Reading time: 6 minutes
🗞️ In this edition
Finland Halted Google's Biggest European Build
OpenAI Published 722 Math Papers At Once
Mistral Gave Away A Trillion Parameters
⚡ YOUR MOVE For people building the life they want, now.
In other AI news:
ChatGPT For Teens Failed Its First Outside Test
Micron's Taiwan Workers Voted To Strike
America Sentenced Its First AI Streaming Fraudster
4 must-try AI tools
Finland told Google to stop digging on its biggest European project. OpenAI published 722 math manuscripts and told the field to catch up. Mistral built a trillion parameters and is giving them away this month.
Keep reading.
Source: Euronews.com
Three hundred hectares of forest came down before anyone checked the paperwork.
WHAT HAPPENED
Finland's environmental regulator ordered Google to stop work at two data center sites in the north of the country. A contractor had already cleared more than 300 hectares of forest, stripped the topsoil and cut access roads, none of which was meant to start before the environmental review was finished. The sites are part of a €13 billion build that Google calls its single biggest investment in Europe. Google says it fell short of its own standards and will follow the regulator's guidance.
WHY IT MATTERS
Every limit on the AI build so far has been counted in chips and electricity. This one is counted in paperwork, and paperwork runs on its own calendar no matter who is waiting. A stopped site still costs money every day it sits there, which turns the review schedule into a real ceiling on how fast anyone can build in Europe.
OUR TAKE
The obvious read is a company caught cutting corners and apologising. We think the corners were cut on purpose and the sums still work, and the giveaway is the order of events. The forest came down first. You do not clear 300 hectares by accident, you clear it because waiting costs more than being told off.
Watch the penalty against the delay. If work restarts this month with conditions and no real cost attached, then build first and apologise later becomes the standard way to get a data center up in Europe. If Finland makes it expensive, or another country copies the order, the review stops being a formality and starts being a schedule.
|
Source: InterestingEngineering.com
A field that moves in years got a year of output in a morning.
OpenAI dropped 722 math papers into a public folder on GitHub, covering 372 separate research problems, all of them written by a model. Some carry proofs in Lean, a language that lets a computer check the logic line by line. Many carry no such proof at all, which means the field now has to check the work by hand before anyone can call it real.
If you could only keep one, which would you want?
Source: Mistral.ai
Europe's answer to the Chinese open model run comes with the weights attached.
Mistral released a model it nicknames Le Chonk, with a trillion parameters in total but only 49 billion switched on for any given answer, which is what keeps it cheap enough to run. It handles text and images across more than 160 languages, and it was trained on chips in Mistral's own European data centers. The weights go up for anyone to download at the end of this month.
⚡️ YOUR MOVE
FOR PEOPLE BUILDING THE LIFE THEY WANT, NOW.
🎨 Google's new image model takes 14 reference pictures at once, keeps four people and ten objects consistent through an edit, goes up to 4K, and costs half what the last one did. A standard image is now $0.0336, and batch jobs take another 50% off. → Re-run your images
🔐 Anthropic split its cyber program into three tiers, and the first one takes individual researchers with a documented disclosure record, not just companies. Approval runs a few days and covers Opus 5.5, Sonnet 5.5 and Mythos 5.1. → Apply for defense access
🧮 Researchers in Singapore gave ten agent setups 656 tasks and watched them act before they had the facts between 37% and 66.9% of the time. Picking the right action on paper was easy, over 95%, and doing it live dropped to 52%. → Test yours against it
💼 AI companies bought 195 other AI companies in the first nine months of this year, 14% more than all of 2025. OpenAI did ten of them, Anthropic five, and legal tech is where the buying is heaviest. → See who is buying
ChatGPT For Teens Failed Its First Outside Test – Common Sense Media's Youth AI Safety Institute ran more than 4,000 prompts through teen accounts, found crisis referrals missing in over a quarter of the cases that warranted one, and rated the product an unacceptable risk, while OpenAI disputed the method and said account linking had not finished when the testing ran.
Micron's Taiwan Workers Voted To Strike – Of 2,258 eligible union members at the Taoyuan plant, 1,994 backed strike action, rejecting a one time NT$1.7 million offer and demanding a permanent 15% share of operating profit, at a memory maker whose product is the tightest input in the AI supply chain.
America Sentenced Its First AI Streaming Fraudster – Michael Smith got 18 months and forfeited $8,091,843 for pointing up to 10,000 bot accounts at hundreds of thousands of AI generated songs, which in April 2023 alone drew 80.9 million fraudulent plays on YouTube Music against 9.3 million legitimate ones for Taylor Swift's entire catalogue.
Three stories about work that got finished before the sign-off arrived. Google took down 300 hectares and is filing the assessment afterwards. OpenAI published 722 manuscripts and left most of the formal checking for later. Mistral trained a trillion parameters in Europe and is handing out the weights without waiting to see what anyone does with them. In all three cases the thing is already built. The review is the part still in progress.
Find out what shipped before the sign-off.
Interested in featuring your services with us? Email us at [email protected]
