6 ms·
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. "
by klyrs 5y ago
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 you mean by [noun in previous sentence]
bot> why are you calling me mean? That really hurts my feelings
It has been a few years, and I feel like a smart bot-writer might be able to leverage something like GPT3 to at least include a little bit more of the current transcript into the prompt to at least superficially address this gaping flaw. Have they?
- hotpotamus 5y agoI'm not trying to be mean or break any HN rules, but did you read the article? It basically covers what you asked and was quite a revelation to me. Others in here point out that GPT-3 is not a chatbot which is good info, but I also wonder if there is anything out there that can even "remember" the context of a conversation it's supposed to be having. Particularly interesting is the question from the article, "who is the current president". Seems like a slam dunk but it could not be answered. Interestingly this is a common question paramedics will give to people who have suffered head injuries to assess consciousness.
- klyrs 5y agoHidden in my comment is a question: has anybody even tried to included previous lines into the prompt, and does that not help? Asking here because there are a lot of ML nerds who could probably do a proof of concept in a few hours, if not a few lines of code, so there's a decent chance I'll get an informed response.
- hotpotamus 5y agoMy guess (and I'm certainly no expert here) is that not only have people put a few hours or lines of code into it, but probably many a PHD thesis as well. It's an intensely researched topic which makes the article even more surprising to me.
- DrammBA 5y agoSo that's a "no" for the "but did you read the article?" question?
- klyrs 5y agoThat is interesting. Please go on.
- AshamedCaptain 5y agoA lot of chatbots from 20 years ago did that. They would even use it as a escape route, and would just bring back the previous topic of conversation (from say 10 lines ago) when they could not figure out something to come up for the current prompt. It really didn't help the illusion; rather it was even more obvious.
- YeGoblynQueenne 5y agoIn most of the conversations with GPT-3 that you see online what people do is that they take the output from the first prompt and include it in the next prompt. That is how GPT-3 can keep the thread of the conversation going without changing subject constantly. This is also why those conversations are relatively short, because as you can imagine, if you keep feeding the output to the language model as input, you run out of input size very quickly. So, yes, that's the done thing already.
- thaumasiotes 5y ago> I also wonder if there is anything out there that can even "remember" the context of a conversation it's supposed to be having. IBM's Watson project involved an internal model of the world, though I have no idea if that got specialized into a local model of an ongoing conversation. Same type of thing though. Having a context of this type available is actually necessary if you want to synthesize sentences in a language with definiteness-marking, such as English, because definiteness is marked according to whether the noun in question is or isn't already part of the context.
- Sohcahtoa82 5y ago> Particularly interesting is the question from the article, "who is the current president". Seems like a slam dunk but it could not be answered. Interestingly this is a common question paramedics will give to people who have suffered head injuries to assess consciousness. Not sure you'd still be able to use that question considered the insane QAnon followers that would answer "Trump". Though I suppose even that answer would still at least prove some level of consciousness.
- lupire 5y agoGoogle Meena https://ai.googleblog.com/2020/01/towards-conversational-agent-that-can.html?m=1 https://ai.googleblog.com/2020/01/towards-conversational-age...
- jhbadger 5y ago>who is the current president? I wonder what happened to IBM Watson's technology that was good at answering trivia -- it actually won a game of Jeopardy! a decade ago (although of course it answered things with a question as per the rules). I know that they weren't that successful at applying it to biomedical research as they had hoped, but it would seem it would be better at chatbots than GPT-3 and other deep-learning models.
- Spooky23 5y agoThe problems with chatbots have nothing to do with tech. You implement chatbots to bend the demand curve to higher cost channels. Period. If you can get 5% of the user base to wait 15 minutes by sending them to a chatbot tarpit, you will have eliminated the need for a dozen call center agents at peak times. Providing good service with a chatbot is possible, but the work and organization required is extensive, and it’s cheaper to throw semi-skilled humans at it!
- jazzyjackson 5y agoI suppose it didn't work with less than 15 terabytes of RAM
- AussieWog93 5y agoFor anyone wondering, this is not an exaggeration. Watson actually had 15 (or 16, depending on the source) TB of RAM, as well as a 2880-core processor.
- dane-pgp 5y ago> this is a common question paramedics will give to people who have suffered head injuries to assess consciousness. I can't help wondering what the assessment would be if someone answered "Who is the current president?" with "Watson Bakery". https://www.youtube.com/embed/cS1Tmz816hk https://www.youtube.com/embed/cS1Tmz816hk
- aidenn0 5y agoTo be fair, if I were asked the current PM of Japan the only name I could come up with is Abe, and that's only because he was PM for so very long.
