The number of permutations you can try to decode a book cipher is near infinite. Just "a book cipher" does not tell you how the numbers map to words or characters in each book.

You just try creating random book ciphers and see whether an LLM would solve them without hints.

> The question for me is where it leaves innovation. Have we really peaked as a species that we so easily want to give up on the young and hand them over a tool that would substitute reasoning?

This is a crucial question but I'm afraid the ship has sailed. We should focus on how to prevent atrophy and cultivate our skills in spite of LLMs. For me and many others this means doing thins the old way with occasional assistance from LLMs that actually increase our understanding. Occassional because if you do it often, you get intellectual atrophy because you stop even trying.

If we think that machines that recombine and interpolate our prior out are intelligent then yes we have probably peaked as a species. Anyway evolution is not telological. It may surprise HN posters but people were as intelligent if not more so as present thousands of years ago. When a problem is solved is more a question of attention, reward, available knowledge base, etc. intelligence is an adaptation to solve the problems of biological life. It would be hard to argue that solving this cypher would have increased anyone's reproductive fitness that much hence it was unsolved. Unsolved problems are often so less because of intrinsic difficulty but lack of conditions that impell their resolution. It is not at all surprising that llms with the full human knowledge base at their disposal, ample computational resources, and the programmed reward of finding the best completion are solving unsolved problems nobody needed to solve anyway. This however is not intelligence in the human sense. Humans solve problems for biological advantage in a dynamic landscape. Llms solve problems to optimize a completion function in a fixed pre trained one.

They cannot yet outmatch humans in critical thinking. Heck, they cannot outmatch a mediocre person like me yet. How many times have we awen "You are right...." Even on sota models?

They are great tools for research and tasks as of now.

There seems to be a fundamental misunderstanding here that many LLM advocates fall prey to. No one claims (anymore) that LLMs are completely useless. We agree that these tools can boost productivity. But there is a fundamental difference between human creativity and the statistical processing of large amounts of data in an artificial neural network. The latter can yield new insights for the person operating the LLM. But if the insight was more or less contained 1-to-1 in the training data, this confirms the observation that the original contribution of LLMs consists mainly of unearthing relevant facts, but is otherwise very limited when it comes to synthetic judgments (in the Kantian sense).

If the claims made by LLM manufacturers regarding these systems' own capabilities repeatedly turn out to be little more than a hoax, critics of this practice shouldn't simply be dismissed as Luddites. Instead, we should acknowledge that, despite their remarkable capabilities, these systems ultimately deliver far, far, far less than what the snake-oil salesmen—cruising through the valley in their Koenigseggs—repeatedly promise to the public and investors. Anyone who still hasn't grasped this, even though it's so obvious, should urgently focus more on what is actually the case and less on what is being sold to us as the future. Anyone who thinks that technological leaps can simply be extrapolated linearly like this is just naive!