5 Thoughts on LLM Capabilities and Limitations

There are different views on the fundamental limitations of Large Language Models. On one hand you can hear that Transformers trained on next-token prediction (NTP) are not the right paradigm for building intelligent machines, that language by itself is too limiting, as a reductive projection of a much richer space (human experience of the real-world), that these machines do not truly understand (whatever understand means), or even that they are « stochastic parrot