Wavebox Podcast: Work Life Browser. Episode 2 (Transcript)

What is ChatGPT, and why is everyone talking about it? Wavebox Co-Founders Tom and Rhys provide a background to ChatGPT and discuss how browsers can use it to help people be more productive. There are also updates on the recent revamp of the Wavebox Settings UI and Google's Manifest Version 3 roll-out for Chrome Extensions.

Transcript

[00:06] Rhys: Welcome to the Wave Box Podcast. This podcast is a discussion between the cofounders of Wavebox. Now, Wavebox is a browser for work. It's focused on productivity, and it's pretty revolutionary. It's a feature-rich Chromium browser and it's built for productive working across things like Google Workspace, Microsoft Teams, Monday, Atlassian, Asana, the list goes on. Pretty much every app you use to get work done will work better in Wavebox. Now, today, it's myself and Tom.

[00:40] Tom: Hello.

[00:41] Rhys: And the main course of what we're going to be discussing today is GPT, or ChatGPT as it's known on the internet. But I guess first of all, we should say Happy 2023.

[00:53] Tom: Happy 2023. So we've kind of got a few things on the go at the moment. In the last podcast, we spoke about the new App Store and the management screens, which we were hoping to ship before Christmas. And but they're looking closer to the end of this month. Now, the things we did manage to ship over the New Year period was the new setting screens.

[01:12] Rhys: They look really good.

[01:13] Tom: They do look good, don't they?

[01:14] Rhys: So the difference here is you imagine a normal setting screen, which is like a list of things you can toggle off and on, very functional, which is what we had. But unless you knew what you were doing, maybe a little bit complicated.

[01:26] Tom: Yeah. Well, you see, I remember when the settings fit on one screen, that day has long gone, there's pages and pages of them. So having that discoverability, because now you've got kind of mock versions of the UI and you can pick the colors or turn things on and off, and it shows you in real-time exactly what's changing.

[01:44] Rhys: Yeah, it's a good way of discovering things that maybe you didn't know. You could fit it in Wavebox. You kind of scroll through, you see a little picture of, oh, I can change how this icon looks like, and I can use the toggles. And like you say, it gives you the feedback there, and then it does.

[02:00] Tom: So that's shipped out now, starts out available to everyone. And the other thing that we managed to ship out just after Christmas was Chrome 109. So that's going up to stable for everyone.

[02:10] Rhys: That's good.

[02:11] Tom: That brings us usual kind of Chrome improvements, fixes and security.

[02:15] Rhys: That does, yes. And one actual Chrome improvement is a more pragmatic approach to manifest version three.

[02:21] Tom: It is. So, for those of you that don't know, Chrome extensions have this manifest which defines what they can do and the functionality that they have. Google has been quite heavy in pushing this new version of the manifest, which adds some new features, but also removes some others. And you might have heard about it, you might not, but it has been quite controversial because they do remove some of the things, the things that ad blockers use, and replace them with other things which are more performant but can be less capable. Google did have a deadline for this, didn't they, for not accepting any new extensions?

[02:59] Rhys: They did. And they're also going to basically have a hard mode of only running version three in kind of the beta and kind of dev versions of Chrome this month. But they've relaxed that and they've basically said there are no new deadlines, but we're just going to take a pragmatic approach. Everything new should be on version three, but we're not going to kick off version twos kind of prematurely. So basically, as every extension is updated, over time, those updates will move to version three.

[03:35] Tom: Version three isn't all bad. We've got a couple of extensions that run in Wavebox, so things like when you press the Wavebox Connect button or the Smart Notes button, those are actually little extensions and we moved those over from manifest two to three, I think middle of last year sometime. And the thing that we saw was we saw a drop in memory.

[03:55] Rhys: Yes. I think the nice thing about version three, to go technical on these things is the way it persists in the background has changed. So rather than always kind of waiting around to see if things happened, it's now kind of quietly sleeping and then woken up and told of new things as they happen by Chrome. That probably means you can have more extensions. I guess maybe that's kind of does. I guess maybe that's something we'll explore in 2023 and what that means.

