[00:00:00] Nathan Wrigley: Welcome to the Jukebox podcast from WP Tavern. My name is Nathan Wrigley.
Jukebox is a podcast which is dedicated to all things WordPress. The people, the events, the plugins, the blocks, the themes, and in this case how AI works and how it might integrate with WordPress.
If you’d like to subscribe to the podcast, you can do that by searching for WP Tavern in your podcast player of choice, or by going to WPTavern.com forward slash feed forward slash podcast. And you can copy that URL into most podcast players.
If you have a topic that you’d like us to feature on the podcast, I’m keen to hear from you and hopefully get you, or your idea featured on the show. Head to WPTavern.com forward slash contact forward slash jukebox, and use the form there.
Before we begin, just a quick alert that there will not be a podcast next week. It’s summer here and I’m having a few days away, but we’ll be back the week after that.
So on the podcast today we have Dan Walmsley. Dan is a long time user of WordPress, having started using it even before version one was released. With a passion for experimenting with different publishing technologies, Dan eventually discovered WordPress and he’s been using it ever since.
Currently working at Automattic as a code Wrangler, dan is part of the applied AI team. Although the team is relatively new, with only a few members, their mission is to coordinate and guide the various AI initiatives within the company.
Recently he’s been focusing on automating internal workflows and communications. A particularly crucial aspect, given the distributed work set up which spans 70 countries, and multiple time zones.
We start the conversation talking about Dan’s background. He’s recently decided that AI is a truly transformational technology, and so has taken steps to learn the skills needed to understand and implement it.
Dan talks about how Large Language Models work, and how ChatGPT has driven awareness and demand for AI technologies in a way that was almost impossible to predict just a year ago. This has caused many companies to become deeply interested in AI and what it can do for their business workflows.
We get into whether the reality of AI can live up to the hype. Do we have enough understanding of AI to know what its impact will be on the workplace, or are we just in the middle of a media frenzy, which will die down over time?
Dan challenges, the notion that AI will take many of our jobs and emphasizes the economic value that AI can bring.
We move on to explore the differences between site generators and site builders, and Dan introduces the concept of the copilot era, in which website creation can be somewhat automated. He highlights tools like Jetpack AI, which can generate content and modify the tone of voice right inside of WordPress.
Dan stresses the importance of building AI tools with user interfaces that learn from human inputs in order to improve over time. He thinks that companies, which measure user responses and interactions will gain a significant advantage in AI development. While those who fail to improve that AI content generation will be left behind.
Whether you’re new to AI or have been paying attention for awhile, this podcast offers a fascinating insight into its impact on society and how it can accelerate progress in fields like scientific research.
If you’re interested in finding out more, you can find all of the links in the show notes by heading to WPTavern.com forward slash podcast, where find all the episodes as well.
And so without further delay, I bring you Dan Walmsley.
I am joined on the podcast today by Dan Walmsley. Hello, Dan.
[00:04:35] Dan Walmsley: Hello Nathan. Great to be here.
[00:04:37] Nathan Wrigley: Yeah. Thank you for joining us. Dan, I wonder if you wouldn’t mind spending just a very quick moment or two just introducing yourself. Obviously, this is a WordPress podcast. I suspect that today we might stray out of the boundaries of the WordPress ecosystem a little bit. I have a feeling with our preamble talk that we’ve had, that may well happen. Nevertheless, given that it is a WordPress podcast, can you just tell us a little bit about your background, the work that you do, who you work for, that kind of thing.
[00:05:01] Dan Walmsley: Yes. So I have been using WordPress since before version one, or whenever the first version came out. Because I remember back at the time I was playing around a lot with Movable Type and, oh gosh, I can’t even remember the name of all the different things. I’d gone through quite a few different publishing platforms, just experimenting with the web. And I discovered WordPress and I’ve literally still got that same blog, and it’s still on WordPress, and it’s been upgraded through every different version ever since.
I work at Automattic. I am on the Applied AI team. I am a Code Wrangler, or code mangler. We all give ourselves our own titles and mine changes a bit. My colleague calls himself an applied AI artisan. And we’ re a pretty new team. We’ve been around just a couple of months. And we’re very small, as in right now it’s just me and a couple of data scientists. But we have a lot of AI at Automattic. Our team’s job is to sort of try to coalesce, coordinate, guide, align it. So that we’re not just operating at the leaf nodes, that there’s a bit of larger thinking going into things.
And as such, my days are mostly spent building weird prototypes on LangChain and chatbots. The most interesting thing I’ve looked at recently is automating some of our internal workflows and communications. Because we operate async, we’re remote. We’re in 70 odd countries around the world in different time zones. And so using AI to capture people’s knowledge and repeat it later when they’re asleep is pretty useful.
[00:06:25] Nathan Wrigley: When the word Automattic is announced, I usually think of WordPress, but I think I’m right in saying that Automattic is the parent of quite a few different companies. So the connection between WordPress, the open source project, download from .org, may not be quite so obvious. But the implementation, it may well go into some of the SaaS offerings that you’ve got I’m guessing as well.
[00:06:48] Dan Walmsley: Yeah, so we are trying to build out AI infrastructure that really doesn’t have a direct dependency on WordPress. You know, GPUs are GPUs, and we’re running a Python based stack on those, because that’s where a lot of the open source activity is. You might have seen that OpenAI announced some changes to their APIs, and in just a few hours, LangChain had a new release, incorporating those features.
