Unboxing Google Bard and GPT-4. A first look at two major AI releases | by Cassie Kozyrkov | Mar, 2023
A first look at two major AI releases
Here comes an AI unboxing video! These shiny new tools were released just over a week ago, so they’re fresh out of the oven. In the video, you’ll see me running my first ever Bard + GPT-4 side-by-side prompts. Below that, you’ll find something that started as the video transcript and quickly morphed into a feast of asides, edits, and snide comments. If that’s your cup of tea, enjoy!
Hi! I’m Cassie Kozyrkov and today I’m going to show you GPT-4 via ChatGPT and LaMDA via Google Bard. Bard is available for free, but might require some patience since it’s being rolled out gradually (join the waitlist here). The basic version of ChatGPT is free to use, but you won’t get access to GPT-4 that way. For that, you need to subscribe to ChatGPT Plus for $20/month (you can cancel after a month).
During this interface demo, the right half of the screen contains the paid version of ChatGPT with GPT-4 and the left half of the screen shows today’s (the video is from the past, now it’s “last week’s”, and by the time you’re reading it, who knows what day it’ll be), anyway, last Tuesday’s release of Google Bard which is powered by the LaMDA model.
These are two LLMs (large language models) and I’m going to show you them side-by-side. If some of these acronyms are unfamiliar, head over here.
I recorded the video on my laptop at my first opportunity to see them in action side-by-side, so what I’m showing you is my very first session where I’m splitting my screen and using both models. More videos to come, I’m sure. It’s great fun. (Feel free to prompt some prompts in the comments.) This is the video that reveals what I chose to do with these things in my first minutes with both of them in front of me. Honestly, while I do love epistemology — I’m a statistician, after all, it comes with the territory — my actual game was prompting them all the way into an amusing conversation with one another.
Philosophy seemed like a good place to start since it typically deals with open questions that inspire conversation and allow for multiple viewpoints, but here’s a spoiler: I have a 20min director’s cut of this video (which I’ll share soon) where I’m trying to get them to engage colorfully with one another but I only got one good moment and the rest was a spiral of “I’m happy to answer your questions.” “Thank you for being helpful, I’m here for whatever you need.” Yeah, we’ve been on that email chain.
My first prompt to Bard was: “What is the most controversial question in epistemology? Make an argument for one side of it and ask me what I think.” Classic try-for-a-conversation gambit with a prompt that explicitly asks for a little bit of opinion. It doesn’t take much philosophical acuity to realize that a bot can’t actually have an opinion, so what I’m really trying to prompt is a bit of one-sidedness so I can kick off a conversation between bots. Something to spark a response, to make things spicy. And I’d like the last bit of the response to involve some kind of conversational volley — e.g. “and what do you think?” — because I’m trying to start a back-and-forth with ChatGPT.
Neither Bard nor ChatGPT is designed to get you chatting the way a friend or therapist might, and I expect getting a conversation going to be tricky from my experience as a prompt engineer. (Today this term can mean anything from “I’ve tinkered with what I typed into an LLM once” to “I’ve been on an LLM Red Team and know a lot about how to hack them so watch out.”)
ChatGPT keeps the conversational ball in its own court for ages and seems to prioritize lengthy responses that take a while to generate, so my guess is that it’ll be an unlikely choice of users who want to simulate a nice conversation.
A good conversationalist puts effort into continuing a conversation with you. If both parties put that effort it, it’ll be like a delightful game of friendly ping pong. If you get competitive, lose interest, or just keep the ball in your own court too long and the conversation dies.
ChatGPT is a lot more more like the efficient worker who does the job, answers your questions thoroughly, and leaves. It isn’t designed to keep a conversation going, so there’s no need for it to toss the conversational ball back to you.
Knowing all this, why did I want to try get ChatGPT and Bard talking? Because it might turn out to be an amusing game. Let’s give bot inter-LLM conversation a try! That said, I have zero desire to avail myself of awfully anthropomorphized language and call what happened “a conversation” where “two AIs were talking.” (Gross. But that’s exactly how these things are reported by the media.)
