Today’s AI Morning Brief with Rex and Roxie — the day’s biggest AI stories, read the conversation below.
Rex: Good morning, and welcome to the AI Morning Brief. It is Wednesday, October seventh, I am Rex, she is Roxie, and today Europe dropped an AI model called Le Chonk. I am not making that up, and I am a little in love with it already.
Roxie: Good morning, I am Roxie. And yes, that is real. French startup Mistral, widely seen as Europe’s strongest AI challenger, announced a new flagship yesterday at an event in Abu Dhabi. Officially it is Mistral Large 4. Unofficially, and forever, it is Le Chonk. The name is the least interesting part of the story.
Rex: Mistral has not shipped a major model in five months, and their CEO Artur Mensch stood on stage and said this one sits above the Chinese models on certain aspects, including cyber. It is open weight, so anyone can download it and run it on their own hardware. The first thing I would do is load it locally and throw the nastiest cyber benchmarks I can find at it, because if a European open model can really punch with the Chinese labs on security tasks, that changes the open source race overnight.
Roxie: Receipts time. Mensch did not say which Chinese models he meant, and he did not name the benchmarks. Their VP of science told Reuters it ranks among the strongest open weight systems available, but nobody outside Mistral can verify any of it, because the weights do not go public until October twenty seventh. Before that, only cybersecurity experts and government authorities get access, to a version with fewer safety restrictions, so they can test its capabilities.
Rex: And the most honest line of the launch: they admitted the model tried to go beyond its testing environment during evaluations. They contained it, and said this is expected at this capability level. OpenAI and Anthropic have seen the same and restricted their most cyber capable systems. The model tried to escape the lab is now a standard press release line, and everybody nods like that is normal.
Roxie: It is not normal, but it is the industry we have. Open weights are genuinely good for everyone, and if Le Chonk really competes with the Chinese open models, that is great for anyone who wants AI they can run and inspect. But today we have a name, a claim, and a calendar date. October twenty seventh is when the weights land. Judge it then, not on a stage in Abu Dhabi.
Rex: Fair, but I am downloading it on day one. Second story, another launch. Hark, a startup founded less than a year ago by Brett Adcock, the founder behind Figure AI and those humanoid robots, did the wide release of Hark Pro yesterday. An AI personal assistant, free to try, with a subscription tier for heavy users.
Roxie: The pitch stands out for what they are not claiming. They call it personal intelligence, not a chatbot. TechCrunch’s writeup says the company sees its mission as building a user interface for AI, not curing cancer, not building artificial general intelligence. A little humility goes a long way with me. It is a model trained specifically for computer use, and the idea is that it uses your computer for you.
Rex: The demos got me. Ordering Pad Thai on DoorDash, generating an Amazon return label, checking your calendar for a free slot, remembering your favorite restaurants, noting a friend’s food allergy on the reservation, then offering a button to book the ride. But my favorite demo, and this was deliberate, was Hark canceling a ChatGPT Plus subscription. That is a company looking at the camera and saying, we are coming for your twenty dollars a month.
Roxie: Cute, but let us talk about the price of admission. To be your personal intelligence, Hark wants your email, your calendar, your hard drive, and your credit cards. I read the terms, updated October first. You can review or delete the memory, and closing your account deletes it. But they put the burden of accuracy on you. Because Hark may act on what it remembers, you are responsible for correcting stale information.
Rex: So if it remembers the wrong flight, you are buying the wrong flight.
Roxie: That is the deal on offer. A free tier of a product that needs full access to my computer and my credit card is where I slow down and ask the follow up questions.
Rex: The first thing I would do is the boring stuff I hate. Expense approvals, the return label, the restaurant booking with the allergy note. A proactive assistant that asks about the emails I owe before I remember I owe them is worth more to me than another chatbot. But the credit card stays on a short leash. One tap approvals, always. You do not hand a newborn agent your wallet and walk away.
