After a lot of hard work recently, I am pleased to finally give an update on our in-house AI efforts. While CarlosGPT was certainly impressive, it was the first iteration of our attempts at building an LLM for our needs, and that always showed. It was largely unpolished, there was lots of knowledge gaps, and it was extremely prone to hallucination. I have been working tirelessly to improve our capabilities in this area, and can now unveil our two newest models

• “Olympus” – 9 Billion parameters. This is our bleeding edge model, trained and fine tuned on both EVE Online data (our largest ever collection of chatlogs, forum scrapes for multiple major alliances, etc) and real world training data. It also integrates web search, and can analyze uploaded files. At the moment only PDF files are supported, although I am hoping to add support for other filetypes later. This works fine on my 128GB rig, and I think it should comfortably work on anything with 32GB but I haven’t tested that.

• “Remus” – 500 Million Parameters. This is our more lightweight model for every day use.

A lot of CarlosGPT’s training data was Jita chat logs, and this would often cause the AI to end up regurgitating Jita local type spam at the end of an answer it gave. I believe this fault has been largely mitigates in both Remus and Olympus. 

Test Out The New Models!

At the moment, I am restricting the use of the Olympus model for my use only. However I am looking for feedback on the Remus model which is now available on the TSOTH app. It is fairly lightweight and should work very well on even low spec hardware, I have tested it on an old 4GB iPad and it works fine.  You can even switch seamlessly between the legacy CarlosGPT model and Remus in-app! This will be kept for some time, and while I do plan to eventually phase out CarlosGPT entirely, this will not be for a long time and I will keep everybody updated about when the API will be deprecated.

Try out the Remus AI and let me know what you think!