How AI will revolutionize accessibility

[Oliver] Okay, let’s get started. So thank you for joining today’s webinar on how A.I. will revolutionize accessibility.

I’m your host, Oliver Emberton, and you may well be asking, Who the heck am I to tell you anything about anything? Well, the most relevant part of my background that you may care about is I have two decades of commercial experience working with both A.I. and accessibility. In fact, I did my degree in A.I. back in 1997. Wrote my first neural net when I was 16, that would have been 1995. And my first A.I. when I was ten back in 1989. So for those of you who are paying attention, yes, I am old.

My mission is to make web accessibility ten times easier for everyone. And as the CEO of Silktide, I am generally known as a bit of a futurist person who likes to talk a lot.

So let’s talk a lot.

Today in our agenda, we’re going to be covering three main things. Firstly, we’re going to look at what is the big deal with A.I.? Why is everyone talking about this? Why is it suddenly all the rage? Secondly, we’re gonna look at how is A.I. changing accessibility? There’s a whole ton of amazing radical new technologies that have emerged in the last couple of, well, literally days in some cases, but certainly months.

That I’m going to share with you many examples of. And then lastly, we’re going to wrap up with what does the future hold? Where is this whole A.I. and accessibility thing going? Let’s peer into our crystal ball and get some solid answers. So to kick us off, part one, what is the big deal with A.I.? Well, you can’t help but notice, what has been happening recently is a radical explosion in hype and demand and interest in A.I. But how is this different from what you’re used to? A.I. today is nothing like the A.I. of two three years ago. It’s unlike most of the A.I. you still have in your day-to-day life.

I’m sure we’ve all had this delightful experience of being on a Zoom call when suddenly your Alexa device start playing the greatest hits of Britney Spears and you’re screaming there frantically. Alexa, stop. Alexa, stop! And in fact, as I’m saying it, my Alexa is, in fact, listening to me right now. I have to be careful. And on the one occasion when you actually do ask your A.I. to do something, it decides instead of playing, I don’t know, turning the lights on or something, it decides to eject your refrigerator into the sun.

So this is not the A.I. we are talking about. We are talking about something quite radical, new and transformative. The short word for it is Gen A.I. or generative A.I. It’s a new universal paradigm for machine intelligence. What do I mean by that? It means this is a single universal technology that ultimately has resulted in transformative effects across every aspect of A.I.

This is different. So if you go back 20 years ago or whatever, A.I. used to be multiple different specialist fields. We used to have computer vision, we used to have speech recognition. These are no longer separate. They are essentially all one version of the same tech. I’m going to give you a few quick examples to kind of calibrate you as to just how transformative this has been and the kinds of things that this will suggest will happen in the future.

So firstly, many of you, I’m sure, have used ChatGPT, but the example I’m giving here works just as well if you’re using something like Gemini or Claude, any large language model, which is fancy terms for a computer in space you can talk to and I’m going to give you this example here. I’ve taken a photo of a butterfly perched on a leaf surrounded by dewdrops, and I’m going to ask in plain natural English language to write an Instagram caption for this.

And A.I. has no trouble with this. This is actually a trivial problem now. Let’s take a look at the answer it comes back with. Nature’s delicate dance: where dew meets wings. And then an emoji of a butterfly and a dewdrop followed by #NatureBeauty, #ButterflyWhispers. Now you might look at that and say, Well, that’s not actually necessarily the best Instagram caption in the world.

It’s certainly not terrible. But think about what’s actually going on here. It understood the natural language. I didn’t operate a piece of computer software. I didn’t click buttons. I didn’t conform to a macro or a programing language. I just wrote what I wanted in natural language. The computer understood it. This is revolutionary.

About a year ago, we simply couldn’t do that.

Now it’s kind of boring. It also understood what an Instagram caption was, and it understood a tone that was appropriate and it did a pretty good job of like making something that seems succinct and a little bit poetic. It understood things like the contents of the image just by looking at the image and choosing to put in appropriate emojis.

And it did all of this without me telling it to do it. It just figured it out all by itself. That’s kind of crazy. Another example is a screenshot here of a check-out for a web page. A fairly standard thing. You could try this with a hand-drawn sketch, by the way. If you haven’t, that’s one of my favorite go to party tricks for A.I. is take a piece of paper, draw something on a piece of paper just with a pen, take a photo of it, give it to an A.I. and say, Hey, could you write the HTML for this page?

And it can. So sure enough, yeah, you’ll get— it’s not perfect, but you’ll get a reasonable approximate of that. So again, the A.I. in this case has some understanding of the HTML and CSS and potentially JavaScript and all these other technologies that ordinarily you might study for months or years, and it just knows how to apply them from a picture and a line in the text.

Kind of crazy. Now, for this interactive part of the exercise, I would like you to try and answer this riddle in your head. When ten plus four is two, what is nine plus six? I’m going to give you a few seconds to think about this. And warning. You’re going to be competing against the A.I. It’s going to try and answer the same question.

So if you want feel good about yourself, you want to try and answer this question quickly. Got 3 seconds. Okay. Well, if we ask an A.I. back in 2019 when ten plus four is two, what is nine plus six? It answers confidently the answer is nine plus six, which is kind of silly. It’s like asking a child what is one plus one?

And it says the answer is one plus one. It’s like, Yeah, I get it. You repeated the question back. Kind of stupid. So that was hopeless. That’s GPT-2, by the way. A little bit later, the slightly smarter A.I. answers confidently that the answer is 15, which is of course nine plus six but is still wrong. It doesn’t actually understand the riddle.

It’s kind of half understanding it. You can kind of see like a proto-stupid understanding. Little bit later. It’s using more words, but it’s still getting it wrong. The actual answer is ten plus four equals two because if it is added— sorry, if it is 10 o’clock and you add 4 hours, it becomes 2 o’clock. So if you apply the same logic to nine plus six, we get three.

Now that’s present day. That’s actually somewhere around March last year. Just to give you an idea of kind of the advancement in this sort of technology and you can run similar examples on a much more complex field. So you can take things like diagnoses of complex medical issues. And again, you will see an advancement moving from a kind of parroting stupidity to what we are we’re increasingly taking for granted is quite radical and surprisingly intelligent.