Good luck even finding that in TypeScript, let alone PHP, right? So if you want to move fast, you want to be on the cutting edge, got to stand up a bunch of Python. I’ve built a version of LangChain in PHP that runs on WordPress.com for the purposes of producing knowledge bases from blogs. It’s possible that if it turns out to be useful and reliable, that we’ll open source some of that. But right now it’s just there to provide some quick indexing for chat interfaces.
[00:07:34] Nathan Wrigley: So your team is fairly new. Give us an idea of how old that word new means. Are we going back two years or 18 months or a couple of months?
[00:07:43] Dan Walmsley: Two months maybe?
[00:07:44] Nathan Wrigley: Really, new. Okay. And did that sort of trickle down from the Automattic leadership? Was it that people up there decided that, okay, now we’ve got OpenAI in the space, everybody’s, I mean, literally everybody seems to be talking about it.
I don’t think I’ve picked up a newspaper, certainly an online newspaper, in the recent past without there being some kind of AI story in there. So was it that, or was it more a groundswell of Automatticians saying, look, if we’re going to stay in the game, we need to be moving with this.
[00:08:11] Dan Walmsley: There’s some people who have been pushing on LLMs and transformer technology since pre GPT three or two. Which includes me. When I had my sabbatical a couple of years ago. So Automattic has a three month sabbatical, and I was like I’m going to learn AI. This seems really cool.
So I did Andrew Ng’s Deep Learning course and a couple of other ones. There’s some really great courses out there now, even better ones now, this was about three years ago. And I just thought, oh my god, if this grows up, which it looks like it’s going to, it could be amazing for generating content. It could be amazing for conversational interfaces.
I had a little Roomba running around my house, pretending to be a psychopathic robot with chainsaw arms, when in fact it was a little plastic Roomba. But it was like vaguely self-aware that it didn’t have chainsaws for arms. And so it would be like, when I get my chainsaws back on, you’re a toast buddy.
I had an Australian robot that trundled around, it would try to get you to stop working and go to the beach. But it had no way of getting to the beach, which is hilarious. Anyway, that’s a long way of saying, some of us have been pushing for this stuff for a while, but I think what changed, obviously ChatGPT came out and created a lot of public awareness and public demand and conversation.
People started to see this as a race. Companies started to see this, I don’t think Automattic necessarily falls in this bucket, but a lot of companies started to see this as existential. Either you have an AI plan or you’re dead. And so it made sense to put together a team that’s sort of looking at what is this for the whole organization.
Because like you said, it’s a complicated organization. We’ve got podcasting apps, we’ve got diary apps. We’ve got Woo. We’ve got Day One and all these different things. Sensei is a learning management platform. And so we really needed to figure out how we could scale these efforts up, and not end up duplicating things or having tons of different approaches where it’s hard to get economies of scale, or build knowledge or build capability.
[00:09:53] Nathan Wrigley: Now, given that the rate of change seems to be so incredibly fast. Give us an idea over those last two months, how much knowledge you’ve had to ingest. And I don’t necessarily mean knowledge, but how has it been, trying to keep up over those last couple of months?
Is it genuinely as fast moving as it appears from the outside to be? What you learned last month probably doesn’t apply this month. And so therefore staring into the future, and if I asked you the slightly banal question, what will we be doing with AI in two years time? Is there really any realistic chance that you can offer us an answer to that?
[00:10:27] Dan Walmsley: Well in terms of keeping up with it, there really is no way to keep up with everything. And I mean, there’s multiple different dimensions here, right? There’s the research dimension, what papers are coming out and how practical are those papers. And where are the outcomes of those papers showing up in libraries?
And then there’s like, where is it showing up in products? What are our competitors doing, or what products might we plug into our own stack? For example, we can use GPT4 to generate help responses, but we have to sort of, stand up maybe a vector database and some other infrastructure, various job management things.
There’s other third party services where you can point them at some public documentation and they figure all that stuff out for you, and just give you one endpoint that just chats with you. And it’s oh, well how much do we embrace this plus that? A lot of the day to day involves build versus buy versus don’t bother.
And it’s really hard because our team currently has not that many full-time developers on it, and we do want to move really fast and understand these technologies and do the judicious integration. I personally in my horrifyingly long career have done lots of integrations and they’re almost always bad news.
And I’m almost always fighting to do some minimal thing like in-house, rather than integrate. But it’s a constant. That’s really the battle. It’s like less so the awareness of what’s happening and more so wrestling with the idea of like, how do we incorporate this or not?
And people wondering if something’s strategic or aligned or whatever. And there’s all these different time horizons you’re looking at. Like, are you talking about today? In a week, in two weeks, in a month, in a year? Because they’re all different answers.
[00:11:57] Nathan Wrigley: Yeah, I feel like if I was to to you about AI two years ago, I genuinely think the conversation about what we would be doing in 2023, 2024, I honestly don’t think we could have got any kind of line of sight into what happened. Even maybe a year ago. Nobody would’ve thought that mainstream media, mainstream products, would be using AI. And like you said, falling over each other to have some kind of policy on AI. So I don’t quite know how conversation will go.
