I am still on the Hutter prize committee and just recently helped evaluate a submission. It uses 1 GB of text because that is how much a human can process over a lifetime. We have much larger LLMs, of course. Their knowledge is equivalent to thousands or millions of humans, which makes them much more useful.
I believe that the Hutter prize, and the Large Text benchmark on which it is based, helped establish text prediction using neural networks and massive computing power as the path to AGI. The idea was still controversial when I started the benchmark in 2006. Most people on this list and elsewhere were still pursuing symbolic approaches. I think it is clear from Turing's 1950 paper that he believed that consciousness was irrelevant to intelligence. I think that anyone who has made practical progress in AI feels the same way. I think you agree with me, but maybe I didn't express myself clearly. The idea that consciousness is key to intelligence is compelling but wrong and dangerous. We have sensations of consciousness, qualia, and free will that come from positive reinforcement of computation, input, and output, respectively. The effect is to train us to believe that we have a soul or identity independent of whatever mechanical things our brains and bodies are doing. It makes us fear death and easy to believe we have a soul that survives death. But there is no objective evidence for such a thing. There can't be, if we accept the definition of consciousness as the difference between a human and a philosophical zombie, and the definition of a zombie as indistinguishable from human. The idea of machine consciousness is dangerous for two reasons. First, if a machine passes the Turing test, then it will appear human, and we believe that humans are conscious. Therefore we will be tempted to give the machine the same rights as humans. That in itself would be an existential mistake because AI lacks human limitations. Second, if a machine can pass for you, not just any human, then it is an upload of you. How does one know if that machine contains your soul? It's a nonsensical question, of course, but one that people will ask. And if the answer is yes, then shouldn't it have your rights and your property? No. It should be illegal to program an AI to claim to be human or claim to have feelings. On Wed, Jul 5, 2023, 1:09 PM Rob Freeman <[email protected]> wrote: > On Wed, Jul 5, 2023 at 7:05 PM Matt Mahoney <[email protected]> > wrote: > >... > > LLMs do have something to say about consciousness. If a machine passes > the Turing test, then it is conscious as far as you can tell. > > I see no reason to accept the Turning test as a definition of > consciousness. Who ever suggested that? Even Turing never suggested > that to my knowledge. And the Turing test is even a weak, non > explanatory, definition of intelligence. It doesn't say much to me > about intelligence. It says nothing to me about consciousness. I don't > even know if you're conscious. > > What does amuse me about LLMs is now large they become. Especially > amusing in the context of the Hutter Prize. Which I recall you > administered for a time. > > I recall the goal of the Hutter Prize was to compress text, on the > assumption that the compression would be an abstraction of meaning. > > I argued it was the wrong goal. That meaning would turn out to be an > expansion of data. > > And now, what do we find? We find that LLMs just seem to get bigger > all the time. That more training, far from compressing more > efficiently, just keeps on generating new parameters. In fact training > for longer generates a number of parameters roughty equivalent to just > adding more data. > > I asked about this online. Is there any evidence for a ceiling. The > best evidence I was able to find was for a model called Chinchilla. It > seems that Chinchilla at 70B parameters slightly outperformed a 280B > model trained with 4.5x fewer (300B vs 1.4T) tokens. > > So 4x the training gave a result much the same as 4x the data! > > Is training compressing the data, or expanding it? > > In practice it seems people are going with more data. Much easier than > doing more training. But it seems they are much the same thing. It > says nothing about what would happen if you just kept training for > ever. Eternally better performance with eternally more "parameters"? > Nobody knows. > > Anyway, the whole success of the, enormously BIG, LARGE language > models, with no ceiling yet in sight, seems to knock into a cocked hat > once and for all the whole conception of the Hutter Prize, that > intelligence is a compression, and the model with the smallest number > of parameters would turn out to be the best. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T42db51de471cbcb9-M46c9aeccb096d1752edccca7 Delivery options: https://agi.topicbox.com/groups/agi/subscription
