Are There Any Full Turing-Class Entities in the House?
Reasoning about ancien régime French aristocrat family relationships, House-of-Orléans edition. No. ChatGPT4 is not even close to being a proper full Turing-class entity. But the fact that it is so effective at mimicking our use of language and persuading us that it is a Turing-class entity should greatly shake our belief in our own intelligence—that we ourselves are Turing-class entities, rather than simply jumped-up monkeys trying to fake it in the hope that we can then persuade enough things we interact with that we make it…
Things like this from the very sharp and highly energetic Timothy B. Lee make me want to bang my head against the wall:
Timothy B. Lee: Google Gemini and the future of large language models:
‘My guess—and at this point it’s only a guess—is that we’re starting to see diminishing returns to scaling up conventional LLMs. Further progress may require significant changes to enable them to better handle long, complex inputs and to reason more effectively about abstract concepts…
This last phrase is utter bilge.
There is no way that conventional LLMs can “reason more effectively about abstract concepts” because they do not and cannot reason about abstract concepts at all.
They make one-word-ahead predictions.
They make them very plausibly.
But they do not make with enough accuracy and sophistication to create some patterns of emergence or supervenience that would make the words “concepts” and “reasoning” plausible descriptors of what they are doing.
Now do not get me wrong. It would be wonderful and exciting if ChatGPT4 had, in the process of learning to do one-word-ahead predictions, found somehow that an efficient way of doing this was to construct internal representations of abstract concepts and carry them around, reasoning about them—that a swapping of nouns turns “uncle” into “nephew”, that applying “son” twice produces “grandson”, and so on.
But there is absolutely no evidence anywhere that they are doing that.
If “Louis”, “Philippe”, and “Orleans” are in the mix, then there is a good chance ChatGPT4 will take a passage ending in Louis Philippe and recursively predict that the next tokens are , duc d’Orléans. And there is a good chance it will take a passage ending in Philippe, duc d’Orléans and recursively predict that the next tokens are (1640-1701). Louis Philippe, duc d’Orléans and Philippe, duc d’Orléans (1640-1701) are both token-strings relatively common in the training dataset.
However, there never was any personage as Louis Philippe, duc d’Orléans (1640-1701).
But ChatGPT4 is very happy to output that string.