Title is incomplete. It should be "No real advances in Open Source LLMs since 2019's T5". I think that is fairly close to being true. The real advances have come from OpenAI (and to some extent Anthropic and Cohere). I've played around with some of the open source models and as much as I want them to be "almost as good as GPT-3.5", they're simply not. On the flipside, you can do A LOT with T5 and Flan-T5.
Google 'secret' 'leaked' memo: 'We have no moat, OSS will rule all!'
Google actual thinking: 'meh, we had the some performance as the OSS models in 2019.'
The leak had the double purpose of shielding Google Research internally and disrupting investment in competitors, it never made any sense if you thought a bit about the market.
It's always the same if there is a technological breakthrough someone that worked in the field pops up and says "I could have done that years earlier." and the only thing you get as information out of it is that he didn't do it.
