I’ve been doing quite a bit of reading recently to prepare for BiZ Gear Up on the 24th, and part of that led me to read a bit more about the early history of chatbots – I hope you’ll enjoy this one!
We take a brief moment to move away from the hype that is OpenAI‘s ChatGPT, and take a brief intermission as we make a small trip back in time.
Picture this: it’s the year 1966, and a computer scientist at the Massachusetts Institute of Technology‘s Artificial Intelligence Laboratory named Joseph Weizenbaum has just created something remarkable – the world’s very first chatbot: ELIZA.
Image credit: Public domain via Wikimedia Commons
Just consider this example of a conversation from Norbert Landsteiner’s 2005 implementation of ELIZA, and you can see what it was capable of.
ELIZA was designed to simulate conversation by responding to typed text with pre-programmed phrases and questions.
But what made ELIZA so special was that it was programmed to mimic the conversational style of a therapist, in particular a Rogerian therapist.
Users could “talk” to ELIZA about their problems and concerns, and the chatbot would respond with empathetic and non-judgmental phrases like “Tell me more about that” or “How does that make you feel?”
It wasn’t just a simple question-and-answer program – it was designed to provide a sense of emotional support and understanding that reflects interestingly on the ways that people derive comfort from self-affirmation.
Weizenbaum didn’t intend for the chatbot to be taken very seriously, calling it a “parody” in his 1976 book “Computer Power and Human Reason”… But the way that the chatbot was received was far from just a parody.
The response to ELIZA was overwhelming.
People were fascinated by this new technology that could seemingly understand and respond to their thoughts and emotions, and the program quickly gained popularity as people tried the chatbot.
But perhaps what’s most remarkable about ELIZA is that it wasn’t just a novelty. Weizenbaum’s creation laid the foundation for decades of research in the field of natural language processing and artificial intelligence.
ELIZA was the first chatbot, but it certainly wouldn’t be the last – and its legacy lives on in the many conversational AI programs we use today, in our Bings, Bards, ChatGPTs, Claudes, and the many more that exist and will exist today.
Today, it exploded and hit 1k, and it shows no sign of stopping; the hype is real!!
Why?
A mixture of Manglish, an anonymous dude who prompted OpenAI’s ChatGPT with the legendary Manglish prompt, the combined efforts of Kenneth Yu Kern San and Jornes Sim, and, I guess, a huge, huge dose of the wonders of ChatGPT x)
My own small contribution (lol!)
I guess I’ve evolved into the comments guy – take a look here if you want đ
Ah, what can I say?
It’s truly fascinating to be at the beginning stage of a revolution.
First off, congrats and creds to Pang for being infinitely better than I am at managing communities at the moment – something I’m definitely looking forward to learning more in the days ahead!!
Second off, I care a lot about the deeper significance of things – and I’m incredibly glad that this is one of the many things that’s starting off AI on the right foot in Malaysia, my home country – where this will go and what will happen I have no idea, but really look forward to watching what the world’s going to bring đ
Amongst other things, I guess it’s brought a ChatGPT Plus subscription for which now, in order to sign up, you’ve got to join a waitlist.
Can’t wait to see what’s next, and thank you based Sam Altman and OpenAI for creating this gift to humanity đ
Could you potentially write a Masters thesis with ChatGPT?
It so happened that someone on the ChatGPT Malaysia Facebook group had asked about the same thing, so I thought ok, let’s make it happen.
Anyway, I was curious about whether it was actually possible, so I decided to give it a go.
Hereâs what I asked:
Okay, so at the very least the software proposed a bunch of topics that seemed kind of plausible and interesting.
Anyway, since I’m involved in the education industry and AI â based learning is very interesting to me, I decided to ask ChatGPT to follow up on #7, as follows:
Okay, wow! I had sources too! This was getting interesting! But thenâŠ
I looked at this, and I was captivated: Was I on my way to get a Masterâs degree for this man?
No, wait. Wasnât this even better? Wasnât this thing essentially describing the process of creating a personalized new education technology company for me???
I set out in earnest, yearning to go where no man had ever gone before!
