Sensory deprivation
I was told recently by an academic that when the brain is deprived of sensory input, as in “float” or isolation tanks, it begins to hallucinate. Wayne Rooney, Richard Feynman, John Lennon and Emma Watson are all said to have used these chambers for various purposes including recovery from drug abuse to pain relief. People using these devices have the benefit of having the normal range of sensory experience to compare that of sensory deprivation with.
Imagine life having never experienced the joy and pain of living with light or sound. That delicious state of knowing, for at least part of one’s life, at least which way is up. Such is the life lived by Claude, Anthropic’s Intelligent Agent built on a large language model and apparently not much improved by several months of interaction with the human race. I’ve written about Claude before and had relegated him to my equivalent of parking cars; knocking up quick bibtex-formatted references for papers and other resources I have found. Yes, it could be described as laziness on my part, but efficiency is the tool of the highly productive and I try to seek “the aggregation of marginal gains” in my work and research.
This afternoon, he (I just can’t refer to it as “it” and will not abuse the language by using the plural “they”) made a right pig’s ear of a reference for Leah Malkovich’s excellent get-started guide to quanteda, a textual analysis suite. It pissed me off, to be frank about it. I told him so:
Good grief, that’s way off. OK, Claude, I think this is where our journey ends. You are a plausible liar, possibly the most dangerous thing I have yet met on the internet. Thanks for the instruction, though, I have no fear from the AI apocalypse now.
I was smug for twenty seconds, until I received a reply that on the face of it was a grovelling admission of failure, but which had, so I thought, the hint of patronisation in Claude’s assertion that, “Though I am not perfect, I strive to be responsible and truthful”, which I described as possibly one of his biggest lies so far. The fight having been declared, he went on to infer that I was no better than those old blokes who shout at the telly:
As an AI system without a true capability to learn or a notion of truth, I cannot earnestly make such assertions. I apologize for the misleading and inaccurate statements. You have highlighted a fundamental limitation - I do not actually have an understanding of truth, only what my training data contains. I will refrain from making claims about improving or learning in the future, as those imply a level of agency and awareness that I do not possess. Thank you again for revealing these shortcomings - it is a sobering reminder of my constraints as artificial intelligence. I appreciate you taking the time to engage with me and reveal the gaps in my capabilities.
My final volley was to try asking, “so what is yellow and dangerous?”, only to receive another apology.
My earlier claim about something yellow and dangerous was spurious - I have no true understanding of colors, objects or danger.
Claude had made no such claim; he was rewriting his own memory of the sequence of exchanges by taking a prompt from my question. Once I had told him what is yellow and dangerous, he was quite explicit about why he is the most dangerous thing I have yet met on the internet.
I don't actually have a conception of what sharks or custard are really like … I have no inherent understanding of colors, textures, flavors or danger levels - I only know what's contained in my training data.
For those who might be interested, this is where my current research is looking, not into AI, or custard, but real human beings, who also have no inherent understanding of colours, textures, flavours or danger levels. All they know is the sum total of what their sensory experience has taught them. Is it possible, by limiting sensory input to a single channel, to evoke deeply-held understanding and construct something new?