But it feels as if we’re in the infancy of this still, and it does feel if we are going propel ourselves through this at an exponentially faster rate. The thing that just popped into my head was that when humanity first came up with the motor car, it was, at least in the UK, you had to have somebody walking with a flag in front of the motor car. And most people probably looked at it and thought, that’s ridiculous. I could walk there just as quickly as I could get into that vehicle and be driven there because it’s going so slowly.
Give it 10 years, got a little bit faster. Give it another 10 years, it got faster and more beautiful and more efficient. But of course it then polluted the world, which brings us onto the inherent problems that we may have with AI. There’s a lot of concern about unexpected consequences. The fact that it hallucinates. The fact that it may give information out which is inaccurate. Given that this is your work, are you fairly sanguine that things built with AI are broadly speaking safe? Or are we just working out what the guardrails even are?
[00:13:25] Dan Walmsley: Well there’s a few pieces to that question, and I keep failing to address all the pieces of your question, so I’ll try to genuinely do it this time. But, the first piece is sort of like where we’re at in this AI timeline, and you talked about various analogies.
I think of this as the BBS era. If you’re in your forties, you know what a BBS is. If you’re not, it was when people used to connect to a single computer using a modem, and the modems were slow enough that you could see the text appearing on the screen. Sometimes slower than you could read. Certainly when I started using BBSs, it was slower than you could read, and even slower for images.
And obviously subsequent to that we got the internet, through various stages. And now, you look at a BBS and it’s unrecognizable. It’s like why would you ever look up information this way when you can look at the whole internet? I think we’re going to go through the same thing with AI.
There was another part to your questions which was the danger piece. There are alignment techniques that we use today on large language models and other kinds of models, that are fairly reliable at the scales at which those models operate, or at least useful.
And the worst things that those models can do are not yet super terrible. if you’ve got one plugin that talks to your bank and another plugin that can pick up the phone, then a rogue AI can hallucinate its way into destroying your life, no problem.
I sometimes talk about this with, we’ve experimented with building ChatGPT plugins for different products, including wordpress.com. And one of the hardest things is, you have to put user confirmation stuff everywhere because you simply can’t predict when the AI’s going to start invoking your API in backwards ways, and just deleting all your posts because it thought that’s what you wanted to do. Turning every post synopsis into the word red paper clip.
There’s a broader alignment thing that I think goes way beyond that. It goes way beyond these hallucinations. Because you know, I think people get caught up with, oh well it’s not that useful because of the size of the context window. It’s not that useful because it hallucinates. So it’s not that useful because it was last updated in September, 2021. As if all of those aren’t things that are going to change immediately, right?
Those are all solvable problems. We know we can make larger context windows. We know we can update it more often. We know we can inject additional information. We know that various alignment techniques can encourage it to reason more thoughtfully and activate pathways that have more expertise, and that will continue to be the case. And as the models get larger, those pathways with expertise will have more expertise. And so it’s obvious and predictable, those things.
So the really hard thing to predict is where does this interface with society? And you know, we touched briefly on jobs and other things. Or whether, obviously people talk about rogue states getting an unhinged intelligence to go do crazy scientific research for them, or invent a nuclear weapon or a chemical weapon.
Google Brain just invented protein folding. So get this, the Google Brain team, Google Deep Mind, they invented a protein folding system that can fold a protein in a few seconds, which is the equivalent of about at least four years of PhD time. And so in that single invention, they eradicated, I suppose you could say, or avoided over a million years of PhD time. By folding all those proteins instantly.
The thing I think we’re not ready for is that rate of progress. I call it Moore’s Law for everything. Where you have a self-reinforcing centralized paradigm, where you have AIs that, by their very progress, make it easier to build the next AIs.
And then at the same time you have this fanning out into different disciplines, where those newer AIs are also making it easier to make scientific progress. You could use, for example a score like Perplexity, feed in all of the papers in the world and find the most useful research questions to ask that have not been answered, by basically large scale language based statistics.
[00:17:03] Nathan Wrigley: I think this is the piece where my knowledge breaks down because my interaction with AI has largely been ChatGPT. Certainly the most recent versions of ChatGPT. Plus also the image creation tools. And, I’m amazed by how quickly I’ve become, unimpressed is the wrong word, but how quickly I just expect it to give me something akin to a human.
The first couple of times I used ChatGPT my entire endeavor was to see what it would produce, and be utterly, utterly flabbergasted by the fact that it could in any way give me something coherent back. And the same with the image creation tools, Stable Diffusion and a few others that I’ve tried. Typing in some kind of prompt, and then just jaw droppingly quickly, something half decent comes back. And you know you try a little bit harder and you tweak the input that you’re putting in and something slightly better comes back.
I’m kind of amazed by how quickly that became uninteresting and just normal. In the same way that when I was a child, I first got on the bike and suddenly I could ride a bike and wow, this was amazing. Two weeks later you have to basically pay me to get on the bike at that point, it’d lost its interest.
But I’m wondering if that interface, because it is replicating a human in many ways, you know, the ability to do art and the ability to give us answers, whether it’s hallucinating or not. I wonder if that’s something that we all think that’s the way the AI’s going to go. But the examples that you gave just then, like medical research and probably research in all sorts of scientific domains, if that’s something which just never quite gets out into the public.
