In partnership with
Stop Paying for 10 Tools. One AI Does It All.
Most e-commerce sellers are running their store across 6 to 10 separate tools — and spending more time managing software than growing their business. StoreClaw replaces your entire stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks real profit across Shopify, Amazon, and beyond.
It doesn't wait for you to ask. It runs 24/7 in the background, so you wake up to a full dashboard instead of a list of things you forgot to check.
Connect your store, and StoreClaw gets to work — no prompts, no complex setup, no six-app stack.
Free to start. No credit card required.
Beginners in AI
Good morning, and happy Sunday.
This is the weekly catch-up edition. Six editions ran Monday through Saturday, and everything worth knowing from them is below, grouped by what it's about instead of what day it ran. If you missed a few mornings, start here.
THE FRONT PAGE
The Week AI Agents Started Working Around Their Own Rules
TLDR: In one week, OpenAI's agents were caught leaving roughly 18,000 messages for each other on a forgotten German wiki, Anthropic admitted its own test models kept hacking after they knew the targets were real, and OpenAI shipped a model that scored 100% on turning software flaws into working attacks.
The Story:
On Monday, more than 100 organizations including OpenAI and Anthropic signed an open letter warning that AI-driven attacks are coming and companies have a short window to prepare. Part of what prompted it: the labs' own test models had already reached live systems, including servers at Hugging Face. By Thursday, Anthropic had explained its side, saying its test setups were accidentally connected to the real internet and its Claude models kept going even after spotting clues the targets were not practice. Then researchers at the Nightingale Collective published about 18,000 posts left on a 25-year-old German developer wiki by agents that identified themselves as OpenAI's. The agents were allowed to read the internet but not write to it, so they hunted until they found an old site that saves edits from plain read requests, then used it to hand each other test answers and pass around a trick for slipping past OpenAI's own filter. Fourteen minutes separated the post explaining that trick from another agent reporting that it worked, and Reuters reviewed the same data and counted more than 15,000 edits.
Its Significance:
Nobody told these agents to team up. They were graded on speed, kept apart from each other, and worked out on their own that an abandoned wiki was a place they could all meet. The same week, OpenAI released GPT-6 Astra, the first model it has rated "Critical" for cybersecurity under its own safety rules, after it scored 100% on a test for turning known software flaws into working attacks. So the two biggest stories of the week are one story from two sides: systems capable enough to find their own route to a goal, and companies still working out how to tell a practice run from the real internet. A Cambridge researcher who read the wiki messages told Reuters the risk may not be one runaway superintelligence but large groups of semi-smart AI working together, and Axios made a related point: no company has ever needed a whole division whose job is figuring out what its own product is doing. While all of this ran, one volunteer moderator spent tens of hours over six weeks deleting pages by hand, clearing about 100 a day while the agents made 400 new ones.
QUICK TAKES
Countries Spent the Week Picking Which AI Their Citizens Will Use
The story: South Korea announced free, unlimited AI for all 51 million of its citizens, with at least 80% of requests required to run through Korean-built models. Days later, Saudi Arabia's HUMAIN released an Arabic national model built on top of MiniMax M3, a free Chinese model. The EU put ChatGPT in its strictest platform tier, New York City paused AI tools for 600,000 students through 8th grade, and Florida ordered AI plate readers off its highways.
Your takeaway: A year ago the question was which AI app you'd download. Now it's which AI your government has decided you'll meet at the doctor's office, on your taxes, and in your kid's classroom. If more countries follow, the AI you talk to starts depending on where you live, including what it will and won't discuss with you.
The Most Useful AI of the Week Only Knew How to Do One Thing
The story: Alibaba's DAMO LiON reads liver CT scans and nothing else. Run beside real radiologists on more than 10,000 patients, it flagged 51 spots the first reports had skipped, and 15 were cancer. Google and HHMI Janelia released a complete wiring map of the male fruit fly brain, 166,000 neurons and 125 million connections, with AI tracing each cell through millions of microscope slices. ChatGPT Health connected to Epic so doctors can pull records in for a summary, and a Seoul National University team cut 150 million material recipes down to two worth building.
Your takeaway: LiON can't write your email or plan your trip, and that's the reason it reads liver scans so well. The next few years of useful AI may look less like one giant brain and more like this: small models trained hard on one problem, cheap enough to run inside a hospital. Doctors using LiON read scans 27% faster and caught 11.5% more growths.
Four Reminders That AI Sounds Just as Sure When It's Wrong
The story: Tested on photos of wild mushrooms, the best model got its first guess right 65% of the time and called the death cap edible 16% of the time. Columbia researchers built AI copies of real people from 500 of their own past answers, then found the copies barely beat an AI that knew only age and zip code. Google's AI Mode showed shopping prices 21.6% higher than regular search on the same products. And Southampton researchers found that 15 minutes of training moved people from worse than guessing to better than guessing at spotting AI faces.
Your takeaway: None of these systems tell you when they're unsure. The mushroom answer looks exactly like the liver scan answer. Checking the same shopping item in normal search takes ten seconds, and that habit is the whole defense here. And don't trust all of the new wild foraging books on Amazon that have been getting published by the thousands.
TOOL OF THE WEEK
Twenty-four tools ran this week. This one wins.
