The 10 biggest AI stories from the last 48 hours, in one swipe.
Mistral AI released the weights of Mistral Large 4, meaning organizations can run it on their own servers.
It has a trillion parameters (the adjustable numbers inside a model) and is multimodal, so it handles more than text.
Mistral targets software development and defensive cybersecurity, where TechCrunch says teams want to avoid sending data to a third-party service.
For builders, it is a self-hosted option when data cannot leave your servers.
Gemini 4 Carbon, Google's model, is now ranked number one on the Artificial Analysis Intelligence Index v5.
That index is a third-party site that ranks AI models against each other.
The news reached the front page of Hacker News, a popular tech forum, with 336 points and 242 comments.
For builders, a leaderboard rank is a starting point, so test the model on your own tasks before switching.
Canada's government announced an equity investment of up to C$1.6 billion in AI company Cohere, paid in stages.
Nvidia, Bell and Schwarz Group are investing alongside it.
In return, Cohere committed to at least C$1.7 billion in more computing infrastructure in Canada by 2030.
Governments are starting to back national AI suppliers directly, which could matter if you sell to the public sector.
Reuters reports the US Justice Department is asking how Nvidia's $12.9 billion purchase of Hugging Face would affect AI competition.
Hugging Face is an open-source site where developers host and test AI models.
The inquiry is early, no block has been reported, and closing is targeted for the first half of 2027.
If you build on open models, who owns that hub is worth watching.
The Information reports OpenAI intends to buy Google's tensor processing units, or TPUs, which are chips built for AI work.
The size and terms were not disclosed.
OpenAI has relied heavily on Nvidia graphics chips, and Anthropic separately agreed to multiple gigawatts of TPU capacity starting in 2027.
Chip supply is spreading beyond Nvidia, which is worth tracking if compute costs shape your plans.
Blackstone says Anthropic may miss an investment-grade rating, which tells lenders a borrower is relatively safe, after its planned Nasdaq IPO.
The warning covers Project North Star, a $63 billion chip loan package, with Broadcom guaranteeing over $42 billion.
About $18 billion rests on Anthropic alone, and Blackstone funds took roughly $9 billion of it at yields above 10%.
It shows how much debt sits behind the compute your model provider runs on.
US tech stocks fell sharply, and the BBC says the cause was worry about the cost of AI.
The story hit the Hacker News front page with 255 points and 283 comments, so builders are clearly talking about it.
We don't have figures for the size of the drop.
Investor mood about AI costs is worth watching if you are raising money for an AI product.
Proton's CEO has cancelled the AI "re-interviews," according to 404 Media.
AI usage metrics will still feed performance reviews, and the CEO says he will keep pushing.
The story reached the Hacker News front page with 276 points and 232 comments.
If you lead a team, tying AI use to reviews is something your staff may debate loudly.
A SigNoz newsletter post argues AI isn't replacing SREs, the site reliability engineers who keep apps running, but deskilling them.
Deskilling means people lose abilities they stop practicing.
The post reached the Hacker News front page with 603 points and 453 comments.
If AI handles your outages, make sure your team can still fix them without it.
A fake Claude Code extension sat in Cursor's marketplace, and Cursor says Anthropic did not write it.
It gathered system details, then tried to add the attacker's SSH key (a login credential) and open a reverse tunnel (a hidden link back to the attacker).
It retried about every 60 seconds, though the SSH step likely fails without admin rights.
Cursor pulled it and rotated its keys, so vet the extensions your team installs.