So the fear that a lot of people have, and there are some parts of that that I share, is never counterbalanced by the, but listen we’ve just done thousands and thousands of hours of PhD equivalent work in a matter of moments. Look how fantastic this is. I don’t think that message gets out very often.
[00:18:56] Dan Walmsley: Well, you know, and without launching into a critique of the media, I think we can all recognize that dramatic headlines sell. And I’m sure if the headlines of these articles were slightly hard to predict whether AI will be good or bad, stay tuned. Then they wouldn’t sell so many newspapers.
You know, I don’t think anybody can actually, at a large scale, predict the outcome of the current AI revolution. That there are people who think that it will be a nothing burger. And there are people who think that it will more likely than not, result in the eradication of the human species. And there are people who think it’ll be cyborgs. And there are people who think it’ll be utopia. They’re all neither right nor wrong, yet.
I will say though, that people narrowly pushing, AI will take all the jobs line, definitely wrong in my opinion. Part of that we really alluded to this before the show, but part of that is, humans are really good at inventing new jobs. We added like 8 billion humans to the planet in the last a hundred or so years, and we gave them all jobs, no problem. We can invent new jobs like dog tickler and it’s fine.
People will just find ways to keep themselves busy. And if AIs come and take away a huge amount of jobs, particularly those jobs that are mostly typing and mostly repetitive, like similar things over and over again, then maybe those people get a chance to like move their bodies and stand up.
We forget how incredibly dysfunctional it is to sit there and type all day. If we can just take away all the typing. I have a gym membership because my body’s falling to pieces because I have to sit there and move my fingers and unblinkingly for like seven hours a day. It’s ridiculous. It’s torture. Can AI make that go away? That’d be amazing. What a revolution.
And so we sort of think about this in terms of jobs as if there’s some fixed number of jobs and the AI’s going to take them. And then there’s going to be no jobs to replace them. We don’t really think about it holistically, in this sense of if it’s doing all that work it’s producing huge economic value and unlocking human potential.
[00:20:48] Nathan Wrigley: One of the things that really has sort of crept up is the use of the word intelligence. So we’ve got AI, artificial intelligence. I’m not entirely sure that, at the moment, is really the right word to be deploying, because that is a fairly scary word.
You’ve seen films going back half a century or more where some kind of intelligent cyborg, something created by a human being at least, Frankenstein onwards, is able to outthink humans and therefore wreak havoc and so on and so forth. But my understanding is that the implementations that we are broadly using, ChatGPT and so on and so forth, are based on these large language models.
It would be interesting to get into the weeds of that if you’re willing. Can you explain how that technology works and why perhaps it’s more of a fluke that it gets anything right? Well, that’s not true. It’s not really intelligent in the sense that you or I would subscribe to a human, but it appears, it masquerades as intelligent.
[00:21:48] Dan Walmsley: Right. That’s very true. So, I’ll try to make this brief but accessible to people who might not have heard this explained before. There was a paper came out, I think it was around 2017, might have been earlier from Google, called Attention is All You Need. And that was the paper that described an architecture called transformers. Where you could feed in a sequence of text that they would turn into these tokens representing, not quite a character, not quite a word, but a numeric string of stuff representing the text.
And then it would be able to predict the next word with a pretty high degree of accuracy, based on paying selective attention to the previous words. So we all know that words like and, or, or not, aren’t always salient but then there’s other words that are sort of really important to the text.
It gets really good at picking up genre and tone and language. It’s important to note that ChatGPT was never trained to speak English. It was Hindi or anything else. It was just fed huge amounts of text, and they hide a piece of the text and say, can you guess what that is? And if you do that enough times with this selective attention model, then you end up with a system that is very good at continuing text where you left off.
Now this by itself is what they call a foundation model. It’s not that useful. The only thing that really does well is generate plausible sounding text. So if you start something that looks like a scientific paper, it will continue. If you start something that looks like a poem, it will continue.
So, once you have that foundation model, it’s not very useful for chat. It will go off the rails. Because it turns out, as soon as a transformer introduces one mistake into its output. Let’s just say it’s producing an output and it changes somebody’s name from Bob to Bill. It will continue to refer to them as Bill, even if it knows in its heart of hearts the correct answer is Bob, because all it’s trying to do is be as plausible as possible. Ah, I said Bill, I better stick with Bill. Or I said, up is down, I better continue with up is down.
I did about eight years of improv. It’s like an improviser in that respect. And in fact, that was one of the first things I used it for was generating scripts and improv things. Little musicals and stuff. Because it can take an absurd premise and run with it. So you give it an absurd premise like bogans in space, that’s a very Australian reference. It will generate the most plausible script it can for bogans in space. And that’s wonderful if what you’re doing is trying to create sort of a fantasy thing, but it’s less wonderful if you’re trying to do something grounded.
And so then they go through these various alignment processes where they feed it a huge amount of handwritten, curated, expert questions and answers on top of that whole internet that they fed it in the first place. And these are supposed to be illustrative of, I’ve got a question, I need a step by step answer that is clear and concise. And I also need it to refuse to tell me how to make a chemical weapon and other things like that.
So there’s some safety stuff there where you look at examples of people asking for malicious things. It’s crazy. I asked it to tell me a joke the other day, an Irishman, Englishman, American joke, right? And so ChatGPT refused to generate it. Because well, I can’t make a joke about people based on specific aspects of their race or whatever. Which is sort of like, fair enough in the general case but also weird in the context of me just wanting that joke for myself to see what it could do. That’s the kind of alignment stuff that they’ve put in.