So we kind of titled this podcast all about GPT, or ChatGPT, because it doesn't matter where you turn on the internet, certainly over the kind of Christmas period into the new year, it was everywhere. It's an AI tool, an artificial intelligence tool, and it's basically allowing people to see a peek into the future of what a capable AI tool can mean in terms of how it works with the different tools that you use and the work that you're doing. And it's surprisingly good and efficient, isn't it?

[05:01] Tom: It's amazing. And Microsoft is one of the backers behind it.

[05:05] Rhys: So it's run by yeah, it's a company called OpenAI. It's been working for quite a few years. Microsoft yes, is a big investor in it. It basically has developed these models where they've trained it on pretty much different data sets available on the Internet. Originally, things like images, but more now on all the things, say in Stack overflow and all the crawled data that you get from common crawl and other sources. And previously it was closed down, you had to apply to try it out, but over the last couple of months, anybody can actually log in and try it out directly with a version called ChatGPT.

[05:51] Tom: But this is a little bit of a downfall because if you try and go there, quite often they're at capacity.

[05:55] Rhys: If you try it today, yes, it's a bit of a potluck because it is so popular. I think the popularity lends itself to why people are amazed by its capability. You think about, say, the original kind of Eliza type AI, where it was very closed, it basically just repeated back what you asked it. Or if you use some of the automated chat tools on some of the support sites, you can get quite frustrated by its capabilities. You know, it's just doing a lookup and trying to cordon you off into some direction through a decision tree. But ChatGPT is completely different. You can actually ask any question to it and it will respond in a quite capable way.

[06:36] Tom: Well, this kind of comes through to one of the uses that we've used it for, and I know one of the big websites, CNET, has also alluded to the fact that they've done this. So when we were playing about with it late last year, we were looking at ways, basically ways that we could use it. And it's very good at writing content. So we tried it with a couple of the blog posts that we've got on our website and decided to see if we could get it to write some of the content. For those, don't get me wrong, you can't just ask it to write a blog post and for it to spit it out. It's not that good and nor is the content great. But what you can use it for is kind of like a starting step. It changes from looking at a blank page, having to decide what to write, to being like a brilliant I can split these four paragraphs out, get rid of one, edit another one, jiggle them around a bit, and hey, I started off with something and I can make something much better. And certainly for us, because we're a small team, it meant that we could basically write content a lot quicker. So we used it or kind of trialed it for a couple of blog posts. And that seems to have worked. I don't think anybody noticed!

[07:44] Rhys: I think this is it. If it blends into the existing content that we have, that means the level of quality is roughly the same.

[07:51] Tom: Yeah, this is it. And we had to edit it. It wasn't like, just let it go and off it goes. It's very handheld.

[07:57] Rhys: But the core of that is the work that you put into the prompt, isn't it? I think, like you say, just asking it to write something and not giving it any kind of guidance.

[08:06] Tom: Don't get me wrong, it will write something for you, but it'll be something like, I don't know about the weather.

[08:11] Rhys: Yes, but if you actually do give it, say, the title and a few of the summary points of what the paragraphs are in a blog post, it will then fill it out for you. And I think there's quite a bit of technique in. I think they call it prompt engineering. How you can actually warm up or excite the AI sufficiently so that when it actually generates content for you, it's the content that you're looking for, rather than more generic content than spew out.

[08:36] Tom: Don't get me wrong, it doesn't know everything. So one of the things that I was using for was to write steps for a knowledge base and asked it to how to do blah, blah, blah, and it gave me some steps and I was like, yeah, these look great, brilliant. I'll add these into the knowledge base. And I was like, well, no, I should check what they are first. And they were wrong. So it's not perfect, you have to check what it's saying. It's kind of there as an aid, as a helper. It doesn't give you the oracle of truth.

