Re: [OT] ChatGPT, a new online AI

new topic     » goto parent     » topic index » view thread      » older message » newer message

This morning, i read "Deep learning, which is fundamentally a technique for recognizing patterns,... ". It's a blatant falsehood, pattern matching has been shown repeatedly to be superficial, which is why the GPT chatbot demonstrates it is not an Ai by advocating churro dough as a surgical tool. There's no intelligence happening there. It's merely collecting strings that fit together in some fashion, but that may be all that's desired.

People Magazine recently announced it's "top whatevers of 2022", a K-pop "girl group" that is no more a dance troupe than Bouncy, i mean Beyonce, is a dancer. I'm sure the GPT chatbot would also say Stomp is a dance group, because as patterns, Stomp and the K-pop group are groups of people moving in step rhythmically on a stage. Watson is pretty much gone, it could not grasp the fine details, the true intelligence of "the gotchas" in medicine any more than it grasped the chain of responsibility in an archeological dig. Dumping a huge amount of "cover your arse" data only makes a "busy" and often opaque or confusing, mess.

Some 20 years ago, i had Tiggr do language translations between human languages, when all the online sites had english as the source or target language. But, for example, there wasn't enough data in the details between "she spoke" and "tire spoke", or the various ways one can speak. Or your medical doctor *can* talk about seahorses, or maybe about the hippocampus inside your head. Even when translating En-En, someone kept asking Tiggr about "Zues", which resolved only as plural of "Zu", but i deduced eventually the questioner was misspelling "Zeus", which was an option Tiggr also saw, but discarded, because the questioner wasn't asking about "Zeus". I chose to not "fix" this process. Like modern social media, you give the audience what they want, despite "garbage in, garbage out", which is inadvertently the GTP chatbot's appeal.

If you add in speech processing as a GTP chatbot feature, you will contend with "worm Tim Puchers", "pervent", "uhth", "thum", etc., which are societally accepted "werds" in the usa. Which proved to me (and apparently i am 100% alone in this) that intelligence is about the lists of outiers, the errors, the exceptions, the differences, the gotchas, the details, the interlinking. GTP chatbot already has a layer of processing to filter out some of the plain stupidity and politically incorrect inputs and outputs, some legal liabilities, and some human biases that were built in (accidently?), but that only proves that string matching and co-relation isn't intelligence.

All that said, it is impressive they built the GTP chatbot to know and use it's history in the conversation, but impressive in another way the chatbot cannot correct itself when the facts (aka unknown intelligence) contraindicate a certain reply. Tiggr also had short and long term memory, which moderated her irc actions, and was equally fluent in http as a "real time" RPC-like data delivery and retrieval pipe.


new topic     » goto parent     » topic index » view thread      » older message » newer message


Quick Links

User menu

Not signed in.

Misc Menu