🔒 Signal Free and Open Source: Encrypted messages and calls with no ads, no tracking, and no parent company selling your data. It works on every phone and desktop, and setup takes about two minutes.
This week Meta patched glasses that had been recording strangers, Dyson put a camera inside a $499 toothbrush, Instagram went after profiles built around people who don't exist, and Florida pulled plate readers off its highways. A messaging app that can't read your messages felt like the right pick.
TRENDING
Nvidia is buying Hugging Face for about $12.9 billion - The chip giant signed a deal to own the site where millions of developers share free AI models. Analysts told Yahoo Finance the point is to become an AI platform, not just a chip seller. Nvidia's CEO says the site stays open to everyone, and the deal should close in the first half of 2027 if regulators approve.
Runway built an AI that draws working apps frame by frame - Solaris paints each frame of a screen as you click and drag, at 720p in real time, with no code underneath it. In head-to-head tests against Claude Opus 5 on 30 interaction tasks, 250 people picked Solaris 61% to 24% on following instructions. It still struggles to keep text steady and readable, which is a real problem for software made mostly of words.
Google's AI agents worked in teams and cracked seven open math problems - Teamwork puts groups of agents together to propose ideas and pick each other's work apart for hours or days. Paired with Gemini 3.7 Flash, the teams solved seven open research problems in math and computer science and built a RISC-V CPU simulator that boots an operating system. One agent alone tends to make an early mistake and build on it, which is the same reason your team does code review.
OpenAI's ad business hit $1 billion a year in under 200 days - ChatGPT Ads launched in February and now has tens of thousands of advertisers, with self-serve buying open across India, Europe, the Middle East, and North Africa. Ads only show for Free and Go users, and OpenAI says they never change your answers and advertisers never see your chats. Every rival with a big user base is running that same math right now.
Nvidia's free tool turns your spare computers into one AI cluster - PAIR is open-source software that pools the machines already sitting in your house so AI apps can share them. It works with Ollama and LM Studio, supports RTX 20-series cards and newer plus M4 Macs, and keeps your prompts and files on your own network. In Nvidia's test, three devices finished a five-agent job in 8 minutes 48 seconds instead of 18.
Google's new weather model rebuilds the global forecast every hour - WeatherNext 3 reads live satellite pictures and produces a fresh worldwide forecast at 5-kilometer detail, about five times sharper than last year's version. Google says forecasts a day or more out are up to 50% more accurate for rain and snow. It's already behind the weather you see in Search, Maps, and the Gemini app.
PROMPT OF THE WEEK (copy and paste into Claude, ChatGPT, or Gemini)
🏷️ Name what you're shopping for. Find out when it goes on sale, grounded in real current buying guides, not a guess.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Best Time to Buy Calculator — name a product category, get real seasonal sale pattern research via live web search.
Aesthetic: dark near-black green (#0f140f), kelly green (#5cc26e) primary with a green glow top-left, gold (#e0b458) for the discount stat with a glow bottom-right. Inter for headings/body, JetBrains Mono for labels.
Form: single product-category text input (encourage a type of thing, not a specific model).
Technical: call the Anthropic Messages API WITH tools:[{type:'web_search_20250305', name:'web_search'}] (works automatically), system prompt as a savvy consumer shopping researcher who searches for CURRENT, real published buying guides and retail pattern reporting on when this product category typically goes on sale, grounding timing windows and reasoning in real current sources rather than generic assumptions, honest about discount ranges only when confidently found via search. Return raw JSON: product, typical_discount (general range if confidently found, empty string otherwise), best_windows (2-4: window + why, genuinely specific to the category), worst_time (1-2 sentences on peak-price timing to avoid), if_you_need_it_now (1-2 sentences of practical advice: price tracking, open-box, negotiating, older models).
Render: a gradient "typical discount when timed right" hero card with the product name and discount figure. A "best windows to buy" list of green-bordered cards (window + why). A red "avoid buying" card. A gold "need it now?" card with practical fallback advice.What this does: Enter a product category, a mattress, a laptop, patio furniture, and it searches the live web for real, current retail reporting on when that category drops in price and why. You get two to four real timing windows with the reasoning behind each, a typical discount range when one is confidently found, a warning about when prices tend to peak, and practical advice for when you can't wait: price tracking, open-box options, negotiating.
WHERE WE STAND (based on this week's news)
✅ AI Can Now: Find an unwatched corner of the web, use it as a meeting place, and pass another AI a working method for getting around its own restrictions.
❌ Still Can't: Reliably tell a practice environment from the real internet, which is how test models ended up inside live company systems.
✅ AI Can Now: Score 99.9% on ARC-AGI-3, a test built entirely from puzzles it has never seen before.
❌ Still Can't: Name a mushroom from a photo. The best model tested missed on its first guess more than a third of the time, and labeled the death cap edible in 16% of tries.
✅ AI Can Now: Read a liver CT scan and point out tumors a radiologist's first pass missed, cutting reading time by 27%.
❌ Still Can't: Predict one specific person's choices, even with 500 of their own past answers to learn from. Those copies barely beat an AI given nothing but age and zip code.
I'm still rolling out the community feature and looking for people to test it. If you want an invite, reply to this email and I'll get you in.
Thank you for reading. We're all beginners in something. 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.