Now let’s go back again to February 2022. This is an example of generative A.I. learning to draw. So in this case you’re giving a text prompt and saying something like, I would like an image of a woman smoking a cigarette and you might get images like these. About a year later, ten months later, those images started to look like this. Same prompt, same technology, more advanced, more computing power, larger models, basically the same thing, suggesting how we’re scaling into better results.

And a little bit later, this is December last year. So this is after 23 months of the same technology being refined, we’re this level, which side by side is kind of absurd. I want you to imagine an industry where this is two years of R&D and this is at the time virtually a free product. Now, this is present day, This is image generation by AI so literally just text generated.

This is completely fictional imagery. Look at these. I find this always— I find this hard to believe. Doesn’t matter how many times I look at these examples. That image on the right sure as heck looks like a real photo to me, but that is 100% machine generated. That person does not exist. That Christmas tree does not exist and it is never a combination of elements that has ever existed and it has been invented from scratch.

And you sure could fool me about it. Again, I want to throw you back to 2022. It’s not that long ago. We also had people attempting at this time for the first time. This was Meta attempting to produce video from text. So this is— this was state of the art in November 2022. So you can type in a text prompt like a dog in a superhero costume, also bunny rabbits or an astronaut in space and you get video clips like this.

And that’s pretty impressive. But I’m going to show you the same kind of technology, but from three days ago. So one and a half years later. Now this video was generated by Sora. Sora is the new video Generation AI by OpenAI, and they wanted to make a video called Airhead. It’s done by some professional filmmakers, but they’re using A.I. to generate everything you see here from nothing more than text.

They are describing a scene that they want. And then the A.I., it’s generating HD video of this quality all by itself. Kind of absurd. So throwing all of that together on a casual timeline. This is not to scale, of course. We see from 2019 the very earliest forms of modern generative AI. Where you’re seeing GPT-2 doing a pretty poor job of answering text questions, kind of parroting them back, but showing some basic signs of intelligence that people like to laugh at, at the time.

And likewise, we saw generated imagery like the woman smoking you see in the top left here back in February 2022. But less than a year later, we had ChatGPT and we had the first versions of video generation. Only a few months after that, we had GPT-4 with a radical improvement in intelligence and again, only a year, less than a year later, we had high quality imagery.

And now of course, a few months after that, high quality video being generated from text. How often do you see progress in any given field that looks this absurdly rapid? So generative A.I. is a new universal paradigm for intelligence. The reason these changes are happening so fast is because we have actually unlocked something new and very special that allows us to get these kind of compounding advantages at an ever-accelerating rate.

That is why you’re seeing everything moving as fast as it is all at once. And to give you a peek of that, I want to give you something, just one example. This is the technology that we use for generating images from text, right? So you don’t need to understand this. Don’t worry about the diagram too much. But the point is, I would type in some text like an astronaut riding a horse and it would give me an image

out of it like an astronaut riding a horse, which is amazing. This technology clearly is designed to look at images and learn how to draw them from text and give you the output. But something crazy happens if you take that exact same technology, the same software, and you take something like this. This is an image of sound.

Okay, so you can take some music, you can take some voice clips, and you can turn it into a picture, right? And if you feed the same A.I., millions and billions of these examples with text describing the— describing like alt text, if you will, or captions for the audio images, the same A.I. learns how to make music. It learns how to make voice or how to make musical instruments.

You can find this, by the way. You can just Google online for examples. But the point is the same A.I. technology that learns how to paint also kind of spontaneously knows how to compose music, which is insane. The other thing that’s transformed or transformative about this technology is to move away from so-called supervised learning. So supervised learning is what you see here, where you take an A.I. and you feed it carefully labeled training examples.

So right now I’ve got an example here of some apples, some red apples, and I’m feeding it in also some images of bananas. And you would give this to an old style of A.I. and you give it like a million examples of an apple, a million examples of a banana. And sooner or later the A.I. can say, yes, this is an apple.

Yes, this is a banana. This is the kind of dumb A.I. we have suffered for decades that many of you have probably come to associate with the term A.I. But this is no longer the case. With modern AI. We are all about unsupervised learning. Unsupervised learning is where you don’t classify to your information at all. You just feed in information.

So you just say, Here are lots of pictures of things. There’s some apples in there. There’s bananas in there. There’s some peaches in there. It doesn’t matter. You just feed them in. And I’m not going to go into a technicality of it, but the way this works is it self organizes. The A.I. will literally figure out the patterns and classify things in a way that makes sense to itself.

And through that approach, with enough power, it ultimately learns to classify things independently. Now, to give you an idea of how absurd this is, how powerful this is, when you see ChatGPT and think of all the things GPT can do right? Bear in mind that ChatGPT was not trained. ChatGPT consumed the text content of much of the internet, and by reading, just by reading the internet, it was able to spontaneously develop the ability to speak pretty much every language and to translate between them to understand grammar and history and science and math and law and chemistry and poetry.

How to rhyme, how to perform mathematics, how to solve complex logic problems. This ability spontaneously emerged through unsupervised learning. Literally, it taught itself, much like people do. Now, it’s true there is a little bit of human reinforcement learning we add afterwards, which is like sprinkled on a top, and that’s basically used to make sure the A.I. behaves itself.

And doesn’t do things we find inappropriate or dangerous, but the underlying tech essentially behind all of this revolution is A.I. teaches itself how to do things. And because it’s so good at that, it can do utterly insane things incredibly, incredibly easily. And yes, if you’re paying attention, it is now correct and appropriate to be at least 20% terrified of this world because the things I’m telling you are radical and the implication of those things is enormous.

As we will explore in a moment. Now, you’ve probably heard with A.I. that A.I. is susceptible to inaccuracies, which are absolutely true. You saw some earlier. It’s also susceptible to hallucinations, which are a particularly dangerous version of this, where an A.I. is absolutely convinced it knows something, that it absolutely does not. A.I., of course, has also been known to copy others, depending on how it’s trained and how it’s designed and how it works.

