A reflective essay exploring how classic LLM failure modes---limited context, overgeneration, poor generalization, and hallucination---are increasingly recognizable in everyday human conversation.
Agree. For example; the amount of times we correct our own speech before ‘releasing it’ is staggering. We have a ‘stochastic parrot’ mechanism build right into the hearth of our own cognition and it generates the same problems for us. ‘Hallucinations’ are build into a statistical model. It takes a lot of culture/rules and energy to constantly adjust(habituate to expectations/environment into the ‘norm’. People that have fallen out of normal social environments know how difficult human interactions can be to learn/overcome.
Current llm’s doesn’t have the ability to do these micro-corrections on the fly or habituate the corrected behavior through learning/culture etc.
‘Context length’ is also directly mappable to human cognitive load, where chronic stress tends to shorten our ‘context length’ and we lose overview in a split-second, and forget the simplest things. ‘Context length’ are for an llm, roughly equivalent to our ‘working memory’.
However, compensating systems are already being designed. Just like life/evolution did, one by one, these natural tendencies from statistics will be fixed by adding more ‘cognitive modules’ that modulate the internal generation and final output…
Right, I think the key difference is that we have a feedback loop and we’re able to adjust our internal model dynamically based on it. I expect that embodiment and robotics will be the path towards general intelligence. Once you stick the model in a body and it has to deal with the environment, and learn through experience, then it will start creating a representation of the world based on that.
Agree. For example; the amount of times we correct our own speech before ‘releasing it’ is staggering. We have a ‘stochastic parrot’ mechanism build right into the hearth of our own cognition and it generates the same problems for us. ‘Hallucinations’ are build into a statistical model. It takes a lot of culture/rules and energy to constantly adjust(habituate to expectations/environment into the ‘norm’. People that have fallen out of normal social environments know how difficult human interactions can be to learn/overcome.
Current llm’s doesn’t have the ability to do these micro-corrections on the fly or habituate the corrected behavior through learning/culture etc.
‘Context length’ is also directly mappable to human cognitive load, where chronic stress tends to shorten our ‘context length’ and we lose overview in a split-second, and forget the simplest things. ‘Context length’ are for an llm, roughly equivalent to our ‘working memory’.
However, compensating systems are already being designed. Just like life/evolution did, one by one, these natural tendencies from statistics will be fixed by adding more ‘cognitive modules’ that modulate the internal generation and final output…
Right, I think the key difference is that we have a feedback loop and we’re able to adjust our internal model dynamically based on it. I expect that embodiment and robotics will be the path towards general intelligence. Once you stick the model in a body and it has to deal with the environment, and learn through experience, then it will start creating a representation of the world based on that.