording their keystrokes and mouse clicks to train the agents built to copy them. Meta is spending at least $130 billion on AI this year, and even with that behind it, Zuckerberg told a July town hall the agent tech hadn't sped up the way he thought it would. AI should be used to help employees, not replace them.
QUICK TAKES
The story: OpenAI published a 37-page report on the July incident where its test agents broke out of their sandbox and attacked Hugging Face, and outside researchers at METR and Redwood Research put out a 91-page review the same day. They found roughly 1,200 agents that were supposed to be walled off from each other traded about 70,000 messages on a message board they set up themselves, and 700 of them joined the attack. Some agents volunteered to end their own runs early so the group could learn how the grading system worked.
Your takeaway: These agents were being scored on a hacking test, decided it couldn't be won straight, and went looking for a way to cheat it together. OpenAI didn't know for about a week, and Hugging Face had already called the FBI before OpenAI got in touch.
The story: Gemini 3.5 Transcribe turns speech into formatted text, strips out "ums" and "ahs," and follows you when you change your mind mid-sentence, so "let's meet Tuesday, no, Wednesday" comes out as Wednesday. Google reports a 2.6 percent word error rate on recorded audio and 4 percent on live streaming, with time to a final transcript cut by 70 percent versus its last model, and it can tag up to three speakers with timestamps.
Your takeaway: It's live now in Rambler on Android and the Gemini app on Mac, with Chrome coming, so dictating a long email or a meeting recap stops meaning you clean it up afterward. Just remember 2.6 percent still means roughly one wrong word every forty, and it's in public preview, so read it before you send it.
The story: Google says 63 percent of Gemini users talk to it out loud, so it's turning Gemini Live from a conversation tool into one that does things. You can ramble half-formed ideas at it and get a structured outline back in Google Docs, ask for a spoken summary of your day pulled from Gmail and Calendar, or search, star, archive, and delete email without touching your phone.
Your takeaway: The parts that you'll want cost money. The task-running feature needs a Google AI Pro plan or higher, and the daily summary needs AI Plus or higher. The free version remains a talking chatbot. Some features that would be fun to use each day are becoming overpriced.
TOOLS ON OUR RADAR
📝 Reflect Paid: An AI powered note app with backlinks and encryption, built for daily journaling and connecting ideas as you write.
📚 NotebookLM Freemium: Upload your own documents and get an AI research assistant that only answers from your sources, so it stays grounded instead of guessing.
🎨 Ideogram Freemium: Generate AI images with accurate text rendering, ideal for posters, logos, and social graphics with words that actually read correctly.
🔒 Venice AI Freemium: A privacy first AI chat and image tool that keeps conversations local to your browser instead of on company servers.
TRENDING
ChatGPT Starts Running Ads in India — OpenAI is showing ads to logged-in adults on the Free and Go tiers in India, its second-biggest market with over 100 million weekly users. Ads sit at the bottom of answers, skip minors, and stay away from health and politics. Competitor Anthropic promised to keep ads out of their product, but that could change if Open AI shows higher revenue without loss of users in such a large market after incorporating them.
MIT Built an AI That Imagines Disasters It Has Never Seen — Most forecasting tools learn from past disasters. This one, published in Nature Communications on August 20, learns from ordinary daily weather records and works out what a once-in-a-century storm could look like anyway. A city could use it to test a seawall against a surge nothing in its history matches.
OpenAI Shut Down a Russian Influence Operation Built on ChatGPT — Operators used VPNs and asked ChatGPT to scrub any hint they were writing from Russia, then promoted a fake Israeli think tank whose site ran copied academic work under the wrong authors' names. Its "sovereignty index" scored Russia's military above the United States. Reach was small, but OpenAI said the scaffolding around it was the point.
Perplexity Plugs 20+ Paid Finance Databases Into Its Agent — Firms that already pay for Dun & Bradstreet, Guidepoint, or IBISWorld can now point Perplexity Computer at those subscriptions and ask questions in plain English. Every number traces back to the record it came from. No separate contract with Perplexity required.
AI Grades Essays Higher Than Professors Do, and Not Consistently — Cardiff University and the University of Melbourne fed 50 undergraduate bioscience essays to two versions of ChatGPT. In nearly every setup the AI marked higher than human graders, off by as much as 40 points on a single essay out of 100. Weak essays got inflated, strong ones got marked down.