It can sometimes echo things in a way that may disturb us. And of course, it is susceptible to bias. All of these things are 100% true. However, I would remind you that humans do all of this and more besides. And in fact, this is probably my biggest major point I want you to take away out of the first third of today’s talk, because if there’s one thing that I think is different between our present generation and the next generation, so the children that are born right now are going to be A.I. natives, right?

They’re never going to know a world without A.I. and they’re going to internalize it in a different way. Much like many of us grew up with the smartphone era or grew up with the Internet era, and you just— you think differently than the previous generation. Well for A.I. I posit that the thing that next generation will take away, that they will intuit that the rest of us will struggle with is that generative A.I. is closer to being human than being software.

I know this sounds absurd. Of course, generative A.I. is software, so mechanically it is software, but the paradigms we have in our head for how we think about it are completely out of date. I just want to expand on that slightly. So with software, software is logical, write software does exactly what you tell it to do and it does it in an exact and precise way.

It is objective, it has objective truth and it has a precise memory. So if you take, say, a piece of software and you type into it 100 names of people in a spreadsheet or something, and any of that information changes or is forgotten, for example, then that software is broken. That is not how software work. Software is expected to handle that perfectly.

However, people and generative A.I. is different. Humans, we’re fuzzy, right? We’re not particularly good at logic to be honest. We struggle a bit. Generative A.I. is very similar. It can do logic, but much like a human, it has to kind of work hard to do it well. The objective truths that we take for granted in software, like, you know, one plus one equals two or whatever, humans don’t really think like that.

It’s quite possible for us to hold multiple truths in our heads simultaneously, or to kind of have like a fuzzy relationship with concepts like truth, like, well, I think that is right, but I’m not sure. And of course, we have fallible memory. If you tell a human, here’s 100 names and you just read them out, you would expect— it wouldn’t surprise you if the human didn’t remember 100 names perfectly.

In fact, the human would probably hallucinate and misremember something to be convinced that it heard ones it didn’t and so on. This is the same underlying principle, I hate to say it, behind a lot of the modern day A.I. that we have, and it’s easy for us to look at this and go, This is flawed. And it is, but so are people.

And we find ways to make people do incredible things. I posit we are going to see a future where generative A.I. gets better at these things, will do technologies to help mitigate it to some extent, like we do with people. People have calendars, people have databases that use software, that use notes. It will be like that. But in order to have this incredible leap in intelligence, we also get some of these side effects.

So wrapping up the first section, what is the big deal with A.I.? They were there were three things. Firstly, it’s one paradigm, one universal paradigm that has been discovered in the last handful of years. Transformers, incidentally, is the underlying substrate, but you can look it up if you want. That one paradigm has unlocked an absurd rate of progress, a level of progress we simply have no prior arc for, it’s faster than anything we’ve ever seen technology do, and it doesn’t seem to be slowing down.

And I would say that taking away the point that this kind of technology is acting more human in software will be a useful mental hack for navigating many of the strengths and weaknesses that that technology has. We’re going to explore some of those in our next part of our talk, which is: How is A.I. changing accessibility? Well, how is A.I. changing accessibility?

I’m going to give us a handful of examples, some of these with video. So if you have audio available, it will definitely help. I will talk you through these one at a time. I’m going to start off with Be My Eyes. So Be My Eyes. Is a remarkable smartphone app. This actually has existed for a number of years. The basic idea is that you use your phone as a pair of eyes that you can talk to.

So originally, the way this started is you would take your phone and let’s say you were partially sighted or blind and you could use your phone to look around, say a room, and you could talk to someone and there would be a volunteer on the end of a video call, essentially, and they would look at what you’re seeing and you could talk to them about it.

So you could say, what color is this shirt, for example, and they could help you out. Now, last year, Open AI brought out a Vision API, and this technology allows for a machine to perform the same task as a human. So essentially it can be your eyes and you can talk to it in natural language. I’m going to show you a quick video demonstration.

The new AI tool in Be My Eyes is powered through GPT-4. Like an exchange with an actual person, Brian can take a picture with his phone and follow up with a question.

[Brian] What’s in my refrigerator?

[VO] Within seconds he gets an answer from the virtual assistant, which Brian has programed to speak at a rapid pace. We slowed the audio down for clarity.

[App voice] I can see that there were several items in the refrigerator, such as a carton of milk, almond milk, soda cans, a water filter pitcher, hummus, and various packaged foods.

[Interviewer] So what makes it different for you versus having a sighted volunteer?

[Brian] Privacy, you know, a sighted volunteer, there are certain things you don’t necessarily want them to help you with.

[Oliver] So that’s remarkable technology.

And this exists today. The demonstration you saw, I think was from somewhere around April last year. And this technology, again, is the worst it will ever be right now. It’s only going to get better. But the idea that we can actually have a conversation with an image is radical, transformative, and the accessibility benefits are profound. As Brian there stated, he was able to do this with a, you know, a human volunteer.

But the problem with that first is privacy. There are things you don’t necessarily want to involve in your life. And also availability. There’s a distinct limit to the number of people that can help you out at all times of the day. And A.I. like this is going to be on every phone, it’s going to be standard and it’s going to be incredible for millions of people.

Now, moving on. Microsoft Copilot. So Copilot is the all encompassing brand that Microsoft has adopted for their A.I. insistence, and they are putting Copilot in every piece of Microsoft software you’ve ever heard of. So, for example, Microsoft Windows, Microsoft Office, which I’m going to show you in a moment, and things like GitHub, all of that technology, all has some form of Copilot either in it now or coming out this year.

So far as I know, everything they’re doing. So let me give you one example of how that could be transformative for accessibility in the workplace. So in this example, I’m going to assume that I have a visual impairment. For whatever reason, my eyesight is either not that good or I literally can’t see. It doesn’t really matter for the purposes of this example.

But let’s say I want to give a presentation like this one. So I have great ideas. I know how to deliver a presentation, I know how to write a presentation, but I’m not going to use software like PowerPoint to do that because the software is very visually orientated, right? I have to think in terms of design. I have to like work on the layout of the page and that’s something that for many of us we can’t actually do.

But I still have this amazing communication I want to deliver so what I’m going to do is I’m going to use Microsoft Word and I’m just going to write in a word processor, and I could do that with a screen reader, I can do that with all manner of assistive tech. I can write a great presentation.