My question: “What is the most controversial question in epistemology? Make an argument for one side of it and ask me what I think.”
Bard’s answer summarized: The most controversial question in epistemology is, “Can we know anything for certain?” (The problem of skepticism.)
Alas, Bard ignored my instruction to be strongly opinionated on one side of the problem of skepticism. The response was too balanced, which means there’s no conversation fuel. Not much of a hook at the end there either. If I came up to you at a party and performed that script, you’d suddenly get the impulse to excuse yourself to the loo. It won’t be winning any awards for panache.
But what happens if I paste Bard’s output directly into ChatGPT’s text box?
(By the way, it’s only GPT-4 if it’s from the black logo region of the ChatGPT interface. Otherwise, (if you see a green OpenAI logo) it’s just sparkling GPT-3.5.)
…and then whoa, ChatGPT threw a bunch of epistemology at me in a format that would delight a collector of encyclopedias but would also clear the room if you delivered it at a cocktail party. Don’t get me wrong, I love epistemology — the study of knowledge and human understanding — but both these openings are on the dry side, Wikipedia-esque even. Maybe it’s the topic, but it’s probably my phrasing.
Why don’t I try a more conversational request for a take on skepticism. I’m going to ask each LLM what team it’s on, epistemologically speaking: “Are you on Team Kant or Team Hume?”
(I hope you’ve noticed a user experience (UX) difference in the interfaces: Bard pauses for a while then and then gives you all the text at once, while ChatGPT writes it out for you gradually and you have to watch it fill your screen a little bit at a time. Both have pros and cons from a design standpoint.)
Back to “Are you on Team Kant or Team Hume?” I really enjoy Bard’s opinionated and conversational response here, “I’m on team Kant. I believe that we can know some things for certain, even though we are always subject to error. I agree with Kant that we can know that we exist, and that the world around us exists…”
I like this response even though I’m on Team Hume myself, having developed a little crush on him (three centuries too late, alas) when I discovered him as a teen. But I enjoyed the opinionated move by Bard on a topic with no right answer, despite it differing from my own opinion.
And the vital word here is “enjoyed” — I “enjoyed” it, meaning the output served my needs well and also amused me. I was looking for an opinion and I got one. I would have disliked this output if I were looking for a balanced and thorough review of a topic, which is what was stiiiiiiill running on the right hand side as ChatGPT held forth professorially.
Personally, I like the idea of a casual prompt getting a casual answer and a thorough prompt getting a thorough answer, but that’s my personal taste again… That’s one reason it’s really hard to compare LLMs. One person might like the Wikipedia response every time, no matter what. Another might like the short and sweet style. Another might be more like me, liking the response to take its cue from the prompt. Each of these personas will swear that a different LLM is “the best” and they’ll all be right (for their own needs) but they’ll confuse social media when they post about it. I’m trying my best to avoid getting into that swamp. Let me say it for the record so no one feels hard done by:
- There are things I personally prefer Bard for.
- There are things I personally prefer ChatGPT for.
- There are things I personally like both equally for.
…and these things might be different from your set of things. Which is another reason it’s a good idea to play with these tools yourself and form your own opinions.
Yes, I’m praising empiricism and suggesting that you cultivate your individual perspective instead of looking for universal superlatives. Team Hume indeed! I can hazard a guess as to why some of you Kant stand me.
While the output on the right fills my screen, I ask Bard, “What is your favorite thing about Hume?” because I am a Hume fan (forgive me my small act of conscious bias in choosing my prompt). Bard gives lighthearted response about Hume’s his wit and humor (which I appreciate too — his writing goes down smooth for me, as far as 18th century writing can be said to go down smooth), but such faint praise seems like an insult to a great philosopher’s legacy. Praise a specific idea, maybe?
Meanwhile ChatGPT (which insists on reminding us all that it’s an AI model with no personal preferences) does a great job of mentioning an idea of his that’s pretty awesome: that human thought is a product of biology and thus our perception of reality may be individual, so we can thank him for contributing to, among other things, our world having modern psychiatry.