[09:06] Rhys: No, this is it. But I think content writing very good niche that chat GPT or GPT in general is very good at. And if you're listening to this and you haven't tried it out, go to OpenAI and you can just sign up freely and use ChatGPT for some content that you're trying to write yourself. Now, other than that, we've been asking the question and we did this over the last year and actually just a little bit before about if you take a tool like GPT or chat GPT, what are the implications for a browser such as Wave Box. And what are the intersection of useful kind of extensions you can apply to the kind of tooling around a browser that ChatGPT can help. And we did a closed kind of trial of this, didn't we, in Wavebox, using the API access we were given by OpenAI to GPT. We actually called it Brainbox. So some of you may have been involved in that trial. And what we did there is exposed GPT at the browser level.

[10:09] Tom: Yeah. And it's more like a playground to kind of play about with a few ideas and to see what was possible. It's kind of like the blank piece of paper versus the guided tour type territory. The playground that we went for was more the blank piece of paper. I think we found some good uses out of it, and some things that we can definitely apply in terms of actually helping you do things.

[10:35] Rhys: Yes. And some of the examples that we tried out was, for example, being able to suggest some alternative content from what you've selected, use the general method of selection in the browser to then feed that into ChatGPT. So we had about five or six kinds of different canned queries, if you like, or canned prompts that would take what you've selected and then run them from there. So we found that that was useful in some instances, but the question in our minds was always what works better at the browser level and what should work better at the app level? And I think what's come clear now with some of the apps themselves incorporating GPT. An example here is Notion. It had GPT added late last year, and it can do things like help you generate a list of things in Notion that you want to do, or maybe summarize something that you want to do in Notion, and that I think works quite well at the app level.

[11:43] Tom: Yeah, and it's quite targeted there as well, because you're writing something, you need more input on that.

[11:49] Rhys: Yes, but I think there's valid things that have come out from that trial is to on the browser level, chat GPT can be a sidekick to you, basically, that can be there as a way of helping you through work that is more generic. And I think one of the main ones that we saw that was useful, that we'll be looking at this year in 2023, is things like summarization and general question and answering, kind of more like a chat around a subject that you're working on. To give you some concrete examples, let's say I'm working in a SAS tool, and I've actually got a whole range of output from this because it's given me everything in the last year that has been generated as like a report. And I just really want to summarize this because what I'm meant to be doing is kind of generating a paragraph that summarizes this whole big document that I'm working on. Now, if at a browser level, I can just say, look, take what I'm working on now, so in this app or tab that I've got, just summarize into one paragraph for me, and then allow me to have that in my clipboard, because then I can hop over into another tool. It could be I'm feeding this into maybe I'm going into Notion from this reporting tool and I can summarize straight away. So having ChatGPT in Wavebox or in a browser of that level we think would be quite an advantage as a useful tool. Because again, it avoids you having to do the hard work of generating that summarized paragraph. Summarizing it for you, but then allowing you to edit it because it's just in your clipboard as you paste it into the new tool of your choice. So we think things like this would work quite well.

[13:47] Tom: Yeah. So these are things that we're going to be kind of experimenting with a bit further this year.

[13:51] Rhys: It is, and from the early experiments we're doing, one of the other benefits we've seen, we're based in the UK, we're multilingual because we have Welsh and English and ChatGPT, or GPT works across different languages automatically. You don't need to tell it okay, this is a Spanish source, I want it summarized in Spanish, please. It knows the language, it detects what it is, and will actually continue the language for you. So this is the reason we're thinking it's applicable as a generic tool on the browser level, because it doesn't need you to guide it so much in terms of the contents that it's looking at. It also is very good at knowing the structure of these things as well. So if you're looking at a structured kind of table of things, it can actually understand a little bit around that table. So if you're looking to summarize the totals in something, then it'll actually understand the format for you and guide you that way.