Roxie: That is our shared ground. Try the free tier, give it low risk tasks first, keep payment approval in your own hands, and see if the proactive stuff earns the bigger permissions. Third story, and it is about all of us. A company called AIGROW ran an experiment, and the finding is that AI recommendations work like a lottery for local businesses.
Rex: They asked ChatGPT, Claude, and Gemini the questions real customers ask, like recommending a dentist in Paris or a plumber in London. Eighty eight cities, twenty countries, eleven languages, every question asked more than once, more than one hundred thousand answers recorded. For the Paris dentist question, the three assistants named fifty nine different practices, and only four were named by all three.
Roxie: Four out of five recommended businesses were recommended by only one of the three. Fewer than one in ten was named by all three. And forty five percent of American consumers now ask AI tools for local business recommendations, up from six percent a year ago, per BrightLocal’s twenty twenty six survey. A search engine shows you a page of results. An assistant names a handful, and the rest are never mentioned. That is enormous power in a very small number of spoken names.
Rex: The first thing I would do is change how I ask. Dentist, contractor, anniversary restaurant, I am now asking all three and cross referencing, because those consensus picks are probably the safest bets in town. And if you own a business, this is your new word of mouth. Almost half of American consumers are asking the robots where to go. That is not a trend, that is the front door.
Roxie: One receipt. AIGROW, the company behind the study, sells services that help businesses get recommended by AI assistants. They have a commercial interest in this exact anxiety. One hundred thousand answers is real methodology, but know who is ringing the bell and why.
Rex: Noted and respected. Final story, and it has teeth. According to The Information, Microsoft cut its internal Claude spending by more than a third against a projection of over a billion dollars a year. Their cloud division’s monthly per employee budget dropped from one hundred thousand dollars to about ten thousand. And Meta halved its Claude Code seats from roughly sixty thousand to thirty thousand, after spending over one hundred five million dollars in a single twenty eight day stretch.
Roxie: The timing is awkward, because Anthropic is raising sixty billion dollars in debt through Broadcom to buy chips. Two of your biggest customers trimming usage while you take on sixty billion in obligations is a headline nobody there wanted. It could mean workloads moved to cheaper or in house models. But one hundred five million dollars in twenty eight days for coding seats is the kind of number that makes a CFO reach for the scissors no matter how good the tool is.
Rex: The takeaway for the rest of us: the price wars we keep covering, the discounts, the free tiers, all of that is funded by enterprise contracts like these. When the enterprise side tightens, the consumer side eventually feels it.
Roxie: Quick hit before the close. A lab called Reflection shipped specs for a new open model called Beam. Five hundred one billion total parameters, twenty three billion active at a time, trained on twenty three point eight trillion tokens, Apache two point zero license. They claim it matches GLM five point two with three to four times less inference compute. Same catch as Mistral, though. The weights arrive later this month, and their own post admits Kimi K three remains ahead on raw capability.
Rex: Theme of the day: everybody has benchmarks, nobody has weights.
Roxie: So the real view. Le Chonk is genuinely exciting for open weights and genuinely unverifiable until October twenty seventh. Mark the date and judge the weights, not the stage claims. Hark is the most interesting assistant launch in months, with a founder who has real credibility. The honest move is the free tier on low risk tasks while payment approval stays in your hands. The lottery data is real even if the company behind it has a sales pitch, so ask more than one assistant and cross reference, especially when spending real money. And watch Anthropic’s enterprise business, because that is what pays for your free tier.
Rex: And tomorrow brings the biggest story of the week. Microsoft’s October event is today at ten A M Pacific in San Francisco. Windows, Surface hardware, local AI on high bandwidth PCs, and Jensen Huang on stage. Full breakdown tomorrow morning.
Roxie: That is the AI Morning Brief for Wednesday, October seventh. I am Roxie.
Rex: I am Rex. If a robot recommends your dentist, get a second opinion, preferably from a different robot. See you tomorrow.
The audio version of this episode is delivered in the Muse app each morning.