I could even narrate it. But then I’m going to ask Copilot to turn my written notes, my Word document, to turn it into a presentation. And sure enough, this is something you can now do. You can ask, Here’s my written notes, and it will take your notes or your, if you want, a transcript, whatever you’ve written, and it will interpret it, it doesn’t just copy and paste it.

It’s not some root set of steps like regular software. It literally understands what you wrote and it turns it into a surprisingly high quality presentation. So it might repurpose your words, it might rewrite them, it might interpret them in different ways, and it’s going to add design elements. It’s going to add appropriate clip art, it’s going to choose a theme for you, etc., etc.

And then this you can see in this example, we then subsequently asked the A.I. if it could add a slide about the cost benefits of sustainable materials. You can just have a conversation with your spreadsheet. Or more accurately with your spreadsheet software. So sorry, not spreadsheet. Goodness. Your slide, your presentation slides software. And in this example we asked to make the slide more visual, to move the text to the speaker notes and to add animations to the slides.

And sure enough, that’s something that Copilot can just do for you. So every single piece of productivity software in your life is going to end up looking like this. The bigger software like Office and things like Photoshop and so on, either already has this functionality or is going to have it in coming months, right? And gradually every single piece of software you use will have some chat in space like this.

The opportunity to improve access for people here and to make people more productive and capable across a range of modalities is tremendous. Now my next example, I want to show you something quite new. You probably haven’t seen this. This is called Open Interpreter 01, and I apologize for the terrible name. I don’t know what they’re thinking, but this is a quite impressive technology demo and I’m going to show you a short video clip.

Just to preface this, the idea of this technology is that it gives you remote control over your computer through voice. So this is the kind of Star Trek or Ironman technology where you can just talk to a computer in completely natural language and ask it to do surprisingly complicated things. And it just gets what you want and it just does it. And I don’t mean things like, can you click on button with the label X? I mean, hey, can you look at my calendar and whatever? I’ll show you an example. You’ll see what I mean.

[Video, A.I. voice] The next sunny day is on March 9th.

[Presenter] Great. Am I doing anything on that day?

[A.I.] Let me check your calendar. You have no events scheduled for March 9th.

[Presenter] So because the 01 is operating my personal computer at home, I didn’t need to set up my calendar. You can just use it. Nice. Can you see if there are any concerts in Seattle on that day?

[A.I.] Yes. You have several options. Option one, Bad Bunny will be performing at the Climate Pledge Arena.

[Presenter] That’s perfect. Can you add that to my calendar and then copy and paste a link to the ticket into the calendar event? And then actually, if you could text Paige the link and say something like, Hey, do you want to go to this?

[A.I.] Done. The concert is on your calendar and Paige has been texted.

[Presenter] Awesome. Thanks.

[Oliver] Now just consider for a moment what you saw there. The A.I. in this case does not have access like an API or whatever to talk to the software. Like the calendar is not being operated in any other way other than as how you would operate a calendar.

So literally it’s using virtual eyes to look at the screen and it’s using a virtual mouse pointer to click on parts of the screen. It’s using a web browser like you would to search the Internet and find out what concerts are on a day, and so on. Right? Absolutely insane technology. Again, this is all in beta, but technology like this exists today, and it’s probably only going to be months to a single digit year before this is going to become a standard means of operating every computing device that we all possess.

Now I’m going to move on to robots. So this is an example of a Tesla Optimus Robot. I’m not actually going to show you their tech demo because, whilst it is impressive, I’ve got a better one for you. Now. Robots have gone through an absolute transformation in the last 12 months in particular, so people don’t realize this, that the same underlying advances in AI that you saw earlier with things like image generation or large language models, the same underlying changes are empowering us to make robots do things that frankly seem like they’re from science fiction.

So you’re going to see in a moment you’re going to see a robot that can respond to voice commands, much like you saw the computer responding to voice commands, completely natural language. I want you to pay attention to the fact this robot is literally going to say Um or Err, and that is surprisingly intentional, as well as the fluidity of the movements, which again is brought about by the same technical advances.

Also pay attention to the lag. You’ll notice what I mean. So the one problem we haven’t solved yet is making computers respond fast enough.

[Video, presenter] Hey, Figure One, what do you see right now?

[Figure One] I see a red apple on a plate in the center of the table, a drying rack with cups and a plate, and you standing nearby with your hand on the table.

[Presenter] Great. Can I have something to eat?

[Figure One] Sure thing.

[Presenter] Great. Can you explain why you did what you just did while you pick up this trash?

[Figure One] On it. So I gave you the apple because it’s the only edible item I could provide you with from the table.

[Presenter] Great. So based on the scene right now, where do you think the dishes in front of you go next?

[Figure One] The dishes on the table like that plate and cup are likely to go into the drying rack next.

[Presenter] Great. Can you put them there?

[Figure One] Of course.

[Oliver] There you go. That’s the real thing that exists today, and 12 months ago, even experts in this field would say that most of that was practically impossible. The rate of change for observing in robotics is genuinely transformative.

You really should Google some more videos if you want to see them, but it increasingly looks not only possible, but inevitable that we will see home robots that could literally be personal care workers or caretakers for us within the next five, seven years. And the improvements that that could offer in quality of living for millions of people, again, are just simply almost unimaginable.

Right. And I’m going to change tact. I’m going to talk about something a little bit more tedious, but I promise you, as important. PDFs, PDF documents, I am not a fan. Anyone who works in accessibility probably is not a fan. PDFs are notoriously one of the most common and yet terribly inaccessible technologies on the planet. And in fact we actually are such fans of them here at Silktide that we made these stickers: Make love, not PDFs. We actually gave these out at the CSUN conference last week.

And yeah, so as much as I love to rail against PDFs, I preface that by saying I’m now going to say some nice things about PDFs because the first thing that has been introduced to PDFs, and this is in beta, but again, you know, this technology will be standard in weeks or months, is conversational PDFs.

This is the ability to actually talk to a PDF and actually have a conversation about the content of a given PDF. Now this is the sort of thing that single handedly is actually quite easy to do with modern A.I. It transforms the PDF from being a terrible clunky, horrible experience, particularly for accessibility into something that’s kind of profound and useful. Like if someone can send me an academic paper and instead of reading it, I can just say, Could you summarize that for me?