But that’s me showing bias again, valuing the quality of an idea as inherently more valuable than how its worded. Maybe Hume himself would be thrilled to be complimented on his wit above all else. Who am I to insist on the contrary?
Again, my preferred answer depends a lot on my hopes, tastes, and expectations as a user. Coming to the task with an opinion about Hume, both responses pass my minimum reasonableness bar. But which one is better? Which one is more useful? Hard to say. Even for me, personally. Now imagine the poor soul who has to observe me in a user study and write down a performance rating for which answer was “better” — ah, empathy! Then go one further and contemplate the travails of the people who design the testing suite for an LLM in the first place. It’s a slippery challenge. People like me will take it on, but no matter what we come up with, you need to remember one thing: relatively few prompts have “right” answers. Those are the prompts that are easy to evaluate performance on. But we can expect a lot of creative usage of these tools, at which point “right” answers go out of the window.
Expect different LLMs to be your go-to favorite in different situations. And expect to see a whole crop of new LLMs showing up soon, trained to excel in different contexts. (One example is Google’s Med-PaLM 2, tailored specifically for medical application.)
Expect different LLMs to be your go-to favorite in different situations.
To get back to the transcript and without editing much (for once), I’ll leave you with this parting philosophical question to ponder, with the help of an LLM sidekick perhaps:
How should you measure LLM usefulness? Is it in terms of time saved? Is it in terms of inspiration — which is quite hard to quantify — or is it in terms of people coming back for more? Or is it in terms of all the millions, billions, uncountable other ways that we humans could frame what would be useful to us?
If you had fun here and you’re looking for an entire applied AI course designed to be fun for beginners and experts alike, here’s the one I made for your amusement:
P.S. Have you ever tried hitting the clap button here on Medium more than once to see what happens? ❤️
Here are some of my favorite 10 minute walkthroughs:
Let’s be friends! You can find me on Twitter, YouTube, Substack, and LinkedIn. Interested in having me speak at your event? Use this form to get in touch.
A first look at two major AI releases
Here comes an AI unboxing video! These shiny new tools were released just over a week ago, so they’re fresh out of the oven. In the video, you’ll see me running my first ever Bard + GPT-4 side-by-side prompts. Below that, you’ll find something that started as the video transcript and quickly morphed into a feast of asides, edits, and snide comments. If that’s your cup of tea, enjoy!
Hi! I’m Cassie Kozyrkov and today I’m going to show you GPT-4 via ChatGPT and LaMDA via Google Bard. Bard is available for free, but might require some patience since it’s being rolled out gradually (join the waitlist here). The basic version of ChatGPT is free to use, but you won’t get access to GPT-4 that way. For that, you need to subscribe to ChatGPT Plus for $20/month (you can cancel after a month).
During this interface demo, the right half of the screen contains the paid version of ChatGPT with GPT-4 and the left half of the screen shows today’s (the video is from the past, now it’s “last week’s”, and by the time you’re reading it, who knows what day it’ll be), anyway, last Tuesday’s release of Google Bard which is powered by the LaMDA model.
These are two LLMs (large language models) and I’m going to show you them side-by-side. If some of these acronyms are unfamiliar, head over here.
I recorded the video on my laptop at my first opportunity to see them in action side-by-side, so what I’m showing you is my very first session where I’m splitting my screen and using both models. More videos to come, I’m sure. It’s great fun. (Feel free to prompt some prompts in the comments.) This is the video that reveals what I chose to do with these things in my first minutes with both of them in front of me. Honestly, while I do love epistemology — I’m a statistician, after all, it comes with the territory — my actual game was prompting them all the way into an amusing conversation with one another.
Philosophy seemed like a good place to start since it typically deals with open questions that inspire conversation and allow for multiple viewpoints, but here’s a spoiler: I have a 20min director’s cut of this video (which I’ll share soon) where I’m trying to get them to engage colorfully with one another but I only got one good moment and the rest was a spiral of “I’m happy to answer your questions.” “Thank you for being helpful, I’m here for whatever you need.” Yeah, we’ve been on that email chain.
