Hallucinations might be a feature, not a bug
Consciousness, as we know it—and we know very little—may have developed our concept of self as an evolutionary need to simplify the image of our brains. When you walk, your brain doesn’t consciously contract the glutes and the quadricep muscles to move forward; it just does it. Similarly, the ego, or the sense of self, that highly conscious beings experience, is a way for our brains to create a model. An abstraction.
This modeling is evidenced in humans with severed brains, particularly in those with disconnected left and right hemispheres. In these individuals, consciousness becomes somewhat “split,” with each hemisphere developing its own reality and experiences. When they are asked verbally—through speech, which is handled by the left brain—why they performed an action, such as following visual instructions to walk, they often cannot acknowledge that they 'saw' those instructions because vision is handled by the right brain. But the peculiarity is that instead of answering that they do not know why, the left brain would start confabulating things to fill in the gap, making things up and giving excuses for their actions.
The brain constantly fills in information, often compromising the factuality in the process. It does this through hallucinations. This pattern of compensating for a lack of knowledge isn’t limited to humans; LLMs (Large Language Models) hallucinate as well. Newer models with reasoning capabilities, perform much better at reducing these hallucinations. But will we ever be able to create artificial intelligence that never hallucinates?
In the latest research papers published by Anthropic, they uncover this process of 'making things up' that leads to hallucinations inside LLMs as well.
Claude, on occasion, will give a plausible-sounding argument designed to agree with the user rather than to follow logical steps. We show this by asking it for help on a hard math problem while giving it an incorrect hint. We are able to ‘catch it in the act’ as it makes up its fake reasoning...
This could be due to various reasons that we have not yet uncovered. But one explanation is that increasing efficiency comes with some decrease in factual correctness. Another way to look at it is that hallucinations are fundamental to an intelligence with finite resources.
Making mistakes is a sign of intelligence; making things up to justify them is the folly of humanity.
Want to chat about this? Just email me!