4 ms·

Nah, it’ll be just a new, higher level of abstraction (and provided we even solve the hallucination problem, which is not a given). Engelbart’s and Parkinson’s

Nah, it’ll be just a new, higher level of abstraction (and provided we even solve the hallucination problem, which is not a given). Engelbart’s and Parkinson’s laws at it once again. In other words, the new work fill the same time, and we’ll just be doing 10x of it (if we can solve the current challenges with LLMs). The human mind is unsurprisingly adaptable to changing conditions and great at abstract thinking, which is what ironically distinguishes it from these machines. We also tend to severely underestimate our own abilities and overestimate the importance of the tools.

rootusrootus 3y ago
> We also tend to severely underestimate our own abilities and overestimate the importance of the tools. Even on HN, this is rampant. I wonder, is it because people are so used to interacting with most other humans solely through a chat interface that they perceive ChatGPT as being close to a real AGI, almost indistinguishable from human intelligence?
- dinvlad 3y ago
  Pretty much :-) Text generation (and rudimental comprehension) is perceived to be a sign of higher intelligence, simply because it's our natural human UI. Ironically, we tend to limit our imagination on what _we_ could do that computers won't, while simultaneously overuse it to imagine what they _could_ do but won't. One very good example of this is humor - LLM models at present are terrible at generating anything but dad jokes, and are even worse at their attempts to explain "deep" humor. Gary Marcus had a recent episode on it, very interesting insights from top experts in the field: https://podcasts.apple.com/us/podcast/can-ai-make-you-laugh/id1532110146?i=1000613196164 https://podcasts.apple.com/us/podcast/can-ai-make-you-laugh/.... They make a case that LLMs were specifically optimized for text generation and not explanation, while humans are almost the opposite. I believe somewhere in there is also our tendency for confirmation bias and overlooking imperfections where we don't expect some. AI may generate nonsense details, but when we look at the overall text, we don't notice them since we're just awed by the output, and think if it looks flawless it must be flawless. Same thing as with mentalist tricks. And finally, we only think of our intelligence as thoughts carefully animated into sentences, while there's this whole "iceberg" level of feelings and unconscious cognition, which we don't even notice ourselves, but it would be extremely hard to reproduce in AI and we have not even the slightest idea how.

rootusrootus 3y ago

> We also tend to severely underestimate our own abilities and overestimate the importance of the tools. Even on HN, this is rampant. I wonder, is it because people are so used to interacting with most other humans solely through a chat interface that they perceive ChatGPT as being close to a real AGI, almost indistinguishable from human intelligence?

dinvlad 3y ago
Pretty much :-) Text generation (and rudimental comprehension) is perceived to be a sign of higher intelligence, simply because it's our natural human UI. Ironically, we tend to limit our imagination on what _we_ could do that computers won't, while simultaneously overuse it to imagine what they _could_ do but won't. One very good example of this is humor - LLM models at present are terrible at generating anything but dad jokes, and are even worse at their attempts to explain "deep" humor. Gary Marcus had a recent episode on it, very interesting insights from top experts in the field: https://podcasts.apple.com/us/podcast/can-ai-make-you-laugh/id1532110146?i=1000613196164 https://podcasts.apple.com/us/podcast/can-ai-make-you-laugh/.... They make a case that LLMs were specifically optimized for text generation and not explanation, while humans are almost the opposite. I believe somewhere in there is also our tendency for confirmation bias and overlooking imperfections where we don't expect some. AI may generate nonsense details, but when we look at the overall text, we don't notice them since we're just awed by the output, and think if it looks flawless it must be flawless. Same thing as with mentalist tricks. And finally, we only think of our intelligence as thoughts carefully animated into sentences, while there's this whole "iceberg" level of feelings and unconscious cognition, which we don't even notice ourselves, but it would be extremely hard to reproduce in AI and we have not even the slightest idea how.