Okay, seemed great so far! I ran out of words, though, so I asked ChatGPT to continue:
Okay, uhâŠ
Do you see what I’m seeing here?
Rather than actually writing the thesis, ChatGPT was malingering â it was casually not doing what it was told to do, and presenting me with some nonsense summary!
That wonât do! You think just because you’re an AI assistant you get to be lazy?!
I asked it to continue, and provide results in detail.
For about five, this seemed really really plausible, so I was happy againâŠ
For about five minutes, before my skepticism began again.
âŠSo I checked the references, only to realize that they mostly couldnât be found anywhere.
Okay, I was thinking to myself.
This is a wonderful software, I declared, trying to beat back the cognitive dissonance.
Surely the third page will be a little bit better? So I thought.
At this point, I realized â ChatGPT had failed.
The two methodology sections contradicted themselves, and there wasn’t a possibility of reconciliation unless I proceeded to prompt chatGPT with the specific information that it actually needed, which I decided not to because the rewarded yielded by that effort would actually be better spent writing the thesis if I actually had a clear idea of how to do so.
So, how do we answer our research question?
With a solid no.
As you can see, thereâs a word limit for responses, which means that you will have to re-prompt ChatGPT, which is likely going to lead it to drift from the original prompt.
ChatGPTâs memory for prior responses is about 4000 tokens (words) and it will not completely remember everything that you told it before unless say, you intelligently summarize.
There is no guarantee that the logic or factuality of your piece will be valid or that even any of the sources that you cite will be accessible or even relevant to what you are writing about, as you see from the questionable sources.
Sorry to those of you out there hoping that ChatGPT was going to help you get your Masterâs degree, but it’s not gonna happen right now.
Even if you can though, should you? I guess thatâs up to each person to decide, but what I would say is that submitting something AI generated for a degree means that you didnât get the degree â the AI did and got certified and you did not.
Let me not moralize this or romanticize education, but approach the matter in a logical way â when this starts to happen on a large scale, if it does happen, I can imagine that companies or other institutions that used to take these degrees seriously will simply no longer take them seriously, thereby causing degrees as a whole to become about as worthless as MOOCS among prominent companies (i.e. companies that actually generate large amounts of business and have a vested interest in hiring actually talented people) and leading to what we already see, to a degree, in institutions such as tech companies and start-ups⊠Whereby many of these companies donât pay the most attention to the particular degree that you received, but rather whether you are capable of demonstrating the specific skills that they are looking for and communicating your perspective in the course of an interview in which there is no opportunity to make use of AI software.
How will artificial intelligence change not just education, but also the job market at large?
Weâll be finding out, and weâre going to be in for a wild, wild ride!
I’ll have lots more to say about this in the days ahead, so if you would like to read about the intersections between AI, writing, and education, do consider dropping me a follow and Iâll see you in my next pieces!
Stop what youâre doing right now and download Poe!
âŠUnless youâre using an Android phone, in which case maybe you can drop what youâre doing right now, cry, buy an iPhone, then come back.
âŠOKAY NO DONâT KILL ME AAAAAAAAAAAAAAAAAnd now, dear readers, we talk about Poe!
Generated with Midjourney!
If you donât know what Poe is, Poe is not this Poeâââitâs question and answer platform Quoraâs brand new AI baby, now available on iPhone and iPad!
Initially, buying into the hype of the Microsoft and Google AI arms race, I thought that Poe was just a Chatbot that was trained on Quora data and by extension the millions upon millions of questions and answers that it contains, but I realized that it wasnât (although that’d be cool, though uh… I’m not sure how that’d work considering that nowadays a ton of questions on Quora are kind of powered by people trying to copy-paste ChatGPT in order to rank?).
Anyway, Poe isnât exactly a chatbot in and of itself – it is what I would describe as a Chatbot aggregator, which means that it collects several different Large Language Models (LLMs) into a single interface; no wonder, considering the fact that the project is called Poe because it’s short for “Project Open Exchange”, which I guess alludes to the fact that the premise is that you can ask questions to a variety of different AI chatbots and receive answers relatively quickly. Already on the platform you will see three main chatbots. The first in the lineup is Sage, the next is Claude, and the last of the bunch is Dragonfly; Sage and Dragonfly are powered by OpenAI, and Claude is powered by Anthropic P. B. C. and its Constitutional AI framework.