[14:52] Tom: Again, this kind of like leads into the old kind of Excel world where you would have a bunch of data and you need it transformed into another way. It can kind of be helpful in those kinds of things, can it?

[15:03] Rhys: Can, yes. Because the experiments we've done is things like, let's say I've got the output of one tool, which has given me a JSON structure. JSON is like a JavaScript kind of structure or squiggle brackets, but it's a common kind of way of moving data from one system to another. But let's say I want to change that into something that is more row orientated, like Excel CSV is the term for that. That's quite a complicated process. If you wanted to code that, you'd be looking at something like a python or a node script to do it for you. But in our tests, we can actually give that JSON input to GPT and then one example row of what you're looking at and it'll actually pick up the data for you. So in a JSON structure you might have something like coupon code when it was qualified as a date, and a number of entries, let's say, as different kind of worded elements in that JSON structure. And then in my CSV, my row, I actually give the titles of that and the first example of what I have. And then you just throw the data at it and then actually it's almost.

[16:16] Tom: Like magic, isn't it? It's like having a little helper that just kind of comes along and does it?

[16:19] Rhys: Yeah. So we think as a way of looking at allowing you to take data from one web app into another web app. GPT has a role there to play with again at the browser level because it can see what one app is doing, understand what the other app needs, and be that middle person, that's kind of going through those format changes between them.

[16:37] Tom: I mean, one example that you could do with this is I've kind of moved a bunch of stuff from Todoist into Click Up this week. And I did that manually. I just copied and pasted it. It took me about ten minutes, but that's the sort of thing that you could do. It's like, hey, I've just copied a bunch of stuff. It's not very well formatted and I need it kind of in bullet points because ClickUp knows how to pick up bullet points and create ten tabs.

[17:00] Rhys: Yes, that's right.

[17:00] Tom: It's that kind of thing that it's like, yeah, that took me ten minutes, but with a bit of a help from a tool, it might only take me two.

[17:07] Rhys: Yeah, these things again, from the output of our initial trials that we did kind of last year and the year before, these are the kind of things we're looking at seeing if we can add to Wavebox this year.

[17:19] Tom: Do you know what it is really good at? I just had a little play the other day and I wrote some code, intentionally made it so it had errors and I thought it'll never be able to tell me what the errors are. It's just not clever enough. And I thought I'll try it anyway. And do you know what? The results were surprisingly good. It was like, no, you can't do that. You can't do this. That's an error. And I was like, what? How does it know? I really didn't think that it would be able to do that, but you know, that's kind of one. You know, just as an aside, one thing that it is good at, it.

[17:50] Rhys: Really is a good helper. And I think it's because it's trained on all of the questions people have asked on the internet and it's seen all the answers. It can measure all that together and know what you're trying to get at and then quickly give you the output that it's seen to work in the past. So again, thinking about the browser level, one thing we're looking at again from a Wavebox point of view is could we make a GPT general helper available? So Wavebox knows that you're working in a certain SAAS tool. Say it's zero and you don't quite remember how to do something. Now, normally you'd go into Google, you'll type a search engine, you'd kind of get the first few results, click through those tabs and try to figure it out yourself. Well, the alternative is that in waybox we basically tell GPT, it looks like you're working in zero on this kind of a page. And then you'll just type one question.

[18:48] Tom: You see, this works great until in the scenario, whereas right in the blog post it gives you a bunch of.

[18:54] Rhys: Instructions which aren't right to be qualified.

[18:58] Tom: Yeah. And then you're looking for the settings button which doesn't exist.

[19:01] Rhys: This is it. But I think there is something there.

[19:04] Tom: No, there is.

[19:05] Rhys: When you are looking for a solution for something, if it's better than a Google search and hunting through five spam blogs and a weird old one that again doesn't have the right answer, then maybe it's a good thing to have.

[19:22] Tom: I knocked it there. But the other thing is, half the time you search Google for something and you'll find somebody on Reddit who said, yeah, you should do blah, blah blah, and it's like, yeah, no, that doesn't exist anymore because it's three years old or whatever.

