Reading time: 6 minutes
🗞️ In this edition
The FTC Came For The Labs And Their Auditor
Google Shipped A Model Almost Nobody Can Run
Tencent Rented Its Way Around The Chip Ban
⚡ YOUR MOVE For people building the life they want, now.
In other AI news:
AI Borrowers Are Paying For The Doubt
Salesforce Bought The Customer Interview
An AI Beat Stratego Without The Compute
4 must-try AI tools
The FTC sent legal demands to OpenAI, Anthropic and the nonprofit that grades them. Google published a model that wins on thirteen benchmarks and handed it to a list. Tencent found the sentence the chip ban forgot to write.
Keep reading.
From Our Sponsor:
The phishing email that used to have typos now reads like it came from your CFO. The "urgent" call now sounds like your mom. Scammers are using the same tools you are.
Antiscam by Surfshark helps you spot it before you get burned.
Flags suspicious links, messages, and calls before you act on them
Plain language guidance on what a scam looks like right now, not five years ago
Works at home and at work, so one bad click doesn't cost you a client or a bank account
You can't out think every scam. You can make sure you're never the easiest target.
Source: ABC News
WHAT HAPPENED
The Federal Trade Commission opened an investigation into Anthropic, OpenAI and other AI companies, along with METR, the nonprofit that runs independent evaluations of frontier models. The agency is using civil investigative demands, which compel the companies to hand over information and compel executives to testify about their products, and it is working under the FTC Act, the statute that covers unfair and deceptive practices. The inquiry sits alongside the voluntary safety accord signed days earlier by Elon Musk, Mark Zuckerberg, Dario Amodei, Jensen Huang, Greg Brockman and Sundar Pichai. President Trump said of that accord that he was seeing tremendous self-policing, and that the companies understand they have to self-police. OpenAI and Anthropic did not immediately respond to requests for comment.
WHY IT MATTERS
Every safety claim a lab made this year was made voluntarily, in a blog post or a system card, with nobody on the other side of the table. A civil investigative demand changes the audience for those documents, because the FTC Act does not ask whether a model is safe, it asks whether what you said about it was true. That turns a year of published evaluations into a year of potential representations, and it puts METR in the strangest position of all, since the nonprofit paid to check other people's claims is now being asked to account for its own. The agency needs no new AI law to do any of this, which is the thing Washington has spent eighteen months arguing about. It only needs a lab to have oversold a result.
OUR TAKE
The obvious read is an agency contradicting a president who had just praised self-policing. We think the accord and the investigation are the same policy arriving in two envelopes, and the giveaway is METR. You do not pull in the auditor unless the plan is to make the voluntary system binding by holding everyone to what they already wrote down.
Watch what the FTC asks for, not what it says. If the first public outcome is an order requiring a lab to publish evaluation results on a fixed schedule, before the end of 2027, then the accord grew teeth and nobody had to pass a law. If it ends in a fine over marketing copy and no lab changes what it discloses, this was an ordinary consumer case that happened to land on AI companies.
Source: VentureBeat.com
The gaps are not small in places. Argon scores 19.6% on Harvey's legal agent benchmark against 5.4% for GPT-6 Astra and 3.8% for Claude Opus 5.5, and it takes AutomationBench at 51.3%, DeepSWE v1.1 at 77.9% and LVBench at 91.7%. Output runs to one million tokens, up from sixty four thousand. Introductory pricing is $2 per million input tokens and $10 per million output, doubling to $4 and $20 once the introduction ends, with cached input discounted 95%.
How did today's stories land?
Source: Forbes
The chips never enter China. They sit in Oracle data centers across Southeast Asia, and Tencent reaches them remotely to train models and build agent tools, with roughly 30% of the money paid upfront. The Financial Times reported the arrangement, which follows the shape of the Japanese data center Tencent used to reach Nvidia Blackwell parts in December 2025. Oracle shares rose 1.86% in premarket trading on the report. Neither company commented.
⚡️ YOUR MOVE
FOR PEOPLE BUILDING THE LIFE THEY WANT, NOW.
🩹 Google's Threat Intelligence Group found that bugs discovered by AI reach remote code execution 50% of the time, against 26% for everything else, and disclosures doubled from 5,045 in January to 10,477 in July. → Shorten your patch window
💾 Micron says memory and storage will be much tighter in 2027 and 2028 than they were this year, and that it has no line of sight to when supply and demand come back into balance. → Buy the RAM early
🔎 Cohere's Embed 5 Pro and Fast share one embedding space, so you index once with Pro and query with Fast at 2.4 times the throughput, from $0.08 per million tokens, at 128K context across 100 languages. → Re-index once
🎨 Ideogram 4.5 is built for editing the same image over and over without the artifacts, pixel shifts and colour drift that pile up in other models after a few passes. Open weights are promised. → Edit it twice
AI Borrowers Are Paying For The Doubt – Low rated firms have raised $88 billion of AI linked debt this year against $20 billion across the first eleven months of 2025, SoftBank paid 9.75% on its longest notes, and one $2.25 billion issue traded seven points below its price within weeks.
Salesforce Bought The Customer Interview – The roughly $2 billion deal brings in Listen Labs, a three year old platform that recruits from a network of more than 50 million people, interviews in over 120 languages and builds digital twins that predict what customers will do.
An AI Beat Stratego Without The Compute – Ataraxos went 15-1-4 against top ranked humans and 39-2 against world championship players on less than a hundredth of DeepNash's training examples, and the work is published in Nature.
Midjourney: Makes the image in your head before you can describe it to anyone.
Jasper AI: Writes the campaign in your brand voice instead of a stranger's.
MurfAI: Narrates your video without you hearing your own voice back.
Canva Image Generator: Fills the empty frame in the deck you present tomorrow.
Three stories about rules that only bind the people who agreed to them. The labs signed a safety accord in public and got civil investigative demands in private, because a promise is free right up until somebody writes it down. Google published a benchmark sweep and then handed the model to an approved list, so the scoreboard is public and the product is not. Tencent read an export ban that covers shipping chips and rented them by the hour instead. The rule is almost never the thing that stops anyone. The enforcement is.
Check who is actually bound.
Interested in featuring your services with us? Email us at [email protected]