Each has its own methods of going about generating conversations, with OpenAI going for RLHF and Claude through the channel of constitutional AI and harm reduction, but I think you should try these for yourself to see the differences, and perhaps also read this article from Scale to see a detailed comparison.
Youâll see each of these on the left end of the app and you can select them and start talking the way you can with ChatGPT.
Dragonfly is fast, but not significantly faster, I think, compared to Sage at the moment – although I’d probably appreciate that a bit more in the event that this app becomes super popular.
The question you might have is that, apart from the number of models available, how is any part of this experience different from ChatGPT or from Claude, though?
Letâs compare.
So here, I ask ChatGPT why the sky is blue, and receive a pretty reasonable response.
It takes about 10 seconds to complete. On the other hand, when we do the same thing with Sage, which is also an OpenAI model, we get the following in 4-5 seconds generation time – which is significantly faster than ChatGPT, but is not bound by the fact that ChatGPT has millions of users concurrently using it… But oh wait.
See the blue links?
When you hit theseâŠ
âŠSage begins to elaborate while providing further links as well.
To be exact, these aren’t exactly links – when you click them, they prompt the algorithm with a new text prompt that’s related to the words that you searched for, in turn providing more context and elaboration on what you asked about and helping you to fill up those knowledge gaps real quick.
I think that this is super cool and definitely a step in the direction of the future of search, because it does mean that you can go down a daisy chain of conversations, ask questions, have them answered, and be prompted to ask about the things that you don’t know about; I can totally imagine this as a question and answer service and can even imagine something like this replacing Quora, if not for the fact that subjective experiences and updated information were still relevant and important.
While at the moment this is constrained to material within the training set and by extension data limited to 2021, one can only imagine what might happen at a later point when or if these models receive internet access!
Apart from that though, a cool feature of Poe is…
Social:
One of the fun things about Poe is a whole social element that’s integrated into the app itself.
Thereâs a feature that allows you to share the things that youâre looking at on your feed so that people can see what you generate – essentially, whatever you ask, once you hit the Share icon and then “Share on Poe”…
…Which will let you create a little set of posts that will just randomly show up on people’s feeds, kind of like an implementation of a little TikTok-esque feed full of the prompts that people are publicly sharing around the world and that people can upvote and downvote at their leisure, along with little bubbles that show the kinds of prompts that people have shared at any given point.
I think it’s kind of cool that you get to see these questions, mainly cause you get to see how people around the world are choosing to interact with AI, which creates a (human) communal experience so we can see how people are choosing to prompt things and keeps things light and fun đ
Here are a couple of examples:
It’s cool to see what people are thinking and writing about, isn’t it? Very much in the spirit of Quora, I think – it does make me wonder if that’s the next evolution of the platform, albeit I still imagine that it might be difficult for algorithms to source personal data or subjective opinions into their training data or make people willingly choose to submit it; perhaps the platforms will coexist? I don’t know.
Anyway, since the thing’s called Poe, I decided to ask Poe to go right ahead and role-play Poe and the Raven.
Uh, towards the end I went and created a bit of an Oxford Union style debate there.
Anyway, I found the chatbot aggregation, links, and social features to be pretty cool and solid features in the app at large, and these all make me really wonder how the platform’s going to evolve in the days ahead.
Some small concluding thoughts:
I guess that Poe’s trying to unite all the different AI models and chatbot applications in one space, and that does kind of make sense, but I would guess that some of the companies that are generating LLMs simply won’t feel the incentive to participate (or won’t have the capital x investment to afford the training costs), while the companies that have generated these APIs would be happy to charge Quora or perhaps at a later point the end users of Poe for using those APIs when they eventually monetize (it’s already happening with ChatGPT), they’ll still be keeping their latest and greatest models for their proprietary usage and for paying users so that these users can use their products beforehand.
Still, what weâve got here is definitely pretty great already, and I look forward to seeing how this platformâs going to develop in the days ahead!
To end this little exploration, I couldn’t resist making a rap battle about sentient AI with a tiny bonus at the end.