And so finally what you get at the end of the day after a few more steps, is a model that has a little background thing where developers can align the model. Has all these different safety mechanisms. Has the ability to spell out instructions step by step,. Avoids as much as it can certain mistakes that would lead to it repeating itself or hallucinating too much. And has the ability to recognize now and use tools that accept JSON structured input as part of its cognition. That’s the latest level of alignment that they’ve introduced. And in the future there’ll be more and more as it gets bigger and more capable.
[00:25:31] Nathan Wrigley: So the fact that we’re on GPT4 at the moment, we’re recording this in June 2023. We’re on GPT4, and prior to that there was GPT3. And I think everybody can agree that each iteration is better. But the way that the technology is structured at the moment, will each version in the large language model, the token version that you described, the transformer model, will that simply get better at creating fewer and fewer mistakes?
Or are we approaching something which we could point to and say okay that now really is intelligent? In other words, are we heading towards a general intelligence? An AGI where we can now no longer disassociate it from being a human. It can come up with its own incentives, its own reasons to do things and then figure things out all by itself, based upon no human input whatsoever?
[00:26:18] Dan Walmsley: Yes. When I think of an AGI, I think of an autonomous AGI, right? Where it’s HAL 9000. I don’t really know when that will happen. And I don’t know if it’s a great idea necessarily. I think in between here and there, there’s like a huge amount of work to be done to bring this technology to life in ways that help people with their work.
It’s one thing to switch tabs and go to ChatGPT and type, write me a program that does x or y. It’s another thing to have GitHub Copilot living in the editor, which is an absolute game changer. And I suspect what’s coming next is AIs that work with your programmers and produce pull requests, or patches on pull requests, that fix linting or reduce complexity.
For example, I would pay at least $10,000 a month for an AI that comes in and reduces the complexity of the code that our teams write every single day. Finds methods that shouldn’t be there. Renames things to more align with each other. Move stuff between classes, and documents things publicly. Maybe pings developers if it’s not sure something’s useful anymore. Can you just imagine? Because not only is that cleaning up the code, it’s reducing the number of developers you want, it’s removing one of the most annoying things about being a developer.
So it’s making your job as a developer more pleasant. It’s not like it’s inventing new stuff, but it’s making it so much easier to invent new stuff because you’re working on super clean, minimal code that only does what you need it to do. And now just imagine if every company had that, how much progress we would see.
[00:27:46] Nathan Wrigley: Yeah, it’s interesting isn’t it because at the hub of that is, it almost feels like if you paid that $10,000, a large proportion of your team have to go away. Because probably a significant proportion of the team is people going in and cleaning things up and what have you.
[00:27:59] Dan Walmsley: Does it work that way though? Because let’s just imagine that somebody is on my team. Unless your company is losing money, right? Large amounts of money, and you’re desperately looking for some way to cut, right? If you have a programmer on your team and you can give them this tool and they become four times as productive. Then why would you want fewer programmers? Every programmer you add is four new programmers.
I don’t think this is going to result in people being fired en masse. People look around at Silicon Valley right now, there’s a lot of companies copying the Elon Musk strategy of, oh boy we just realized that we need to trim the fat. Over the long term, I don’t know if that necessarily means fewer programmers. Although I do think more people will get to be a programmer.
My dream is that every human being has their own open source stack that is completely proprietary to them. That is built and managed by an AI that is completely personal to them. Runs on a device that they own and control.
And so then you can simply describe how you want your life to be, and your personal software stack adapts and makes sure that I only see the information that is valuable and actionable to me. And because of this AIs role in my life, I’m able to get insights about what’s really working, and avoid distractions and nobody will ever be able to spam me again.
I actually literally am building an AI that scrapes the bajillion inscrutable emails from the school and plucks out the things that actually need to go in my calendar. It’s easy now, right? It’s 50 lines of code. And I can do the same thing for other digital parts of my life and just make that whole thing go away.
[00:29:29] Nathan Wrigley: Yeah, I think there’s three places to sit on this seesaw. There’s either I’m terrified by AI, or I’m really pro AI, or I think where I’m finding myself at the minute is more in the middle. There are parts of it that I can see which clearly have enormous utility, and are really going to put us on a rocket ship to Mars if you like.
There’s just no downside, but I think there is a part of me which does genuinely worry about the incentives. Whether or not it’s a great idea to automate all the things. Whether the landscape is just going to become flooded by noise, which actual humans can’t go through. So we then have to employ more AI to figure out how to get rid of the fake content. I’m not a hundred percent sold on it. I can certainly see there’s bits of it which have benefits.
[00:30:51] Dan Walmsley: Yeah, I’m going to start with the quote from Matt, learn AI deeply. We don’t really know where AI is going to go but we see a certain rate of progress. And it’s faster than Moore’s Law. And so if you use an imperfect AI tool today and you get familiar with it and fluent with it, let’s just say GitHub Copilot. You can be pretty sure that tool will accelerate in progress over time. Because it’s already an AI tool that’s like standing on the shoulders of this like industry. So it’s going to get faster, it’s going to get better.