- jazzyjackson 5y ago"chatscript" is a great dialogue engine where language patterns land the state machine into a "topic" which it gracefully backs out of (keeps a topic hierarchy) when it runs out of responses in that topic / branch of conversation. It also includes the most robust, hand-tuned lemmer and stemmer + disambiguation language parser built on top of WordNet. really good stuff, written in C so its light on memory and responses are instantaneous. https://github.com/ChatScript/ChatScript https://github.com/ChatScript/ChatScript
- Sharlin 5y agoThe GPT2 version of the AI Dungeon [1] could keep track of context for maybe couple of lines at a time. I've heard the GPT3 version is substantially better. The problem is, of course, that these "AI" chatbots on websites, marketing buzzwords notwithstanding, have very little to do with state-of-the-art machine learning, and are indeed unlikely to be any more sophisticated than the original ELIZA for the most part. [1] https://play.aidungeon.io/ https://play.aidungeon.io/
- 300bps 5y agoThanks for the link to that! I tried it out just now and I have to say that I am not impressed. What will you do? You get longsword from the ground. You get your longsword from the ground. You hold it in your hands and examine it for a moment. It's a very sturdy weapon. The knight then grabs his longsword and bows. Huh? Why did it grab the bow as well? You inventory. You have a crossbow. Huh? I thought I just grabbed my longsword and bows. What will you do? You don't have enough experience to know what a good shot is. You get longsword and bow. You grab the bow and the longsword. You examine the crossbow. It's a bolt-action weapon. Huh? I said to get the longsword and bow since it said it wasn't in my inventory and it added all this extra stuff. You inventory. You have a crossbow, a short sword, a longsword, a lantern, a flint and steel, and a quiver of bolts. What will you do? You load the crossbow. Huh? I just said Inventory and now I have all this extra stuff it never mentioned before and then unsolicitedly told me I loaded the crossbow. I have to be doing something wrong. Zork from 1980 was better than this.
- mithr 5y agoI think things started to go wrong here > The knight then grabs his longsword and bows. "bow" here probably refers to "a bending of the head or body in respect, submission, assent, or salutation", rather than to the weapon. The knight grabbed his longsword and bowed to you. Your confusion then confused the game. This does expose an interesting common "failure mode" for these types of bots: they assume that all conversations are rational, and are conducted in good faith (as in, no one misunderstands or is actively trying to confuse them). Janelle Shane likes to test AIs that attempt to describe the content of images by asking "how many giraffes are in the picture?", and since most of these have been trained to answer what is in a picture rather than what's not in a picture, they tend to assume that there must be some number of giraffes, and provide answers like "there are 1.5 giraffes".
- imglorp 5y agoThe self driving people talk about time domain occlusions, which seems closely related to conversation context. Say you're driving along and you see a pedestrian about to cross but something blocks your view for a moment. Your mental model will remember that person, continue to speculate on their motion and intention, consider if their vector might intersect yours, and also know when it's okay to flush them from attention.
- throwaway675309 5y agoEliza was modeled after Rogerian styled psychology, it uses an almost Socratic method to illicit and draw out responses from users.
- not2b 5y agoEliza had a little, but very limited memory, and sometimes when no rule would match it would say "Earlier you said <x>". There was some dumb heuristic about which <x>'s to save for this purpose.
- Phelinofist 5y agoTwo months ago I started working at a company that develops the new speech assistant for a large German car manufacturing group. The assistant will support such queries pretty well. [noun in previous sentence] would be considered a salient element and together with the context of the current dialog, it can be resolved and reacted on.
- masswerk 5y agoFor the record, ELIZA does have a short term memory, but it is only used as a fallback, when other rules fail to match. (E.g., in productions like, "Earlier you said *" or "What else comes to your mind when you think of *" by the famous DOCTOR rules set.) ELIZA/DOCTOR manages somewhat to succeed with this, because it pretends a setting, where there are no outside references at all, beyond the very limited purpose of carrying on with a conversation that is meant to explore internal subjects in ever-shifting context. (You could say, the shift in context is the actual progress and success of the conversation, while having to pick up an earlier thread actually marks a point of failure.) Outside of this restrictive setting, the shortcomings of the absence of a consistent world model become obvious very soon.