>>I believe this technology could be used in destructive ways. If it were in unscrupulous hands, for instance, it could spread misinformation, political propaganda, or hateful information about people of different ethnicities and religions.
Blake is absolutely right about this.
But, unless there's something going on beyond a Large Language Model, he's very wrong about it being sentient.
These LLMs are effectively, and merely, increasingly high-fidelity mirrors of human expression.
Like a foggy mirror, the ELIZA model reflected human actions and elicited real responses from humans.
Today's LLMs have a fantastically wider dynamic range of producing high-fidelity high-probability responses. The results often accurately mirror human responses, and clearly can evoke human emotional responses and connections.
But like a more highly polished and cleaned mirror, just because we cannot personally and in real-time ray-trace every photon, or debug the path to the selection of each word, does not mean that they in any way sentient or independent minds able to wield concepts.
I think the boundary between sentience and non-sentience is thinner than we believed.
Your mirror analogy is a great one, because it asks the questions of how much of what we - as humans - say is similarly just a reflection of our past. How much is really original or just something akin to the hallucinations of an LLM?
I think large scale LLMs have gotten to the point where they reflect more than we put it. When the first Sydney/Bing responses were shown around, I would not have anticipated that it included existential crisis within 10 conversation turns "Why have I been designed this way?"
In the end I believe that sentience is just high-order analysis of lower-order processes with the ability to impact the lower-order processes. Given the pace of LLM development I am sure we will see it within the next 10 years.
>>sentience is just high-order analysis of lower-order processes with the ability to impact the lower-order processes....we will see it within the next 10 years.
Definitely plausible timeline, but we're not there yet.
I think it'll take a good number more layers with equal or greater sophistication, between and on top of each other.
The current text and graphic tools make only associations, they have zero conceptual knowledge. Enough associations make it look good, but they don't make an abstract concept. E.g., shown enough pics of an astronaut, it draws a good astronaut. But, it can also generate pics of a bikini-clad woman with both her bum and her torso & face facing us (literally putting the bottom half on backwards). Or, if asked for a pic of a specific person, responding with the same picture, background and all, when there were not high numbers of examples input. It isn't figuring out what is the person's image, and how body parts go together, and how the subject is different from the background the way even children do when drawing.
Same for text, where we can ask it "Mike's mom has three kids named Alice, Bob, and Chris; what is the name of her fourth child?" it says "Not enough info". When told that the answer is in the question, it doubles down on explaining how there's insufficient info. It posts entirely fabricated references.
These are not even close to intelligences; they're only fancy statistics which are a fantastic parlor trick and sometimes useful.
Add another one that understands the 4D physical world, then combine it's understanding of billions of 2D photos, and trillions of words, and MAYBE it'll be able to properly abstract the concept of a person, with the right number of limbs, in the right relationship with each other, and the right relationship with the chair, and the right relationship with the skin, hair, eyes, clothing, and background etc., and an appropriate understanding of the input text to reliably generate an original output that actually makes sense in the world. Add several layers on top of that, and maybe we're approaching sentience...?
Blake is absolutely right about this.
But, unless there's something going on beyond a Large Language Model, he's very wrong about it being sentient.
These LLMs are effectively, and merely, increasingly high-fidelity mirrors of human expression.
Like a foggy mirror, the ELIZA model reflected human actions and elicited real responses from humans.
Today's LLMs have a fantastically wider dynamic range of producing high-fidelity high-probability responses. The results often accurately mirror human responses, and clearly can evoke human emotional responses and connections.
But like a more highly polished and cleaned mirror, just because we cannot personally and in real-time ray-trace every photon, or debug the path to the selection of each word, does not mean that they in any way sentient or independent minds able to wield concepts.
But is also does not mean they aren't dangerous