Inspired by yesterday’s talk I took glm-5.3-flash for a test drive tonight.
This is offered for example by the GWDG in Göttingen
docs.hpc.gwdg.de/services/…
as well as commercial providers all over the place. I created an account at Cortecs which is an Austrian company offering only EU providers with zero-data retention promise:
cortecs.ai/detailedServerlessView/…
We have learned from openai that this promise is worthless in the US, but I do consider trusting EU providers.
Would you?
For openai to train on our chats, they just need to keep them and copy them inside the company because they are both a hosting provider and a lab that develops models.
If one of those cortecs providers wants to keep the data they do need to sell it (on the black market), because they don’t train models themselves. It’s feasible, but one more step.
Do I need to say that the performance of glm-5.3 is amazing? I started from an empty pi coding agent and wrote myself 3 extensions to interact with #DevonThink, #Thunderbird and a litte cost control extension. In Benchmarks this is approach Opus 4.8 level and yes, I remember that opus-4.8 was pretty amazing already.
As the first task I renamed all the 22 generic-name scanned documents in my DevonThink inbox to match their content. Worked flawlessly. One was a todo note I wrote to myself and the model noticed that and turned it a todoist task in the same go.
The GWDG had reliability issues, so I mostly used cortecs. I paid 22 Cents for my 2 hours of fun.