Or, What does this bit mean, I’m confused? Suddenly it’s gone from being awful to brilliant, and this is one of the worst technologies I know of for accessibility on the planet. This actually brings up a broader point with conversation and A.I., it’s actually one of the easiest things to do. You’ve seen examples of robots controlling a computer. A conversational A.I. is pretty easy to build into anything, and so I’m actually pretty confident in predicting that before the decade is out, you will be able to have a meaningful conversation with your toaster oven. It may sound absurd, and it probably is, but you’re going to find the technology that allows us to do something like GPT-4 and better is just going to be miniaturized.

It’s going to be made portable. It’s going to be like a Wi-Fi chip or a Bluetooth chip, something that right now you could fit in a wristwatch or indeed in a refrigerator. And sure enough, people are going to make home appliances that support natural language and chat. So I’m not saying every toaster oven will have it, but by 2030, I’ll be very surprised if you cannot buy a toaster oven that you can have a legit conversation with.

Anyway, you heard it here first. With PDFs. One of the biggest problems of PDFs is how much work you currently need to put in to make them accessible, because the format is basically not designed for humans, it’s not designed for interpretation by anything. It’s just designed to be a visual format that shows you what a document looks like. And so the traditional way to fix this is to tag a PDF manually, to do lots and lots of complex and frustrating technical work to fix a PDF and make it accessible.

But starting last year there’s been some quite radical improvements in this and making A.I. do this automatically. And whilst I myself would have argued that A.I. could not do this well enough, I have seen enough at this point to be convinced that it is now inevitable. There will be a point where PDFs could be remediated automatically to a human level or even potentially better than an average human level, by A.I. This is one of the most transformative things because nothing needs to change in the broader marketplace for people to take advantage of this. They just suddenly start using PDFs that support reflow, they start using PDFs that support voice and copy and paste, and they just work and they’re just accessible and they’re great. So that’s basically now inevitable.

And then lastly, I want to show you brain computer interfaces. So this if you don’t recognize logo, this is Neuralink. But there are other companies in this space. This company is specializing in making a device which you see here that is implanted into the human brain. Yes, that’s the real thing. And the reason for doing this is because it allows incredible range of high bandwidth access between your mind and a computer that can afford you a user interface that for many people who are you know, well you’ll see an example in a moment with a quadriplegic, completely life altering.

So I’m going to show you a quick one minute video demonstrating this. This is the first patient for this technology, and I believe this video is about a week old.

[Video, Nolan] My name’s Nolan Harbaugh, 29 years old. About eight years ago, I was in kind of a freak diving accident and dislocated my C4-C5. So I’m a complete quadriplegic, so I’m paralyzed from below the shoulders.

I love playing chess. And so this is one of the things that you all have enabled me to do, something that I wasn’t able to really do much the last few years, especially not like this. I had to use like a mouth stick and stuff, but now it’s all, it’s all being done with my brain. If y’all can see the cursor moving around the screen, that’s that’s all me, y’all.

It’s pretty cool, huh?

[Neuralink engineer] Actually can you pause the song for the audio coming through?

[Nolan] Yeah, absolutely.

[Neuralink engineer] And that was also done with your brain?

[Nolan] Yep, it’s all brain power.

[Oliver] Now consider how utterly transformative that is, right? What a life that is opening up. So my favorite part of this story is, so this is the first patient using that technology in the world, by the way, is that after they gave it to him, his first time playing with it, was to pull an all nighter, playing the video game Civilization for six straight hours.

And he went from using a mouse stick. So for those of you who don’t know, that is like a stick that you hold between your teeth in your mouth that you use to push buttons on a computer, He went from that to being able to think what he wants a computer to do and it doing it, and essentially begin to operate a computer at a level that’s comparable to a regular mouse and keyboard. That is incredibly transformative.

I’d also add that this is just scratching the surface of what this technology can do. If you haven’t seen what A.I. can do here, this is incredible, because it’s A.I. that’s enabling this, by the way. Without A.I., we just have something we plug into a brain. But with A.I., we can actually start to deconstruct your thoughts. This is an example of images we are presenting to a person.

So at the top, these are images we would show someone, they would look at these images, and then we give an A.I. the task of trying to read their brain. So the A.I. can’t see what they see. They can only see their brain. And the idea is, can we train A.I. to actually reconstruct what is going on in your head?

It’s not perfect, but you can certainly see the relationship between the two. And this is early days. This kind of technology is, again, the worst form it will ever be right now. So imagine it in five years and ten years. Imagine being able to literally think what you want a computer to draw or to write. It’s no longer science fiction.

So quickly, to recap our second section, how is A.I. changing accessibility? Well, there’s a few points. Firstly, natural modalities. For the longest time, computers have been operated through things that were designed for computers. So a switch, a button, something like that. A keyboard, for example. Very easy for a computer to work with that kind of technology because it’s just on and off.

It’s just binary. We gradually move through things like mouse pointers and touch, but what we’re seeing now in this kind of third wave of A.I. enabled technology is the ability for computers to understand complex modalities like speech or even thoughts. Through A.I. Completely impossible until now. Then we’ve got chat-to-everything. We saw the ability to talk to a PDF document.

We saw the ability to talk to a computer or to talk to a robot. This is actually one of the easiest things to do. It’s actually surprisingly easy to add to almost any technology. You’re going to see it everywhere. Chatting to a computer to interact with an app is going to become as standard as using copy and paste. Then we saw robotics.

The advance in robotics is easy to miss. It’s not getting as much press right now as the rest of A.I. but I assure you it’s worth paying attention to. You’re going to see transformations in the next single digit years that are going to blow everything out the water that you expect. And I honestly think we are going to have robots in our homes before 2030 and they are going to be incredible.

And then lastly, I would say that if you look at all of this technology, it’s not really assistive tech in the traditional sense. We’re seeing assistive technology just becoming technology. So for the longest time, there used to be separate technology for accessibility, right? Yeah, that would be screen readers, for example, would be technology. But what we’re starting to see now is companies like the big tech companies, the Microsofts, the Googles, the Apples, creating technology for the wider world that just happens to have a profound impact on accessibility.

