I always been mesmerized by Cyc (and EURISKO). I've born in late USSR, and my father (then "junior staff scientist" in natural language processing) has several translated books (borrowed from Institute's lab, as I understand mow) which were (again, as I understand now, many years alter) abbreviated translations of proceedings of some "West" conferences about AI. As I read everything about computers, I read these books too, and was deeply impressed by description of EURISKO (it worth noting, that fleet-simulation game was not mentioned, but MOS structure was). I was... 10? 12? years old.
I still think, that this approach much better than modern "multiple huge matrix by huge vector. I understand, that DNNs and LLMs won practically. But I still think that these blackbox pure-feedback based approaches are not so elegant as knowledge engineering... Practical. But in some sense "dirty". Like "dirty hacks" in code.
BTW, part about ABBYY is not exactly correct. I know this because my father became one of the founders and life-long CTO of PROMT, company which is (was? It is not clear after beginning of war, unfortunately) author of successful pre-LLM grammar-based translation systems (Stylus, later PROMT) and partner of ABBYY. ABBYY translation project stated much later than 1990s, because in 1990s ABBYY distributed PROMT technology, and ABBYY itself was (is) experts in OCR and dictionaries (dictionaries is not translation system, of course). Translation project in ABBYY started much later, when their ways with PROMT parted due to several business-related reasons.
I always been mesmerized by Cyc (and EURISKO). I've born in late USSR, and my father (then "junior staff scientist" in natural language processing) has several translated books (borrowed from Institute's lab, as I understand mow) which were (again, as I understand now, many years alter) abbreviated translations of proceedings of some "West" conferences about AI. As I read everything about computers, I read these books too, and was deeply impressed by description of EURISKO (it worth noting, that fleet-simulation game was not mentioned, but MOS structure was). I was... 10? 12? years old.
I still think, that this approach much better than modern "multiple huge matrix by huge vector. I understand, that DNNs and LLMs won practically. But I still think that these blackbox pure-feedback based approaches are not so elegant as knowledge engineering... Practical. But in some sense "dirty". Like "dirty hacks" in code.
BTW, part about ABBYY is not exactly correct. I know this because my father became one of the founders and life-long CTO of PROMT, company which is (was? It is not clear after beginning of war, unfortunately) author of successful pre-LLM grammar-based translation systems (Stylus, later PROMT) and partner of ABBYY. ABBYY translation project stated much later than 1990s, because in 1990s ABBYY distributed PROMT technology, and ABBYY itself was (is) experts in OCR and dictionaries (dictionaries is not translation system, of course). Translation project in ABBYY started much later, when their ways with PROMT parted due to several business-related reasons.