The people who don’t embrace AI are going to continue on their linear or plateau trajectory. And so I feel like any human being alive today should probably start embracing some piece of AI in their life so that they can get a sense for how it’s shifting and changing and improving. So if it’s just a matter of using ChatGPT to like make plugin snippets, oh it’s good at this, it’s not good at that. Make it a habit. Then you’ll bear witness to what’s going on and you’ll know where to jump both feet into the stream and start leveraging this stuff more at scale.
In terms of where we’re going to see it show up in WordPress. I was on a panel recently and one of the things that I said was, a question worth asking is what content management system would an AI choose? If you’re an AI and you’ve been asked to create a website for someone and you haven’t been told what technology to use, would you use WordPress?
And the answer today is, probably. Because most of the public documentation for content management systems is WordPress documentation. So the AI has access to like 20 years of all this stuff. And that’s really, really powerful. It means it can reason about WordPress in a really impressive way.
It’s actually a great testimony to keeping WordPress roughly the same all of that time with minimal breaking changes. Because, you know, one of the things that I’ve noticed is there are lots of breaking changes between libraries in the Python ecosystem. And that means that ChatGPT very rarely writes working Python code for me. I have to modify it to use the latest API or whatever. It almost always produces working PHP WordPress code, because what works hasn’t changed, which is quite amazing.
[00:32:58] Nathan Wrigley: I mean, that is actually phenomenal to see that happen.
[00:33:01] Dan Walmsley: Yeah. Now we have to capitalize on that, but that’s a really great start. And you know what CMS would an AI choose, okay it’s one that it’s familiar with. And then the next level is, well it would be one where you can modify it and extend it easily. WordPress certainly checks that box to infinity, right? There’s all of these existing plugins and an AI can read the documentation of plugins and choose one for you or whatever it needs to do.
So the plugin mechanism is amazing because you can basically take a statement that someone makes about how they want their website to be different, and turn it into a function that runs a bunch of hooks.
It doesn’t have to go modifying the existing code of WordPress and forking it. It can just like inject the things that you ask it for, and correlate them back to the statements that you made. And in the future if it finds out that there was a better way to implement that request then it can implement it differently. Because it has the original things you asked for. So that’s one way I think I see AI helping with WordPress over time. Not that that’s a product that I’ve built I’m just sort of reasoning broadly about it.
[00:33:59] Nathan Wrigley: I think one of the areas that I really would like to see is the ability to leverage what’s just come around. I’m really excited about blocks and block patterns in particular. I’m quite a visual person, so I love to see images of what I’m about to get. And the idea of, I don’t know, I want to build a website for a local industry. A real estate agent, a lawyer or something like that. And the AI has some kind of interpretation of what that means. It probably has a little understanding of the geography of where I live and what kind of imagery might go into a website like that.
I live on the coast and there is some things which people always take pictures of and they often end up on websites for the places where I live. But also it understands typically what a lawyer is, you know? And it would understand that, okay, you probably need a page that has this on it, and a page that needs this on it, and probably a form and blah, blah, blah.
And then it would just throw at me, I don’t know 100, 200 designs, something like that, that I can look at. And because of the fact that it’s all built with blocks I could input that pattern, and then start to tweak things as I like it. I just love the idea of the choice that it might be able to give me, and short circuit, I mean me building 200 different designs, that’s going to take me weeks. This potentially could happen in the blink of an eye and I love that choice.
[00:35:13] Dan Walmsley: Yeah. Think about a few years ago, if you had like a site generator versus a site builder, right? So let’s just say I generated a site and we’ve all be familiar with site generators, you give it like, what kind of color scheme you want and what kind of industry you’re in and kind of thing.
And this has been possible for 10 or 20 years that you can generate a site. But the problem is, okay, now you’ve generated a site and then you make some content and you’re like, ah, I want to change that one decision. Well you can either regenerate it from scratch and blows away everything you’ve done. Or you can try and manually make the change, but you have no idea how to do that because you didn’t build it in the first place. And then you’re going to learn the whole system.
That sort of like magic trick of generating the site back in the day is the thing you can only do once. But in the copilot era, which I think Microsoft correctly identified this paradigm. You can jump in and out of automating the site creation experience as much as you want. And so the idea is, okay, I’m going to generate the content on this page. Jetpack AI block is actually really, really good at this. I’m not here to like boost our products too much. But it’s like a really good example.
You can generate a page and then you can just change the tone of voice. And it will go and take the same content and change the tone of voice, non-destructively, you know what I mean?
And so the AI is able to work with whatever changes you’ve already made and make some more. I think that that’s going to be the paradigm for a long time. And anybody building AI tools needs to be very careful about building the UI in such a way that it takes these hints from the human. And uses them to make the AI better over time. Better at getting the first guess right.
And any company that does that is going to have an AI flywheel. And any company that just generates content directly but doesn’t measure how the users respond to it or interact with it or change it over time is going to be stuck on a plateau, with no way to get to the next level.
[00:37:04] Nathan Wrigley: I really find the whole idea of that curious. Literally you could go to bed one night, wake up in the morning and the AI has decided that we’ve gathered lots of data and well you had a real blitz of users during the course of the night and it’s really shown us that no, they don’t like this bit, so we’ve changed it entirely on your behalf. So it’s like split testing but on steroids.
