ROSS

ROSS Intelligence's reply brief is filed. Oral argument is next. 20 highlights: 1/20: “West's answering brief is a hallucination.”

"It seems like we're seeing this rent-seeking process where the world's largest content companies are just adamant that we should not have this AI progress without us acting sort of like a pole through which people have to pass." - FTC Chairman Andrew Ferguson @AFergusonFTC

Large content companies' attempts to claim a slice of the pie whenever other firms train large language models on their data “feels awfully like rent seeking to me.” - Federal Trade Commission Chairman Andrew Ferguson @AFergusonFTC

Chair @AFergusonFTC has copyright monopolists' number: hiding behind individual creators to justify wringing rents from facts used for gen AI training defeats ©'s purpose of promoting "progress" and thwarts innovation. He's not buying it:

.@AFergusonFTC put it plainly: ""It seems like we're seeing this rent-seeking process where the world's largest content companies are just adamant that we should not have this AI progress without us acting sort of like a pole through which people have to pass."

Ferguson directly called out news outlets' overreaching demands that facts be licensed: "[T]he main is, 'no training without paying because necessarily, our content is going to inform things that the LLM says.' And it just feels awfully like rent seeking to me." Well said!

@andrewarruda 👀

👀

Thomson Reuters is building this own AI model, “Thomson” but here's the catch: All AI models are trained on vast troves of copyrighted data. For “Thomson” to compete, fair use would have to be apply to training AI models, exactly what’s at stake in Thomson Reuters v. ROSS Intelligence where Thomson Reuters is arguing fair use does *NOT* apply to training! What TR is implicitly admitting is that TR’s legal data (which may be found to not even be copyrightable in TR v. ROSS) is only really useful on top of the abilities of an advanced base model. Abilities that come from exactly the kind of training TR is arguing against in court in TR v. ROSS. So, TR is free-riding on the AI investments and training of other AI labs, which are only made possible by the same fair use assumptions TR is challenging in TR v. ROSS. TR’s legal position in TR v. ROSS, if successful, would saw off the branch TR is sitting on and needs to compete broadly (see $TRI 's stock performance). Unless TR builds, from scratch, a competitive base AI model, using only their own content, and only outside licensed data, with ZERO unlicensed data, TR’s business lines will slowly dry up and die due to competition, and all this would stem from the precedent TR itself is arguing in TR v. ROSS. Competitive base models, without licensing, cost ~10B+ right now, with licensing, who knows how much that would cost. TR’s $650M Casetext acquisition bought TR a company with relaviely low revenue and a team with experience building with base models (specifically OpenAI’s), but Casetext was a team with limited if any experience building AI base models from scratch. ROSS, the team with experience in both (building models from scratch + building with other models), TR instead sued. ROSS was founded by students from the same University of Toronto labs where deep learning was pioneered, and was actually building foundational AI technology, training models from scratch. And this is all compounded by the US reliance on domestic commercial AI models by the Department of War. The ruling in TR v. ROSS will cut across all AI training: ROSS trained in the same foundational way ALL other models train, which means a bad ruling in TR v. ROSS is existential to US national security. Dale Cendali, TR's lead attorney from Kirkland & Ellis on Thomson Reuters v. ROSS said: “Although this case may have differences from generative AI cases, we think this decision [TR v. ROSS] provides a useful framework for how to think about whether a company’s training of its AI is permitted under the fair use doctrine.” [source: kirkland.com/-/media/news/p…] And I agree, which is why the ruling in TR v. ROSS is so critical. 🔗artificiallawyer.com/2026/03/24/tho…

“Although this case may have differences from generative AI cases, we think this decision [TR v. ROSS] provides a useful framework for how to think about whether a company’s training of its AI is permitted under the fair use doctrine.” - Dale Cendali, @Kirkland_Ellis

“We think this decision provides a useful framework for how to think about whether a company’s training of its AI is permitted under the fair use doctrine.” - Dale Cendali, Kirkland & Ellis 🔗kirkland.com/-/media/news/p…

Brutal

@Techmeme @zijing_wu

OpenRouter data: lower-cost Chinese AI models by groups such as DeepSeek and MiniMax have surpassed US rivals in token consumption since last month (@zijing_wu / Financial Times) ft.com/content/256787… #a260325p46" target="_blank" rel="nofollow noopener">techmeme.com/260325/p46#a26… 📥 Send tips! techmeme.com/contact

The founders of startup Manus have been barred from leaving China and are facing a criminal investigation, simply because they sold their Singapore-based company to Meta. You are no free as long as in China. Your knowledge, work, even your life belong to the Communist Party.

@andrewarruda 👀

@ROSSIntel 👀

Thomson Reuters is building "Thomson" on Meta open models despite Kadrey v. Meta's revelations on how Meta's models are trained. TR is now commercially benefiting from weights trained on pirated materials. Is this a tacit agreement that fair use applies to training AI models?

ladies and gentlemen... China:

and how are china's open-source AI trained? here's what deepseek said, on the record: “we collected source code and compiled PDFs from 1.4 million arXiv articles ... we cleaned 860K English and 180K Chinese e-books from anna’s archive (anna’s archive, 2024)”

you need non-generative AI for humanoid robots to get trained and function. you also need a ton of data to train them well, so, fair use

especially considering ... x.com/ChinaSelect/st…