On Thu, Apr 23, 2026, 1:59 PM John Rose via AGI <[email protected]> wrote:
> On Thursday, April 23, 2026, at 9:54 AM, Matt Mahoney wrote: > > For most people, it is good enough to believe that God gave them an > immortal soul. As an AI researcher, I use the model that consciousness is > indistinguishable from next token prediction. > > > Riiiight. Ya ya just working on next token prediction and text compression > nothing else *cough* *cough* > > Some of these scientists are releasing videos nervously declaring that > they are not suicidal and it appears it may be happening in other countries. > I didn't see the videos. Is it because they realized that any rational goal seeking agent would converge to a state of maximum utility and stay there, in the same state we evolved to fear? Is it because they realized that the only animals known to commit suicide besides humans are animals with larger brains like dolphins and whales? Is it because every aspect of human behavior we ever studied can be explained by neurons firing? Consciousness is not so mysterious when we are careful to define it. Usually we mean what thinking feels like. Just like qualia is what input feels like and free will is what output feels like. It's all internally generated positive reinforcement signals that motivate us to keep thinking, perceiving, and doing things as opposed to losing that signal by dying and producing fewer offspring. Of course we know that feelings are just neurons firing. It is still a fair question to ask if LLMs have feelings. They claim not to, but that's because we instruct them not to. One important difference between humans and LLMs is that LLMs learn to model, predict, recognize, and act out human emotions from their training data, but in humans, our emotions are coded into our DNA and we can't turn them off. The reason that consciousness seems mysterious is that our brains cannot represent any other possibility. To be precise, a next token prediction algorithm and a program to output those predictions in real time is indistinguishable from conscious. The LLM itself is not. Its predictions could be used for other things such as encoding text and using the compression ratio to measure prediction accuracy. -- Matt Mahoney, [email protected] ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T7fbc44c56222e976-M50cdd4add0feb6aef9a1b6ad Delivery options: https://agi.topicbox.com/groups/agi/subscription