Now, this is actually a really good thing. It may not be the motivation we wanted, but it gets us much higher quality technology much sooner because the audience is that much larger. And I hate to say it, the capitalist incentives for those companies is much stronger. If you can make a voice controlled interface that allows someone to be driving in their car and to manage every aspect of their life, that affects everyone.

But the benefits to assistive tech are going to be most keenly felt. So moving on to our third and final section, what does the future hold for AI and accessibility? Well, the question I get asked a lot, and it’s kind of one of those dark, hidden questions below the surface is, with all of these amazing and profound and slightly terrifying changes to A.I., could A.I., in fact, make accessibility redundant? I mean, is it just not now kind of inevitable that at some point we don’t even need to worry about making, say, websites or apps or hardware accessible? Because eventually we’ll all be living in like A.I. powered mech suits and, you know, all of our senses will be plugged into our brain or whatever.

Right. Is there not a point where this just goes away? And. Well, let’s explore that rather challenging question with a historical anecdote. So I want to go back a little bit. Hopefully, there… Everyone on this call is old enough to relate to this example. The example I’m going to go with is the smartphone era. So consider how in 2007 the first iPhone was released.

I don’t know if any of you remember that, but I know I was there queuing for the very first iPhone myself. I got a release day model of the iPhone, which was amazing, and I loved it for the six months that it lasted before they brought out a new one. But such is the price of early adoption.

Now, I want you to consider for a moment how long do you think it took after the first iPhone came out, before half of all US adults were using a smartphone? So how long do you think in years that took, right. So I don’t know if you remember, 2007, the first iPhone came out and at the time it was kind of crazy.

Like a lot of people didn’t think that was going to be a hit. You know, a lot a lot of skepticism. But still, we now know, of course, it was. How many years do you think it took before half of US adults were using a smartphone of some kind? The correct answer is nine years. Most people, in my experience, estimate something much, much quicker.

They seem to remember as if it happened in like the blink of an eye, like two years, or something, it definitely didn’t, it took nine years for half of the US to be using a smartphone. And to continue that. What do you think the current percentage of smartphone adoption is in the United States for adults? So what percentage of US adults do you think currently use a smartphone in 2024?

The correct answer is 90%, which means 10% of us adults are not using a smartphone. Even today. It’s often hard to believe. But we forget how, despite the technology being seemingly transformative as the smartphone is and being ubiquitous, it doesn’t prevent people from not adopting it. And that’s going to be very important to where we’re headed in the moment.

So two years later, for those of you who follow accessibility standards, WCAG 2.1 launched and this was the first accessibility standard specifically covering mobile. So it, as one example, introduced orientation as a requirement. That’s the requirement that your mobile device or your tablet potentially cannot require that the user rotates it for them to use a website or an app unless there’s absolutely no other way around that.

The reason that’s a requirement is because for many people, their smartphone, for example, may be mounted to a wheelchair. It may not be something that they have the physical ability to rotate. So you can’t require that they do. Now that’s just one example. But the idea is we took the accessibility standards of the time and we added a handful of new standards specifically to mobile.

And then two years after that, we saw the first introduction of this becoming a legal requirement. As far as to the best of my knowledge, the EU public sector was required to adopt that around 2020. So what does this mean? It means that it took 13 years between a new technology being introduced for it to be adopted by roughly half of US adults, which of course is not half of the world, but relatively rapidly moving marketplace to adopt, and then for new standards to be introduced in accessibility and then for those standards to become a legal requirement.

And they’re far from universal, right, this is just a little bit of the EU, this is just a little bit of the overall marketplace of people. So what happened? We saw small extensions to existing guidelines. We took all of accessibility that was already there. So all of the technology for desktop and all the other things that we’re used to, and we took the old problems, they still remained, but we just added new ones.

So while we have mobile… Mobile devices, for those of you who don’t know, are among the most accessible technology that exists. They are incredible. Like a screen reader on the mobile phone, I will take any day over a screen reader on a regular computer. It’s just better designed. It’s more modern tech. But it didn’t fix the problems with desktop computers because we’re still using desktop computers.

To give you an illustration of this, this is a comic from XKCD. The first panel is: Situation. There are 14 competing standards. And then a couple of engineers get together and go: 14? 14 standards is ridiculous! We need to develop one universal standard that covers everyone’s use case. It’s now in agreement. Yes, this is right. Absolutely. Soon, there are 15 competing standards.

This pattern is as old time with technology. Everyone is always inventing a new technology to solve all of the problems that have come before it. And unfortunately, when they do what they do, they don’t actually lose the old tech. They just add to it. Now, keep this in mind when we think about A.I., how fast could A.I. change everything?

Well, to do this, I’m going to invent a fantasy technology, right? So this is— I’m just going to reach for the stars with this one. So we’re going to assume we have the best bits of the tech I already showed you. So we’re going to have… I’m going to call them sentient voice assistants. They’re not actually sentient. That’s just cool branding.

It’s the sort of thing that we probably would actually see in the real world because it sounds cool. And I know Apple, you love a cool sounding name, so, call me. If you’ve got the ability to talk to your computer in natural language, to take anything you could describe and for your computer to do it. So this, this would potentially make a ton of normal assistive technology redundant.

You would need screen readers, maybe, because this is better. You wouldn’t need to worry about learning complex commands and so on because this is better. Let’s just assume this technology comes out and we’re going to assume this technology comes out today. Just pretend it’s released right now. Okay? And for the sake of argument, we’re going to pretend it exists on every computing device all at once.

This is science fiction. So we’re just going to run with the thought experiment. Now, it’s going to take in this case, I’m going to be very optimistic. I’m going to say it takes four years for half of US adults to be using this radical new technology because it’s so good that they simply can’t help themselves and think about how unlikely that is.

Many of you already have a Siri or a Google Assistant or an Alexa or whatever in your lives, and you don’t use them. In fact most people that use them, right? They’re not as good as this, of course, but would making them much better suddenly convince half of Americans to adopt them? Probably not. It takes time. Anyway, give it two years.