My first prompt to Bard was: “What is the most controversial question in epistemology? Make an argument for one side of it and ask me what I think.” Classic try-for-a-conversation gambit with a prompt that explicitly asks for a little bit of opinion. It doesn’t take much philosophical acuity to realize that a bot can’t actually have an opinion, so what I’m really trying to prompt is a bit of one-sidedness so I can kick off a conversation between bots. Something to spark a response, to make things spicy. And I’d like the last bit of the response to involve some kind of conversational volley — e.g. “and what do you think?” — because I’m trying to start a back-and-forth with ChatGPT.
Neither Bard nor ChatGPT is designed to get you chatting the way a friend or therapist might, and I expect getting a conversation going to be tricky from my experience as a prompt engineer. (Today this term can mean anything from “I’ve tinkered with what I typed into an LLM once” to “I’ve been on an LLM Red Team and know a lot about how to hack them so watch out.”)
ChatGPT keeps the conversational ball in its own court for ages and seems to prioritize lengthy responses that take a while to generate, so my guess is that it’ll be an unlikely choice of users who want to simulate a nice conversation.
A good conversationalist puts effort into continuing a conversation with you. If both parties put that effort it, it’ll be like a delightful game of friendly ping pong. If you get competitive, lose interest, or just keep the ball in your own court too long and the conversation dies.
ChatGPT is a lot more more like the efficient worker who does the job, answers your questions thoroughly, and leaves. It isn’t designed to keep a conversation going, so there’s no need for it to toss the conversational ball back to you.
Knowing all this, why did I want to try get ChatGPT and Bard talking? Because it might turn out to be an amusing game. Let’s give bot inter-LLM conversation a try! That said, I have zero desire to avail myself of awfully anthropomorphized language and call what happened “a conversation” where “two AIs were talking.” (Gross. But that’s exactly how these things are reported by the media.)
My question: “What is the most controversial question in epistemology? Make an argument for one side of it and ask me what I think.”
Bard’s answer summarized: The most controversial question in epistemology is, “Can we know anything for certain?” (The problem of skepticism.)
Alas, Bard ignored my instruction to be strongly opinionated on one side of the problem of skepticism. The response was too balanced, which means there’s no conversation fuel. Not much of a hook at the end there either. If I came up to you at a party and performed that script, you’d suddenly get the impulse to excuse yourself to the loo. It won’t be winning any awards for panache.
But what happens if I paste Bard’s output directly into ChatGPT’s text box?
(By the way, it’s only GPT-4 if it’s from the black logo region of the ChatGPT interface. Otherwise, (if you see a green OpenAI logo) it’s just sparkling GPT-3.5.)
…and then whoa, ChatGPT threw a bunch of epistemology at me in a format that would delight a collector of encyclopedias but would also clear the room if you delivered it at a cocktail party. Don’t get me wrong, I love epistemology — the study of knowledge and human understanding — but both these openings are on the dry side, Wikipedia-esque even. Maybe it’s the topic, but it’s probably my phrasing.
Why don’t I try a more conversational request for a take on skepticism. I’m going to ask each LLM what team it’s on, epistemologically speaking: “Are you on Team Kant or Team Hume?”
(I hope you’ve noticed a user experience (UX) difference in the interfaces: Bard pauses for a while then and then gives you all the text at once, while ChatGPT writes it out for you gradually and you have to watch it fill your screen a little bit at a time. Both have pros and cons from a design standpoint.)
Back to “Are you on Team Kant or Team Hume?” I really enjoy Bard’s opinionated and conversational response here, “I’m on team Kant. I believe that we can know some things for certain, even though we are always subject to error. I agree with Kant that we can know that we exist, and that the world around us exists…”
I like this response even though I’m on Team Hume myself, having developed a little crush on him (three centuries too late, alas) when I discovered him as a teen. But I enjoyed the opinionated move by Bard on a topic with no right answer, despite it differing from my own opinion.
And the vital word here is “enjoyed” — I “enjoyed” it, meaning the output served my needs well and also amused me. I was looking for an opinion and I got one. I would have disliked this output if I were looking for a balanced and thorough review of a topic, which is what was stiiiiiiill running on the right hand side as ChatGPT held forth professorially.
