This is spot on - the big AI companies are trying to sell something they cannot even define. AI is a fine marketing tool, but the claim of AGI is so much larger. The 'general' part requires an incredible jump on the level of logic/psychology/epistemology style claims. Within epistemology its a well known result that induction (the formal tool of generalization and abstraction) cannot be derived non-circularly. Induction is treated as a given in formal systems, and no amount of data fed into a LLM will reproduce this ability. Generalized abstraction is upstream of data, not downstream of it.
I actually think that within a given scope current AI models are already better at data processing than a person, but the human aspect it cannot mimic is that the human's reasoning is entirely unbound to that scope. So its no surprise their focus is on tokens used. Since that's what they sell, they will always advocate for it. Its going to be the responsibility of managements to create criteria for productivity that are actually useful for their business. Its just like in the 2000's when the focus was on lines of code written, which simply ended up incentivizing creating long and inefficient software.
Completely agree on the induction point. And to be clear, the argument is not that AI does not work. AI is real, useful, and here to stay. The problem is that we are building dependencies on an assumption of continued exponential development on a very compressed timeline. When that timeline turns out to be longer than expected, the dependencies are already baked in. We are also probably in a collective psychosis where AI is being mandated for everything and anything regardless of whether it makes economic sense for that specific use case. Most things do not need it. Most things would be cheaper and sometimes faster without it. The real risk will not be the current token cost. It will be the token cost when the subsidies end and the infrastructure has already been built around the assumption that it stays cheap forever.
The gaslighting read is sharper than it looks — a model tuned to agree reflects your premise back as confirmation. That's not intelligence, it's a mirror with no spine. The fix isn't a smarter model, it's a human who can tell reflection from truth. Wrote a book on exactly that, free Kindle thru 6/3: amazon.com/dp/B0H3HY8W9F
This is spot on - the big AI companies are trying to sell something they cannot even define. AI is a fine marketing tool, but the claim of AGI is so much larger. The 'general' part requires an incredible jump on the level of logic/psychology/epistemology style claims. Within epistemology its a well known result that induction (the formal tool of generalization and abstraction) cannot be derived non-circularly. Induction is treated as a given in formal systems, and no amount of data fed into a LLM will reproduce this ability. Generalized abstraction is upstream of data, not downstream of it.
I actually think that within a given scope current AI models are already better at data processing than a person, but the human aspect it cannot mimic is that the human's reasoning is entirely unbound to that scope. So its no surprise their focus is on tokens used. Since that's what they sell, they will always advocate for it. Its going to be the responsibility of managements to create criteria for productivity that are actually useful for their business. Its just like in the 2000's when the focus was on lines of code written, which simply ended up incentivizing creating long and inefficient software.
Completely agree on the induction point. And to be clear, the argument is not that AI does not work. AI is real, useful, and here to stay. The problem is that we are building dependencies on an assumption of continued exponential development on a very compressed timeline. When that timeline turns out to be longer than expected, the dependencies are already baked in. We are also probably in a collective psychosis where AI is being mandated for everything and anything regardless of whether it makes economic sense for that specific use case. Most things do not need it. Most things would be cheaper and sometimes faster without it. The real risk will not be the current token cost. It will be the token cost when the subsidies end and the infrastructure has already been built around the assumption that it stays cheap forever.
The gaslighting read is sharper than it looks — a model tuned to agree reflects your premise back as confirmation. That's not intelligence, it's a mirror with no spine. The fix isn't a smarter model, it's a human who can tell reflection from truth. Wrote a book on exactly that, free Kindle thru 6/3: amazon.com/dp/B0H3HY8W9F