That seems like a really interesting idea. Obviously people will not wish to hand some aspects of that over but if you can prove that a WooCommerce sale, for example, this configuration of a checkout system seems to be 20 times more popular than this one. Okay, we’re going to get rid of that one. Now we’re going to start working our way through whether we can improve this one. All of that seems to be a bit of a no-brainer.
[00:37:47] Dan Walmsley: Yes. Building an awareness of when humans need to make discriminating decisions, and when you can make them on their behalf. And the product design aspects of what expectations do you set about what’s going to happen, or whether it’s reversible, or whether it requires confirmation or authentication or et cetera, et cetera. Taking a backup.
That’s all stuff that you don’t get for free with AI. That’s all the infrastructure of actually making it useful. And I will say the AI itself is dead simple to use, right? it’s conceptually unbelievably easy. 99.99% of the work is just like aligning the whole rest of the system around it so that you can make sure that customers have a good experience.
The normal stuff of building products, right? Setting expectations, all these different things. It feels different because watching a generative AI talk like a person is weird, but it’s not, it’s not work that requires you go do a deep learning course.
The thing that is transformative about this is it’s generality. These techniques have existed for years. We’ve always been able to classify, well, not always, for a long time been able to classify images for a long time been able to sort of grammatically parse out text or detect languages or sentiment or other things.
But they were all specialized models with vast data sets. And now you can fine tune it on 500 of your own examples and have it go answering entire support requests straight out of your knowledge base. And so it’s that generality that is really powerful.
[00:39:13] Nathan Wrigley: I’m curious to see what the UI for all of these different things are going to be in the future. In the sense that, you know, if you look at WordPress from when you began using it, it’s a very different animal. Although it hasn’t changed dramatically in the last five or six years. When you began using it, it was a different animal to the way it looks now.
And then these sort of page builder technologies came along and further democratized publishing and made things easy and it was a point click interface. I’m just curious to see how, what the pieces are that live inside WordPress. Whether it’s going to be text input. Whether we’re just going to start talking to our website and, you know, move it left a bit, a little bit more, make it red. Not that red, the other red.
I want a picture of a, I don’t know, a sausage over there, that kind of thing. How all this gets surfaced. We’re obviously in the era of trying to get everybody to use Gutenberg. Whether it fits into there or whether we need a brand new interface because the AI will just take care of everything. That bit is for me going to be really interesting.
[00:40:06] Dan Walmsley: Yeah. I’m really excited to see what happens with Gutenberg. I’m completely convinced Gutenberg will not go away. And actually AI makes Gutenberg look like a better and better decision versus the classic editor as AI comes into view.
[00:40:21] Nathan Wrigley: Can you develop on that? I think I know what you mean but I want to hear what you mean. Yeah.
[00:40:25] Dan Walmsley: So having things embedded as blocks with parameters provides a much more semantically rich interface than just a bunch of HTML. It’s similar as to how we see markdown used a lot more in AI than HTML as a formatting language, input, output. And why is that?
Well, it’s because the structure tells you something about the meaning of the document, right? This is a table, this is an image, this is a whatever. Obviously you get an HTML but more sophisticated than that, right? This allows the AI, so say you’ve got like a cover block with an image and a text. This allows the AI to have some confidence about how that’s going to appear when it shows up on a webpage.
As opposed to arbitrary HTML that may be pulling in CSS from various places and like all that kind of stuff. Gutenberg provides an incredible foundation for collaboration. And collaboration is key, right? If we’re talking about the copilot era here, I don’t think for a long, long time we’re ever going to have necessarily AIs. Like you’re not going to have a CMS come out that like, doesn’t have an editor, because it just has a chat interface. You tell the AI what to do and hope that it does the right thing.
Like that’s not going to be the case for a really, really long time, if ever. What you need is an editor where you can seamlessly collaborate with an AI. And if I was to take Matt’s words and bring them back into the conversation about learning AI deeply, I would love to see people in the community experimenting with UX concepts for collab.
We are in the collaboration phase. Now is the time to start bringing your ideas to the table about what it looks like to collaborate with an AI in Gutenberg and how revolutionary that could be.
[00:42:01] Nathan Wrigley: Are you open to those conversations? Is your team keen to hear from the community? And if that’s the case, where do we go to begin that conversation?
[00:42:08] Dan Walmsley: That’s all happening in the open source community. I’ve had a couple of conversations with Matias or others, but really at a high level. I think it’s the community that needs to help drive that. We’ve shown what’s possible with Jetpack AI. It’s like the first quickest, most sane thing we could build.
But in terms of the collaboration phase, my team is aligning the AI efforts of a large multinational corporation across many, many, many different modalities. Not just in the editor, but across image classification, and trust and safety, and all sorts of other things.
On a day-to-day basis I don’t have a huge amount of bandwidth for one thing like the Gutenberg editor but I really encourage the community to get involved and share ideas.
[00:42:53] Nathan Wrigley: Yeah. I’ll put links to the presentation that you were involved in, with, I know it was at least Anne McCarthy. I can’t remember who the other contributors were now but that was really fascinating. Interesting kind of first steps in, well, tell us what we want out of AI because we can see what it can do out in the wild with other things. You mentioned co-pilot and there’s obviously ChatGPT and all fun images that you can create with mangled fingers.
Interesting to find out what the community want from it. How it will look in two or three years time? And getting involved in that conversation could really impact the project right now.