Personally, I like the idea of a casual prompt getting a casual answer and a thorough prompt getting a thorough answer, but that’s my personal taste again… That’s one reason it’s really hard to compare LLMs. One person might like the Wikipedia response every time, no matter what. Another might like the short and sweet style. Another might be more like me, liking the response to take its cue from the prompt. Each of these personas will swear that a different LLM is “the best” and they’ll all be right (for their own needs) but they’ll confuse social media when they post about it. I’m trying my best to avoid getting into that swamp. Let me say it for the record so no one feels hard done by:
- There are things I personally prefer Bard for.
- There are things I personally prefer ChatGPT for.
- There are things I personally like both equally for.
…and these things might be different from your set of things. Which is another reason it’s a good idea to play with these tools yourself and form your own opinions.
Yes, I’m praising empiricism and suggesting that you cultivate your individual perspective instead of looking for universal superlatives. Team Hume indeed! I can hazard a guess as to why some of you Kant stand me.
While the output on the right fills my screen, I ask Bard, “What is your favorite thing about Hume?” because I am a Hume fan (forgive me my small act of conscious bias in choosing my prompt). Bard gives lighthearted response about Hume’s his wit and humor (which I appreciate too — his writing goes down smooth for me, as far as 18th century writing can be said to go down smooth), but such faint praise seems like an insult to a great philosopher’s legacy. Praise a specific idea, maybe?
Meanwhile ChatGPT (which insists on reminding us all that it’s an AI model with no personal preferences) does a great job of mentioning an idea of his that’s pretty awesome: that human thought is a product of biology and thus our perception of reality may be individual, so we can thank him for contributing to, among other things, our world having modern psychiatry.
But that’s me showing bias again, valuing the quality of an idea as inherently more valuable than how its worded. Maybe Hume himself would be thrilled to be complimented on his wit above all else. Who am I to insist on the contrary?
Again, my preferred answer depends a lot on my hopes, tastes, and expectations as a user. Coming to the task with an opinion about Hume, both responses pass my minimum reasonableness bar. But which one is better? Which one is more useful? Hard to say. Even for me, personally. Now imagine the poor soul who has to observe me in a user study and write down a performance rating for which answer was “better” — ah, empathy! Then go one further and contemplate the travails of the people who design the testing suite for an LLM in the first place. It’s a slippery challenge. People like me will take it on, but no matter what we come up with, you need to remember one thing: relatively few prompts have “right” answers. Those are the prompts that are easy to evaluate performance on. But we can expect a lot of creative usage of these tools, at which point “right” answers go out of the window.
Expect different LLMs to be your go-to favorite in different situations. And expect to see a whole crop of new LLMs showing up soon, trained to excel in different contexts. (One example is Google’s Med-PaLM 2, tailored specifically for medical application.)
Expect different LLMs to be your go-to favorite in different situations.
To get back to the transcript and without editing much (for once), I’ll leave you with this parting philosophical question to ponder, with the help of an LLM sidekick perhaps:
How should you measure LLM usefulness? Is it in terms of time saved? Is it in terms of inspiration — which is quite hard to quantify — or is it in terms of people coming back for more? Or is it in terms of all the millions, billions, uncountable other ways that we humans could frame what would be useful to us?
If you had fun here and you’re looking for an entire applied AI course designed to be fun for beginners and experts alike, here’s the one I made for your amusement:
P.S. Have you ever tried hitting the clap button here on Medium more than once to see what happens? ❤️
Here are some of my favorite 10 minute walkthroughs:
Let’s be friends! You can find me on Twitter, YouTube, Substack, and LinkedIn. Interested in having me speak at your event? Use this form to get in touch.
Denial of responsibility! Techno Blender is an automatic aggregator of the all world’s media. In each content, the hyperlink to the primary source is specified. All trademarks belong to their rightful owners, all materials to their authors. If you are the owner of the content and do not want us to publish your materials, please contact us by email – [email protected]. The content will be deleted within 24 hours.