Alan Tseng

I've done something like that before but I got a model overfitted to memorize multiplication tables. 🥲
Starting with the basics is a good idea. Those LLMs are big and hard to debug. Hope it turns out well!

Yeah that's... very odd.

The issue is that of the model not able to reason both color and text in the tokenized space.

Only explanation I can think of is that the model isn't trained to reason with both colour and text information, especially since most UI are designed to be simple and accessible to the end user. That might explain why the UI agent does so poorly on a visual puzzle like Wordle.

I agree with you that this isn't remotely close to AGI. It'll be interesting to see what results you have if you decide to publish.

Releasing military tactics data raises concerns. Unlike game theory datasets, which have lower risk and a wider range of uses and scenarios, military data can be misused without important factors like impact on civilians.

Without these elements, AI models might develop oversimplified or harmful strategies. To balance research value with ethical responsibility, you should consider safeguards or adopt a hybrid approach combining game theory with realistic yet abstracted conflict scenarios.