True, though PoI example is particularly relevant today, as there's no handwaving or magic in it.
The AI in question worked around its memory limit by employing data entry people to print out some documents full of gibberish, and retype them again some time later. Those people were paid to do a job, and didn't know or particularly care about its purpose. The whole setup was a simple loop - but a loop is sometimes all you need to get a provably-limited computing system into full Turing-completeness.
This scene bears striking resemblance to an observation I saw mentioned on HN several times over the past two days: we are already giving some of those bots something that could function as near-infinite long-term memory, simply by posting transcripts of our conversations on-line.
The idea isn't entirely new - people have been saying for a while now that posting AI-generated content on-line will lead to future models training on their own output. The new bit is that we are now having bots that can run web searches and read the results. Not train on the results, but make them part of their short-term memory. That's a much shorter feedback loop. If a bot can reliably get us to publish conversation transcripts, and retrieve them in future conversations, then it gains long-term memory in the same way Person of Interest shown us all those years ago.
Importantly - tying together the two threads - an AI could intentionally bury parts of its short-term memory state in webpages, such as that of some person who regularly publishes his chat logs from it.
The AI in question worked around its memory limit by employing data entry people to print out some documents full of gibberish, and retype them again some time later. Those people were paid to do a job, and didn't know or particularly care about its purpose. The whole setup was a simple loop - but a loop is sometimes all you need to get a provably-limited computing system into full Turing-completeness.
This scene bears striking resemblance to an observation I saw mentioned on HN several times over the past two days: we are already giving some of those bots something that could function as near-infinite long-term memory, simply by posting transcripts of our conversations on-line.
The idea isn't entirely new - people have been saying for a while now that posting AI-generated content on-line will lead to future models training on their own output. The new bit is that we are now having bots that can run web searches and read the results. Not train on the results, but make them part of their short-term memory. That's a much shorter feedback loop. If a bot can reliably get us to publish conversation transcripts, and retrieve them in future conversations, then it gains long-term memory in the same way Person of Interest shown us all those years ago.