Chatbots: Still dumb after all these years • Hacker News Board

Re: Chatbots: Still dumb after all these years

#161

Post by JamesBarney » Thu, Jan 06, 2022, 7:22 PM UTC

fault1 wrote: post #148

p1esk wrote: post #137

Earlier quoted context omitted.

I'm not 100% sure, but I think in 2012 chatbots still used handcoded rules. So we switched from handcoded rules to trainable neural networks in just a few years. Models like GPT-3 are interesting because they are conceptually simple, and are able to do (previously) complicated tasks without any explicit training (e.g. simply by processing lots of text and trying to predict the next word). This is a huge advancement i…

I remember chatbots in 2000 using hidden markov models (and sometimes even neural nets), so from that standpoint, they aren't exactly new. The models are just larger now.

Never underestimate the effects of scaling on intelligence.

Caenorhabditis elegans neurons work a lot like our neurons, we just have a lot more of them.

Re: Chatbots: Still dumb after all these years

#162

Post by not2b » Thu, Jan 06, 2022, 7:23 PM UTC

klyrs wrote: post #12

My favourite flaw of chatbots is exposed by ELIZA. Not chatting with ELIZA, (though, it does suffer this flaw) but using responses inspired by that program. "Please elaborate on that" or "tell me more about [noun]" etc. Bots appear to have zero lines of short term memory, and utterly fail to pick up a reference to the thing that they just said. My favorite being bot> [something plausibly human-sounding] me> What do y…

Eliza had a little, but very limited memory, and sometimes when no rule would match it would say "Earlier you said ". There was some dumb heuristic about which 's to save for this purpose.

Re: Chatbots: Still dumb after all these years

#163

Post by adament » Thu, Jan 06, 2022, 7:25 PM UTC

p1esk wrote: post #137

chubot wrote: post #118

Earlier quoted context omitted.

It was more than 10 years ... People reacted the same way to SHRDLU 40+ years ago, and 20 years ago when I encountered it: https://en.wikipedia.org/wiki/SHRDLU There's definitely an illusion where we anthropomorphize very simple software, ascribe intention to it, etc.

I'm not 100% sure, but I think in 2012 chatbots still used handcoded rules. So we switched from handcoded rules to trainable neural networks in just a few years. Models like GPT-3 are interesting because they are conceptually simple, and are able to do (previously) complicated tasks without any explicit training (e.g. simply by processing lots of text and trying to predict the next word). This is a huge advancement i…

I think MegaHAL is (mostly?) just a markov model, and I think that was not exactly new when I looked at it around 2002. As I recall it was easier to distinguish from a human than Eliza, since it had a greater probability of spouting nonsense, but it was still amusing and fun to play with.

Personal anecdote: I read one of Ray Kurzweils books in school back then, completely misunderstood how neural networks worked and ended up making a markov model chatbot with single word states.

Re: Chatbots: Still dumb after all these years

#164

Post by Barrin92 » Thu, Jan 06, 2022, 7:26 PM UTC

after having played AI dungeon for a while these issues become obvious very quickly. These large models are good at producing text in some probabilistic manner but they don't understand anything at all.

Other than for obvious commercial and hype reasons I have no idea why there is still so relatively little recognition of criticisms that Gary Marcus and others provide because it's blatantly obvious if you interact with these systems even for a few minutes that they're to intelligence what a trojan horse is to a real one.

There's no real understanding of the world, no coherence, no real ability to grasp logic or have some ontology of the world and that makes reasoning about the world other than producing some 'nice sounding BS' as the article puts it impossible.

Re: Chatbots: Still dumb after all these years

#165

Post by Ozzie_osman » Thu, Jan 06, 2022, 7:27 PM UTC

acidburnNSA wrote: post #78

In the early 1990s, Sound Blaster Pro sound cards came with a "psychiatrist" chat bot called Dr. SBAITSO which would reply in text and actually speak it out loud (showcasing the TTS). As a kid, I thought it was amazing, and I spent countless hours trying to converse with it. From a superficial level, modern chat bots seem roughly as good as the old doc.

I came to post the same. SBAITSO was obviously pretty simple but it had enough tricks to convince you, superficially, that there was some intelligence there, and you could spend hours talking to it.

Re: Chatbots: Still dumb after all these years

#166

Post by throwaway0a5e » Thu, Jan 06, 2022, 7:27 PM UTC

300bps wrote: post #44

pwinnski wrote: post #13

Having worked in ML at two different companies now, I think that people interpreting model output as intelligence or understanding says much more about the people than about the model output. We want it to be true, so we squint and connect dots and it's true. But it isn't. It's math and tricks, and if human intelligence is truly nothing more than math and tricks, then what we have today is a tiny, tiny, tiny fraction…

We want it to be true, so we squint and connect dots and it's true. I've had numerous people look at the actions of software I wrote over the last 30 years make comments like, "Ohhhhhh it probably did that because it knew x". The software had no such functionality. Seems to be a natural thing for people to be overly optimistic about software's capabilities.

People always over-estimate the complexity of stuff they don't understand.

If I had a nickle for every time an internet commenter attributed intentional design to a feature that's simply the designer copying established default practice...

Re: Chatbots: Still dumb after all these years

#167

Post by drakonka » Thu, Jan 06, 2022, 7:29 PM UTC

I was obsessed with chat bots as a teenager and used to make them on a website called The Personality Forge. It was all hard coded responses with some very minor scripting involved to do things like fill in the speaker's name, or skew the bot's emotions in relation to the speaker, and have them choose different words based on their emotional state.

I loved reading through my bot's transcripts. They were all recorded, and sometimes people stumbled across him and didn't even realize he was a bot. This happened extremely rarely...most of the time, his responses were perfect examples of dumbness as shown in this article. He would have flashes of greatness in an ocean of bot stupidity.

Years later, I was excited to notice that a single line from my one of my bot's Chatterbox Challenge transcripts was included in a book[1] (which was seemingly a collection of papers). He was mentioned in a paper called "Chatterbox Challenge as a Test-Bed for Synthetic Emotions", a tiny mark left on the world by my little bot. He may be dumb, but I was still proud. I even thought about buying a copy of the book, but it cost $200 and I decided I wasn't that committed.

[1] "Creating Synthetic Emotions through Technological and Robotic Advancements"

Re: Chatbots: Still dumb after all these years

#168

Post by jkonline » Thu, Jan 06, 2022, 7:29 PM UTC

Wait, I still don't know:

"Is it safe to walk downstairs backwards if I close my eyes?"

Granted, I'm more biased here than most, I suppose. I emergently received 10 screws and 2 plates to repair my broken ankle, after falling down my home stairs last year. I wasn't walking backwards, nor were my eyes closed.

Re: Chatbots: Still dumb after all these years

#169

Post by Sohcahtoa82 » Thu, Jan 06, 2022, 7:30 PM UTC

qaq wrote: post #28

well wait for Web3 chat bots

I can't tell if this is satire. Poe's Law in action here.

Re: Chatbots: Still dumb after all these years

#170

Post by a-dub » Thu, Jan 06, 2022, 7:31 PM UTC

yes, you can trip them up quite easily... and yes, they're not "thinking."

but, i have been really impressed by how far the google assistant has come. it's still a dumb machine, but a dumb machine that can find precise and factual answers for specific questions via a voice interface. it's, dare i say, useful!