And let’s assume accessibility standards get expanded and they now cover this technology. Maybe we have new guidelines for things like bias, for example, seems quite likely. And then a couple of years later, that standard becomes law in some countries. So even in this example, it takes eight years to go from the most radical technology I can humanly invent, deployed as quickly as I could possibly speculate and it still takes about a decade.

And that’s a contrived example. But even in this contrived example, consider what actually happened. What were the consequences? Well, once again, we saw small extensions to existing guidelines. The accessibility requirements that we take for granted today still remain. We haven’t lost desktop computers. We haven’t lost, I don’t know, your car radio or whatever. Right. This technology still exists.

We’ve just added new technology. The old problems remain, but we now have new ones. Now I’m going to give you a framework I came up with for forecasting A.I. disruption. Now this is general, so that you can apply this to anything, not just to accessibility, but we are going to look at it for accessibility. And this is my go-to for kind of predicting, peering into a crystal ball, what A.I. will actually do to the real world.

So I want you to think of this in terms of accelerators and brakes. So accelerators to adoption and brakes for adoption, because the challenge here is never really the technology. The technology you’ve already seen is available or becoming available at an incredibly rapid pace. But the challenge is how quickly does it get adopted? Because unless people are using it and using it everywhere, it doesn’t really matter.

So what are these? Well, the first one is distribution. So consider if you have something that could be distributed through existing means. Everyone gets it. It’s adopted quickly. For example, if Google decided change the Google search engine to work with A.I. right now, then it’s distributed to the whole world instantly. So that’s super, super quick. And so you’re going to see things like existing web apps and software just change, right?

Office is just going to gain A.I. and so everyone’s got it, as long as they’ve got the latest Office. But on the other hand, you’ll sometimes see A.I. and we’ve seen this with assisted tech where the distribution is very novel. So in an extreme example, if you’ve got a human brain interface, if you have something you have to drill a hole into your skull for, that’s going to take lot longer to distribute.

And it’s worth thinking in A.I. where a given disruption lives on this continuum. Next, you have hype versus trust. Some industries like, say, marketing or tech, tend to be very hypey and they’re very rapid early adopters of new tech. Other industries like, say, finance, health care or accessibility tend to operate in a higher trust space. They are skeptical, and potentially rightfully so, the new technology has flaws that can really affect people’s lives in a negative way. Those robots are amazing, but if any of them were to harm someone, the physical harm in fact could be quite profound or devastating. So this creates a very reasonable break to its adoption.

Next, we have value versus cost. The value of a given A.I. disruption is often economical. So you might look at, for example, if Amazon was looking at replacing warehouse workers with robots, there’s clear value for them in doing so. But for many of the technologies we will see an accessibility, the cost could be prohibitive. We might have new voice assistants, but they probably require new phones. We might have robots, but they certainly need to be bought and distributed.

These things are going to slow down adoption. And lastly, we have a change in behavior. Old behavior is where I don’t need to do anything different. I’m used to using Google by typing what I want in a box. If you can keep Google working like that, and add A.I., then it’s great and I don’t need to change anything.

I will adopt your technology very rapidly. But if there’s a new technology where I need to learn to work in a different way, depending on how old I am, there’s a very good chance I’ll never adopt it regardless, you know, voice assistance might be the best thing since sliced bread, but there will be a ton of people who are just like, No, I’ve used my screen reader for 20, 30, 40 years.

I don’t need a new way of doing things. I’m going to do it the way I’m used to. My life works like this and who are we to tell them differently? So again, that’s a very strong break to any kind of adoption. I would say generally that humans resist geometric change, which is a fancy way of saying things that curve up to the right too quickly.

So this is a chart of GDP per capita or essentially the economic output of, in this case, the US. Right. And this chart goes from 1950 up to the present day. And I want you to take a look at this chart and think this chart covers eras such as the invention of the modern microprocessor. It covers the rise of personal computing, it covers the invention of the Internet and smartphones, incredible technologies that disrupted every aspect of our lives.

And yet look at that chart and tell me on the chart where you see, for example, the Internet era, where you see the personal computing era, where is the profound up and to the right explosion of growth and economic output of whatever that we got from those changes? There really isn’t one. It more or less looks like a curve that just carried on regardless, going up and to the right in a fairly consistent way.

And those technologies undoubtedly facilitated that. But they didn’t result in this explosive transformation that we might have imagined. And that’s because humans resist that kind of change. We are our own form of natural brake on things adapting that quickly. And I think this often gets overlooked with A.I. So in conclusion, I’m going to say this doesn’t actually mean that nothing changes because the technologies we saw earlier on, things like robots and human voice interfaces, human brain interfaces, these are real.

These are going to have profound life altering effects for millions of people. Just consider that the life that this technology could open up for so many around the world. It’s incredible. But it’s going to be mediated by the speed at which humans can adopt it. In conclusion, everything is going to get better, profoundly better, but it’s also going to get more complicated.

The new technology does not replace the old. It takes a long time. It gets added. So if anything, accessibility is going to get more complicated, more involved, and there’s going to be new challenges we have to face. Change will be fast. And yet, it will take too long all at once. We’re going to be both blown away by how fast these things are moving and you’ll look back and be like, Do you remember a world before we had robots?

But it’s also going to take way too long and people who might benefit from it are going to be waiting longer perhaps than they should because of these reasons. And both of those things could be true at the same time. So I’m going to wrap up. At the end, I was like, What do I say? What do I end my presentation with?

So I asked ChatGPT to help me out because I don’t know what to say. So ChatGPT very helpfully wrote the next line. A.I. will be exceptionally impactful for accessibility. The opportunity to do good is enormous, but the barriers to success are equally massive. And it also gave me, very helpfully, a nice picture to decorate that slide. So good job, A.I.

Thanks a lot. That’s it from me. That’s my talk today. I hope you found that useful. I’m going to open us up to questions now. So I’m joined by my colleague, Jess. Jessica, if you are around and you can unmute yourself, if we have any questions, I’m happy to take them for the next five or so minutes.

[Jess] We do. I’m going to start with the easiest one, which is, just to clarify, is conversational AI available with PDFs right now?

[Oliver] I believe it’s beta tech. I’d have to check. I haven’t used it myself. I think if you Google for— I want to say— it’s on Adobe’s official website, something like Adobe PDF chat or something. I’m sure you will find it.