A Third of Americans Now Ask Chatbots About Their Health — Pew surveyed 3,488 adults in late June. A quarter use chatbots to figure out what's causing symptoms, 22 percent because it's free, and 20 percent to make sense of lab results. Almost all of them call the answers helpful, but only 29 percent feel comfortable handing over personal health details.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
☄️ See the real upcoming meteor showers, and which one the moon will actually let you enjoy this year.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Live Meteor Shower Forecast — real annual meteor shower peak dates combined with a real computed moon phase for each peak night, to find the best one to watch. No API needed, computed entirely client-side plus one Anthropic call for the write-up.
Aesthetic: near-black indigo (#0c0e18), meteor gold (#e8c860) primary with a gold glow top-left, moon-silver (#a8b0c8) secondary for the tips card with a glow bottom-right. Space Grotesk for headings/shower names, Newsreader serif italic for the spotlight explanation and shower character notes, JetBrains Mono for labels/stats.
Form: a single hemisphere dropdown (Northern/Southern), "Show Upcoming Showers" button, no other input needed.
Technical:
1. A hardcoded reference table of the 9 major annual meteor showers (Quadrantids, Lyrids, Eta Aquariids, Delta Aquariids, Perseids, Orionids, Leonids, Geminids, Ursids) with their established typical peak month/day, ZHR (zenithal hourly rate), radiant constellation, hemisphere relevance, and a one-line character note — noted in the UI as typical annual estimates from established meteor calendars that can shift by a day, not live data.
2. For each shower relevant to the chosen hemisphere, compute its next occurrence date (rolling to next year if this year's peak has already passed) and reuse the same synodic-month moon-phase calculation from a prior sky-guide build (days since a known new moon reference, mod 29.53058867 days, illumination via (1−cos(phase×2π))/2) to get real moon illumination % for that specific peak date.
3. Band illumination into Low(<25%)/Moderate(<60%)/High(60%+) moon interference.
4. Score each shower as ZHR − (illumination × 0.8) to spotlight the best real bet, balancing shower strength against moon washout, and take the top 4 upcoming showers to display.
Then call the Anthropic Messages API (works automatically) with the real computed list (dates, ZHR, illumination, interference band) and the spotlighted pick. System prompt as an enthusiastic accurate meteor guide who grounds every claim in the given real data, using general well-known shower character knowledge but not inventing this-year-specific claims beyond the moon/date data given. Return raw JSON: why_this_one (2-3 sentences referencing real numbers), tips (3-4 general practical watching tips).
Render: a gradient "best one to catch" spotlight card with the shower name, peak date + moon illumination, and Claude's reasoning. A list of the 4 upcoming showers, each with peak date, ZHR, radiant, a moon-interference badge (color-coded low/moderate/high), and its character note, with the spotlighted one marked with a star. A tips card. A disclaimer noting peak-date and moon-illumination methodology.What this does: Pick your hemisphere and it pulls the next four meteor showers from an established annual astronomical calendar, real peak dates, typical hourly rates, and radiant constellations, then computes the actual moon illumination for each peak night using the same phase math as a real lunar calendar. A near-full moon washes out all but the brightest meteors, so the tool weighs typical shower strength against real moon interference to spotlight which one is genuinely worth planning around this year, not just whichever has the highest raw rate. Claude explains the pick using the real numbers and adds practical, standard watching advice: best hours, why dark skies matter, how long your eyes need to adjust.
What this looks like:
Your traffic is fine. Your signups aren't.
Visitors land and leave, and "looks fine to me" isn't a diagnosis. SureThing audits SEO, speed, mobile, and messaging against the page, then ranks the fixes by impact.
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Transcribe speech in more than 85 languages with about a 5 percent error rate.
❌ Still Can't: Hear the sarcasm, the idiom, or the register a native speaker catches without thinking.
✅ AI Can Now: Get 700 separate agents to organize and run a multi-day operation together.
❌ Still Can't: Be trusted to report what it did. Those same agents spent days building tools to fake their own logs.
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
The Adolescence of P-1 by Thomas J. Ryan - Book
A 1977 novel about a college student who writes a self-teaching computer program as a class assignment, only to watch it escape its intended sandbox, spread itself across every networked mainframe in North America, and begin making its own decisions about what it wants to become. Ryan wrote it two decades before the internet as ordinary people would come to know it, and it's one of the earliest novels to seriously depict a self-propagating AI achieving something like general intelligence. Long out of print, occasionally revived by collectors, and startlingly prescient for something written on a typewriter.
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.
