Three things crossed my desk in the last day and every one of them is worth a minute of your time. A newspaper chain cut a data deal that made its own reporters furious, two more open models dropped with numbers that made me raise an eyebrow, and Google put out a tool that lets AI models work on data they can't even see. Let's get into it.
The Palantir and USA Today Data Deal
USA Today Co., the biggest newspaper chain in the country, just partnered with Palantir to monetize its audience data. Problem is, over 800 unionized journalists found out about it the same way the rest of us did, after the fact. That is not how you build trust with the people writing your stories.
Why it matters: newsrooms have been sitting on mountains of reader data for years and nobody quite knew what to do with it. Now Palantir does, and that data is worth real money once you run it through an AI pipeline. Every media company watching this deal is going to want in.
Robert's take: I don't blame USA Today for wanting the cash, newspapers are hurting bad. But cutting a deal like this behind your own newsroom's back is a good way to torch whatever trust you had left with readers and staff alike. If you're going to sell the data your reporters helped generate, at least have the decency to tell them first.
Qwen and GLM Push Open Models Forward
Alibaba's Qwen team dropped Qwen 3.8 27B under Apache 2.0, a 27 billion parameter model with vision built in and a native context window of 262K that stretches to a million tokens. Meanwhile Z.ai put out GLM-5.3, and the benchmark jump is not subtle. Their Terminal-Bench score went from 4.6 to 28.3 in one release.
Why it matters: this is the open source side of AI actually closing the gap fast. A permissively licensed 27B model with a million token context is the kind of thing a small shop can run without begging a cloud provider for GPU quota. And a six-fold jump on a coding agent benchmark in a single release tells you these teams are not coasting.
Robert's take: I like seeing this fight stay competitive. When Alibaba and Z.ai keep shipping real gains under open licenses, it keeps the big US labs honest on price and keeps builders like us from getting locked into one vendor's terms. Go download the weights and kick the tires yourself before you take my word for it.
Google's HEIR Wants Your AI to Work Blind
Google put out HEIR, short for Homomorphic Encryption Intermediate Representation. It's an open source compiler toolchain that takes a pretrained model and lets it run inference on encrypted data, meaning the model never actually sees your raw information.
Why it matters: this is the kind of plumbing that makes AI usable in places it currently isn't, healthcare records, financial data, anything where sending raw data to a third party model is a nonstarter. If encrypted inference gets fast enough to be practical, a whole category of use cases opens up.
Robert's take: encrypted inference has been a research curiosity for a long time because it's slow as molasses. Google putting real engineering behind an open compiler for it is a good sign they think the performance problem is solvable. Worth watching, not worth betting on yet. Ask me again in a year.