2 ms·
I find it fascinating that how useful generative AI and LLMs are despite these basic logical limitations. It really puts into perspective the gap between our cu
by kubrickslair 3y ago
I find it fascinating that how useful generative AI and LLMs are despite these basic logical limitations. It really puts into perspective the gap between our current state of the art and more robust models we will see in the future.
Look forward to the day when such flaws are largely eliminated!
- nologic01 3y agoThis marriage of logic with statistical, ML based, LLM is highly non-trivial (if at all possible) and if it happens it will most likely be called another major "AI revolution". In fact the current AI mainstream considers this as a research direction to avoid at all costs (they term it "the bitter lesson"). The strategy and rallying cry is roughly translated: you can achieve gee-wheeze results here and now by ignoring these deep problems that stymied generations of AI researchers.
- kubrickslair 3y agoAbsolutely, the integration of logic into LLMs is far from a straightforward task. Having been involved in both traditional machine learning and common sense AI in my grad school years, I've seen first-hand the limitations of a purely statistical approach. (Some of my past data augmentation work is being used to benchmark LLM reasoning.) While most folks are too fixated by the 'quick wins' achieved by LLMs the trade-off is often a lack of non shallow reasoning. And I worry that many active researchers are glossing over these deeply rooted issues.