Good day . Microsoft says it built a better base model than the one behind GPT-5.6 on a quarter of the compute, and Amazon wants you wearing a $75 listening wearable. Plus Meta hands out the weights to Muse Spark Open, a Harvard survey on who bluffs about AI, and a Linux developer who got Claude to write drivers for hardware nobody documented. Let’s get into it.
Microsoft Says It Beat OpenAI Cheaper Business Insider
What happened: At a Thursday town hall, Microsoft AI CEO Mustafa Suleyman told employees that MAI-Base-2, announced earlier that day, is “a stronger pre-trained base model than the model that currently powers GPT 5.6 Sol,” trained with “one-quarter of the compute that OpenAI used.” He said Microsoft tested it privately against OpenAI models, including one nicknamed “Spud.”
Why it matters: Pre-training is the first stage of building a model, when it learns patterns from huge piles of data, and it’s the most expensive. If Microsoft can do it for a quarter of the cost, it needs outside suppliers less. In August we covered Microsoft telling its engineers to stop burning tokens and moving its internal default to the cheaper GPT-5.6. Now it’s saying it can make its own.
What everyone’s saying: Business Insider found it striking that Suleyman would point at limits in OpenAI’s models, even privately, since Microsoft has put in billions and its stake is likely worth more than $100 billion. He also said the model used “12 times less computation than was used for DeepSeek,” which a public Microsoft blog post backs up for DeepSeek’s V4 Pro Base.
My read between the lines: The comparison that makes the headline isn’t published, and the one that is public only covers open-weight models. Business Insider says Suleyman suggested Microsoft has special access to information on how OpenAI’s models perform, so one partner is grading the other with the answer key. Microsoft declined to comment and OpenAI didn’t respond.
Microsoft says it did the job on a quarter of the compute. Your version of that win: the weekly report that stops eating your Friday. Viktor is an AI agent that lives in your team’s Slack, connects to 3,000+ tools, and does real work: reports, dashboards, code and campaigns. Not a chatbot. A coworker. New readers get $50 off their first month. Hire Viktor →
Amazon Wants a Recorder on Your Wrist TechCrunch
What happened: Amazon’s Bee launched two personal AI devices on Thursday: the $74.99 Bee Frontier, worn on the wrist or clipped to clothes, and the $99.99 Bee Eclipse, a credit-card-sized recorder that sticks magnetically to the back of a phone. Frontier is on sale now and Eclipse is “available soon.” Both turn conversations into summaries, to-dos and reminders.
Why it matters: Free devices keep 14 days of memory; the roughly $12-a-month Bee Pro adds unlimited summaries plus links to Amazon.com, Alexa+ and apps like Gmail. CNBC notes that when turned on, Bee devices are always recording, and Amazon says no audio is stored and transcripts are encrypted so “no one, including Amazon, can access” them. In July we covered AI notetakers joining one in three meetings, mostly without anyone asking. This is that, minus the meeting.
What everyone’s saying: TechCrunch points out Bee undercuts rivals: Plaud’s Note is $159 and Pocket is $129. CNBC puts it in a holiday wave of AI gadgets, alongside Meta’s Muse Charm pendant, Plaud’s card and Friend’s necklace.
My read between the lines: The privacy promise covers Amazon’s access to your data. It says nothing about the person across the table, whose entire disclosure is a small indicator light. And a $75 recorder is cheap for a reason: the $12-a-month tier is where what you said at dinner can reach Alexa+ and the Amazon store.
The Brief is free and always will be. A headline tells you what happened, though, not what to do about it. Members get the paywalled deep-dives behind the stories, plus the full archive. Become a member →
Meta Releases Muse Spark Open Weights Meta AI Research
What happened: Meta released the weights for Muse Spark Open, which it calls “the most capable US open-weight model to date”: 816 billion parameters (42 billion active per token), a one-million-token context window, and built for coding and agentic work. It’s available through providers like Together AI, Fireworks, Groq and OpenRouter, and as downloads on Hugging Face and Amazon SageMaker. Meta says it runs on two Apple M5 Ultra Mac Studios.
