Thomson Reuters spent $40M and 20 months to train Qwen3.6 which is 130…

Thomson Reuters spent $40M and 20 months to train Qwen3.6 which is 130 days old and 2 models behind Qwen3.8: is it rotten?

Steve Hasker, to Richard Tromans on Artificial Lawyer video 👉🏽 https://lnkd.in/dtEgYeAF

They joined the training treadmill with Harvey 👉🏽 https://lnkd.in/dt5BBaGD

Training a model on a larger legal corpus will extend its shelf life more than training it on a few expert lawyer rubrics which is just benchmaxing

Corpus acts like a preservative - Salted Onion™ - stolen from Ross Brodskiy

But using corpus to keep LLM's fresh has been tried before

In 2023 Microsoft put GPT-4 on Bing and went from 3% of search to 4.47%.

Then in August 2025 they retired the entire Bing Search API into Azure.

Perplexity ran the same play, and is now partnering with NVIDIA in a obvious pivot to subscriptions while traffic slides. Btw, they also tried legal and...zip

NVidia motives are to keep demand for chips growing since it may not be growing naturally anymore...

Fair pushback: open web is not the same as KeyCite. Exclusivity still matters.

Sidenote: Google walked in 👉 https://lnkd.in/dQ4vfCP6

Like they did with Document AI for Legal in 2019 and star-studded list of legal tech vendors which resulted in...nothing 👉🏽 https://lnkd.in/d3JJX9fq

Tokenomics still Tricky so please...Enjoy The Weeknd 🎧