Mistral's got a shiny new flagship. The open models beating it? Not European. This is AI Daily Briefing. I'm Bill, with Cassidy. Also today: coding agents writing more code but not more software. I may have opinions. But first, Bitdeer's new Malaysia campus. Big build, big timeline. Who's it actually for? Here's Aakruti Thakkar at COMPUTE FORECAST:
Bitdeer AI, the AI cloud business of Bitdeer Technologies Group, has signed a 10-year data center services agreement for a 67-megawatt (MW) AI cloud campus in Malaysia. Announced Oct. 8, 2026, the agreement covers a facility known as A901 and adds another major site to the company’s expanding AI infrastructure network. The campus will support liquid-cooled, rack-scale NVIDIA GB300 NVL72 systems designed for enterprise AI workloads.
Sixty-seven megawatts in Malaysia, GB300 NVL72 racks, first capacity Q1 2027. And Compute Forecast says no signed offtake on A901. I've read this one before. Wednesday. It was called AM Intelligence. Different logo, same shape. Racks picked, megawatts counted, tenant line blank. Except Bitdeer signed a ten-year services agreement, so the builder side's committed for a decade either way. Ten years of liquid cooling for racks that don't land till 2027. The pitch says 'enterprise AI workloads.' Great. Which enterprise? They haven't even said exactly where the site is or who runs it. And Blockspace has this pushing them to 273.5 megawatts secured, about ninety-five percent of a 350 target. So they're building toward a capacity number. Two of these in one week? I'll call that a pattern. Doesn't mean it breaks. It just means the next Bitdeer disclosure I care about is a customer, not a rack count. Capital & Compute writes:
The verdict: Mistral says it beats any open-weight model built in the US or Europe, but it is not the best open-weight model. The independent Artificial Analysis Intelligence Index scores it 38, behind GLM-5.3 (45), Kimi K3 (44), GLM-5.3-Flash (42) and DeepSeek V4.1 Flash (39). All four score higher, and the two Flash models also cost about a quarter as much per task.
Mistral Large 4. Launch sale is sixty-eight cents in, two-oh-nine out per million tokens, and Mistral hasn't said when the sale ends. At list it doubles, which Artificial Analysis works out to a buck thirteen per index task. For a 38. Sixty-fourth out of two hundred twenty-five. GLM-5.3, Kimi K3, GLM Flash, DeepSeek V4.1 Flash, all ahead of it, and the two Flash models do it for about a quarter of the cost per task. So 'best open-weight model from the US or Europe' is a very carefully drawn map. Yeah, no. If I'm pricing an enterprise deal, that one's over before the second meeting. Unless the buyer needs European data residency, or cyber work the closed models refuse. Then the map is the whole pitch, and honestly? Some customers will pay for it. Sure, once you can download it. Weights are due 'by the end of the month,' and there's still no licence named. So today, 'open' means an API preview with a coupon attached. Ask me again when the licence has a name and the sale has an end date. This one's from Ars Technica:
Overall, the average “review process” time between a pull request getting submitted and it being merged into the codebase balloons 49 percent on average after AI agents are introduced. That effect can be seen in more granular data, too, with “the share of pull requests with changes requested nearly doubl, and the number of comments per pull request increas by 35%” following the AI agent shift, the researchers write.
So. Time from pull request to merge, up forty-nine percent once the agents show up. Share of PRs getting changes requested, nearly doubled. I'm allowing myself exactly one 'told you so,' and that was it. Spend it, you earned it. But the data stops in March, so this is research on last spring's agents. Every vendor's going to say their tooling's moved on since. Okay, but eighty percent of those firms already had AI code review running. And the bots still wrote under a quarter of the review comments. Humans are eating that work. And the researchers can't pin any employment change on it. What they do see is fourteen percent more people doing reviews. So the work just moved into the review queue, with extra staff to man it. From Hacker News:
This lines up with basic manufacturing principals, run the machines faster so as to give away quality leads to costs spiral in rework and inspection efforts. That or real end user demand eventually drops when quality issues are eventually experienced by customers leading to them to react, complain publically and seek alternatives.
Yeah, that's the factory-floor version. Run the line faster and the inspection station becomes your bottleneck. In software, inspection is your most expensive engineers. It's the second half of that comment I'd chase, whether the rework ever leaks out to customers. That's the next number I want. From Christopher Wieduwilt at AI Musicpreneur:
One senator was enough. On September 30, 2026, Sen. Marsha Blackburn (R-TN) asked the Senate to pass the NO FAKES Act by unanimous consent, and Sen. Ted Cruz (R-TX) objected. The bill, S. 4591, has 15 cosponsors from both parties and cleared the Senate Judiciary Committee on June 18. It would give every person in the US a federal right over AI clones of their voice and face.
Fifteen cosponsors, both parties, cleared Judiciary back in June. Then September 30th, Blackburn asks for unanimous consent, Cruz objects, and that's the whole afternoon. And listen to how Glazier talks about it. 'A real shot' in the lame duck after the November 3rd midterms, and then, same interview, he says a 2027 reintroduction is the more likely outcome. That's the RIAA's CEO setting the over-under on his own bill. The number I'd actually circle, if I ran anything that hosts user audio: up to seven hundred fifty thousand dollars per work for platforms that ignore the takedown rules. Per work. That's a takedown queue with its own headcount. And keep the two fights apart. This bill's about clones of your voice and face. Payment for AI training, which Glazier calls one of the RIAA's biggest issues, he's fighting in court. Same industry, two separate fights, and only one of them is stuck on the Senate floor. If you're finding the AI Daily Briefing useful, please subscribe or leave a review wherever you're listening. Reviews help other people find the show, and your support keeps these briefings going.
Looking ahead, Mistral's promised downloadable weights and a license for Large 4 by the end of October, and the RIAA is pushing for another shot at the NO FAKES Act in the lame-duck session after the November 3 midterms. Links to every story are in the show notes, so take a look at whichever ones caught your attention. That's AI Daily Briefing for today. This is a Lantern Podcast.