Why it matters: “Open weights” means anyone can download the trained model and run it themselves, which makes it your hardware and your data instead of someone else’s API. Earlier this week we covered Meta saying Muse Wonder puts it back at the frontier. This is the other half: a model it’s handing out instead of renting.
What everyone’s saying: Hacker News turned it into a 780-point argument. One commenter cited Artificial Analysis benchmarks showing it behind Qwen and GLM models that fit on a single B300 GPU. Another noticed that open-weight labs have shifted from claiming “the best” to claiming “the best in the US.”
My read between the lines: That qualifier does heavy lifting: “most capable US open-weight model” is a category with China carved out. The license also carries a “Model as a Service” clause for companies selling API access above a revenue line, and some commenters argue that means it isn’t really open. It’s open for you and your laptop, and not for anyone hoping to rent it out as a rival cloud.
Harvard: Men Bluff More About AI WIRED
What happened: In a Harvard Kennedy School Shorenstein Center survey of more than 5,400 Americans, 53% of men said they knew of Mira Murati, Susan Milner and Craig Dalton. Milner and Dalton do not exist. Murati is real, the former OpenAI CTO who now runs Thinking Machines Lab. Among women, 40% claimed to know the invented names and 38% said they knew Murati.
Why it matters: Men used AI more and answered more factual questions correctly, but men and women got questions wrong at about the same rate. The difference was “not sure”: women picked it on an average of 3.3 of 12 questions versus 2.3 for men. In July we covered a study finding AI advice made people more wrong and more sure of it. Confidence and accuracy keep failing to travel together.
What everyone’s saying: The report also looked at coverage: across a year of AI news, men were named in headlines more than 14 times as often as women, and three-quarters of the experts quoted were men, against an AI talent pool that’s roughly 29% women. Executive director Laura Manley says “we have to be able to say ‘not sure’ more often.”
My read between the lines: The invented names are the best part of the design, since you can’t be right about someone who doesn’t exist. Confident claims got surveyed, then quoted, then written up, and “not sure” got none of that airtime. If your AI news diet is made of confident voices, the survey is also describing the diet.
Claude Wrote Drivers for Mystery Hardware Christian F.K. Schaller
What happened: GNOME and Fedora developer Christian Schaller used Claude Code to write a Linux driver for a film scanner whose chipset had no Linux support, then to patch the kernel and libcamera for his ThinkPad’s MIPI camera. The trick was a test rig: the vendor’s Windows driver runs in a virtual machine, the USB traffic is recorded from the host, and the AI diffs that trace against the Linux driver’s. He released it as hw-harness.
Why it matters: Writing a driver for undocumented hardware has meant weeks of staring at hex dumps. His first working scan came within hours, and the camera fixes took a day: in a low-light test, frame brightness rose from 31.1 to 90.9. He notes the rig only observes what the hardware does, so every constant traces back to something recorded.
What everyone’s saying: Hacker News piled on with wish lists: an old Wacom tablet losing driver support, orphaned hardware generally, and someone asking for a go at smart TVs, to strip the ads and botnet-style proxies out of millions of devices.
My read between the lines: Schaller’s own footnote is the funniest part: after giving up and then succeeding, he found someone had already committed upstream scanner support. The AI shrank months into hours and the humans got there first anyway. What got cheap is driver archaeology, and what didn’t is the person holding the actual hardware to check it works, which he says in his own words.
📖 Further reading: REA: The Free Binary Ninja Sidekick Alternative That Still Needs a Decompiler — my teardown of REA, an MIT-licensed add-on that lets a coding agent take apart apps, binaries and websites. Same idea as Schaller’s harness, aimed at software instead of a scanner.
That’s your AI Brief for Friday.
—Nicholas from Artificially Intimidating