[19:33] Rhys: Exactly. And I think if you're listening to this and somebody thinking, no, we agree this is actually something quite cool. And if you're happy to try these things out with us, knowing that there are some caveats that's not 100% solution, then please get in touch, tweet us, email us hello@wavebox.io or support@wavebox.io and we will bring you on board the trials that we'll be doing this year in terms of chat, GPT and Wavebox. Because I know that asking the question is a good interface into chat GPT. Yeah, and you can do even things like, let's say I've got Wavebox, I'm looking at a huge Wikipedia article and I'm looking for one specific thing. I could actually then ask GPT to a specific question on the content that I've got in front of me in Wavebox and it is capable then of understanding that and answering your question so that you don't have to comprehend and get the answer yourself. And again, it maybe saves you between, you know, a ten minute task of reading comprehending an answer and a 1 minute task of just asking ChatGPT.

[20:42] Tom: This is it. You know, it's not, it's not the answer to everything, but it can help you along in some of these instances and it might just be a case of that one piece of information that you need out of this massive article. It's enough to point you in the right direction so that you can just search in the article for the bit that it's kind of pulled out or whatever. So there's a lot of scope there for different things that it can do.

[21:04] Rhys: There are the things that OpenAI themselves are kind of opening up more and more now is a way of you taking the core model of GPT and then training further that model on a specific data set that you have. So, again, bringing it back to Wavebox, what we could potentially look at this year is allowing OpenAI's kind of SDK access for data that you have in Wavebox. So this might be a couple of web apps that you have now you're logged in, it has access to that through Wavebox. We can actually use that data to train a model just for you that's locked down to you, that allows you then to have specific GPT, that understands the domain that you're working on, that allows you then to do all sorts of things that it's applicable solely to the work that you're doing or your company is doing now. Currently that's a little bit bespoke and there's costs and stuff involved in there, but I think that's a general direction that it's worth exploring in because it really increases the productivity of what you can do as a company if you have AI that's trained specifically on the data that you have.

[22:18] Tom: I mean, like a prime example of that, of how it could actually work, is we've got our support desk and in there, there's thousands of tickets that people have asked there's questions and there's answers that we've put to it. Let's say we were onboarding a new person in the company and we wanted to get them up to speed quickly. It would be a case of if we could train it over all the support tickets and all the answers and questions and then give that to the support person and say okay, deal with some new tickets, ask the GPT thing to get a new ticket and they can ask it. It might give you something that's a reasonable answer and could even point you to the ticket where that was and it's like, yeah, that's the sort of range of things where it would actually work really well.

[23:02] Rhys: Yes, definitely. And you can extend that further if you also then gave it access to all our knowledge base articles. You're in a point now where you can probably expose a ChatGPT for Wave Box directly on our support page. So the initial point is that you will ask ChatGPT, how do I have a split screen working? Or how do I change the sleep settings on a specific group? It can actually then take you through and answer that support straight away for you. Yeah. So again, all of these things we are looking at in 2023, the question always is what works at the browser level rather than the product or kind of app level. And if it's something we think will work in the browser level, we will try that out this year with Weight Box. And again, if you're listening to this and you want to help us with.

[23:55] Tom: Getting contact, we might have some things that we're happy to kind of try between select a set of users and see what works.

[24:02] Rhys: Yes, that's right. And then the output of course is a productivity tool that is even better, understands your work and can just basically help you get the work done rather than being a hindrance in the way. Yeah, we'll be podcasting again in a couple of weeks, so hope you like that podcast talking more about releases in Wavebox that we have.

[24:23] Tom: And there's also some stuff ongoing that we can't really share at the moment, but there's some new exciting things there is hoping over the next few weeks, so they might be almost in beta next time.

[24:33] Rhys: There we go. If you're using a tool, you can subscribe to this podcast, please do so. Until the next time. All the best. Bye.