> It seems that at the moment LLMs are not able to grok Oberon source code (although they are able to generate it). This seems unlikely. The "knowledge" LLMs

> It seems that at the moment LLMs are not able to grok Oberon source code (although they are able to generate it). This seems unlikely. The "knowledge" LLMs have is inherently bidirectional, so I suspect it's likely that one of the following is true: - any model than can generate it can also grok it (though you may need to know exactly how to ask) - that what it's generating isn't actually valid Oberon - that it can only generate it for a very narrow subset of things (eg specific examples it's been trained on).