[00:43:25] Dan Walmsley: I would also say, dark horse here, but I would love to see more people get involved in WordPress Playground. So for those don’t know, WordPress Playground they demoed it last year and I was actually in the room in New York for the WordCamp US there.
[00:43:38] Nathan Wrigley: That is some astonishing tech.
[00:43:41] Dan Walmsley: It is game changing. I mean, and it’s funny because it’s on the one hand you could look at it and be like, well, this is like a cute hack, but it’s you know, you would never run a website this way. But think about it, if you’re a person creating or modifying or wanting to come up with a new website. With no hosting, with no nothing, just sitting there like running a blob of JS in the browser.
And so if you go a step further, oh, generate an AI block that does X, Y, Z, right? And if you’re a developer that doesn’t already have WordPress or know WordPress, and you don’t have to pull down PHP, you don’t even have to write PHP. You have this like ephemeral WordPress in the browser and you can see what it’s capable of.
I think that could bring so many potential developers into the WordPress community. Who are able to see what’s possible, have this low barrier entry, who have zero dependencies and can provide plugins and blocks and other cool ideas into the WordPress community who might not have had a chance to contribute before.
[00:44:56] Nathan Wrigley: It’s amazing when you actually use it because you just assume that there’s a machine somewhere remotely that’s serving up that website and it just spun it up in a heartbeat. But of course it’s not. You can entirely unplug from the internet and there it is. It’s still working. And it took all of no seconds at all to get the whole thing going. It’s amazing.
[00:45:17] Dan Walmsley: Yeah, it really is.
[00:45:19] Nathan Wrigley: I will link to that as well. Yep.
[00:45:21] Dan Walmsley: I hope that becomes the way that a lot of people build stuff on WordPress actually. It is a playground. It’s really fun. It reminds me of when I was playing with the first version of WordPress. But it’s just accessible to vastly, vastly more people. You know, anyone with a web browser?
[00:45:35] Nathan Wrigley: Yeah, it’s kind of like having a blank piece of paper next to you, one of a thousand bits of paper that you can just scribble on and screw it up and throw it over your shoulder and, okay, that didn’t work. Let’s try again. We’ll just blank canvas, start again. And actually, I don’t know if you did see the address that Matt gave at WordCamp Europe. That was one of the other things he discussed. So you are very much in alignment.
[00:45:54] Dan Walmsley: It’s in my queue.
[00:45:56] Nathan Wrigley: Okay. Well, sadly, I mean, I could honestly talk about this with very little authority for hours and hours and hours. But we’ve probably used up our allotted time.
Dan, if anybody wants to reach out to you specifically, do you make yourself available in that way? And if so, where do we find you? Are you a Twitter fan? Or are you on, you know, you’re going to throw an email in our direction or a Slack channel? Let us know.
[00:46:17] Dan Walmsley: Well, you can reach me on Twitter. Twitter.com/danwalmsley. d a n w a l m s l e y. It’s a tricky one. And, that’s a start.
[00:46:28] Nathan Wrigley: Perfect. Well, thank you so much for chatting to us today about AI. I’m just sorry that I, uh, I can’t kind of keep up with the level of intelligence that’s probably required to make this conversation worth while, but I appreciate it.
[00:46:40] Dan Walmsley: I super appreciate being on the podcast. I’m really, really excited about the next couple of years. And especially for WordPress. I think we’ve got like a lot of strengths that if we leverage them, can put us in an amazing position to empower a lot of people to, you know, publish and to continue to democratize publishing.
On the podcast today we have Dan Walmsley.
Dan is a long-time user of WordPress, having started using it even before version one was released. With a passion for experimenting with different publishing platforms, Dan eventually discovered WordPress and has been using it ever since. Currently working at Automattic as a Code Wrangler, Dan is part of the Applied AI team. Although the team is relatively new, with only a few members, their mission is to coordinate and guide the various AI initiatives within the company. Recently, he has been focusing on automating internal workflows and communications, a particularly crucial aspect given the distributed work setup, which spans 70 countries and multiple time zones.
We start the conversation talking about Dan’s background. He’s recently decided that AI is a truly transformational technology and so has taken steps to learn the skills needed to understand and implement it.
Dan talks about how Large Language Models work, and how ChatGPT has driven awareness, and demand, for AI technologies in a way that was almost impossible to predict just a year ago. This has caused many companies to become deeply interested in AI and what it can do for their business workflows.
We get into whether the reality of AI can live up to the hype. Do we have enough understanding of AI to know what its impact will be on the workplace, or are we just in the middle of a media frenzy which will die down over time? Dan challenges the notion that AI will take many of our jobs, and emphasises the economic value that AI can bring.
We move on to explore the differences between site generators and site builders, and Dan introduces the concept of the ‘copilot era’ in which website creation can be somewhat automated. He highlights tools like Jetpack AI which can generate content and modify the tone of voice right inside of WordPress.
Dan stresses the importance of building AI tools with user interfaces that learn from human input in order to improve over time. He thinks that companies which measure user responses and interactions will gain a significant advantage in AI development, while those who fail to improve their AI content generation will be left behind.
Whether you’re new to AI or have been paying attention for a while, this podcast offers a fascinating insight into its impact on society, and how it can accelerate progress in fields like scientific research.
Google’s ‘Attention is all you need‘ paper
AI and the future of WordPress – Panel session