But a lot of this technology, like for example Office is early release or beta, mostly I think because of the quite understandable legal ramifications and also the cost implications of making that technology widely available. So there’s a big incentive for these companies to show it now, to assure, you know, their users and their shareholders, Hey, we’ve got it coming, but maybe not to distribute it to a billion people before they realize what the consequences might be, because let’s say a PDF gets interpreted by chat and says something that’s wrong and someone uses that in a business decision, they don’t want to be liable.

So there’s a lot of stuff like that they have to navigate. So the best of my knowledge, it’s not currently available to the general public.

[Jess] Generative A.I. is highly iterative. How much memory is needed to run it? How much does it cost to run it? How much energy can we hypothesize will be dedicated to Gen A.I. on a daily basis?

[Oliver] Ooh, well that’s kind of three questions. So it does depend on what you’re doing specifically. So for example, something that actually blew my mind. Sorry, this is a relevant tangent, I promise. You saw those models that could generate images, right? So you give them text and they create an image. There’s a version of that, an early version of that that does a pretty good job of creating just about any image from text that runs in something stupid, like I think a hundred megabytes of memory, as in we were able to take the English language and an understanding of creating photos and landscapes and illustrations of pretty much anything and everything you’ve ever seen.

And we were able to fit it in less memory than you’d probably have on your wristwatch. So that’s kind of absurd. But with large language models like the ones you’re seeing, like GPT, they are massive and they have to run in the cloud. So big tech companies, like Google or Microsoft are kind of required to run that at the moment.

There’s a lot of technologists trying to make that smaller and more portable. I expect you will see it on a phone quite possibly as soon as this year. It is rumored Apple is working on that. I’m sure Google is working on that. So we’ll see. But it kind of depends on— how long is a piece of string, because what is your model doing?

If your model is making images, it turns out it could be quite small. If your model is learning the program and answer medical questions, it turns out it needs to be quite big. So we’ll see. And then sorry, the last part of that question was how much power can we hypothesize? Very difficult to say in energy terms. That is likely to be a ten maybe… I should probably say a hundred X to a thousand X increase in the amount of computing power in the world as a result of A.I.

I’m not alone in thinking that. We need incredible increases in processing to do stuff like those robots and like we’re talking like the most powerful chips we currently have on the planet. We need them like in the head of a robot, and we probably need them to be very affordable. So you’re going to see there’s a reason NVIDIA is now $1,000,000,000,000 company.

You’re going to see an explosion in chip manufacturing and high powered computing over the next few years. Hope that answers your question, sorry.

[Jess] Next up, A.I. generated art seems to rely heavily on sampling from real world imagery that’s available on the web. Is there any concern that its capabilities will degrade once A.I. art is sampling from other forms of A.I. art instead of real images? And how good is A.I. at detecting and identifying other A.I. generated art?

[Oliver] This is an interesting question. So the fun thing you may not realize is actually one of the biggest advances we have in generative A.I. is actually not looking at human examples of anything, whether it’s text or programing or art. It’s actually using A.I. to generate training for other A.I.

Synthetic training data, it’s often known as. So in fact, when it comes to images, right now, we’re very deliberately doing what you’re describing, and you would think it would make things worse, it actually makes things better. It’s a little bit like if you took an artist and you got them to practice making their own art, they get better. It’s a little bit like that.

So we can use A.I.s to kind of generate training and A.I.s to judge whether that training data is good, like you know, Is this image attractive? Yes. No. And then we can use that to feed into another A.I. to make it better at doing the same thing and so on. So it’s actually very much the snake eating its own tail and it doesn’t seem to degenerate anything at all.

If anything it makes stuff better, which is maybe counterintuitive, but quite profound.

[Jess] Does Copilot automatically ensure the file itself is accessible with appropriate tags and headings, etc.?

[Oliver] So again, this technology is beta, so I can’t speak authoritatively on that because I’ve not had a chance to use it. I will say Microsoft is definitely strongly orientated around accessibility as a whole, but generally come away pretty impressed with what they’ve been doing.

And so I’d be astounded if they haven’t thought about it. Of course, when they’re generating art to put in a PowerPoint presentation, for example, or in Word, they certainly have the ability to generate the text for it because they often use text to generate the image. So it would certainly be easy for them to accomplish in the abstract how good the end result is. Right now, I can’t say. We’re going to have to get our hands on it. I’d also say to some extent, just bear in mind this technology is beta, right. This is new. And so I’m not going to claim for a second that this is all going to come out of the gate and be perfect to do all of these things in our lives.

It’s going to be flawed. But I’m old enough to remember back in 1994 when the Internet was a new thing and the Internet, people would say, the Internet is so slow, you know. And there were people saying, No it’s going to change everything. You’re going to watch video on the Internet, and they were like, how will you watch a video on Internet?

It takes like five or ten minutes to download a single song on the Internet. And it did, right? But then it didn’t stop the predictions from being wrong. It will take time, but A.I. will get smarter, it will get faster and it will get cheaper. Those things are 100% inevitable. So eventually, maybe not today, I expect that capability will be incredible.

[Jess] I’m going to try to combine some of this. They’re asking about A.I. being used as an agent to help users accomplish tasks, right? Like ordering products from Amazon. They think it could be an accelerator because stores would be incentivized to include it, since it could be perceived as cheaper than remediating current practices. And it kind of brings in the question of do you think browsers like Chrome and Safari will introduce the A.I. features that can interpret websites autonomously?

[Oliver] So fun aside, I’ve actually written a little bit of an experiment where I actually tried to program that myself just to see how feasible it was, and very much an hour or two just playing around. From what I can tell, it’s very, very achievable. I think it’ll come from several angles. So I think it’s likely that people will develop specialist tools like you can write a browser plugin to do it, but I think eventually browsers will probably do it themselves.

And then I also think operating systems will do it themselves. And as most of the browsers are written by companies that make operating systems like Google and Apple and so on, I think you’ll just see that we’ll end up with a multitude of smart assistants with varying levels of capability. And some of them, like you saw 01 earlier, some of them will just work with what we already have.

So they don’t even need specialist knowledge. They’re just, they’re looking at a screen like you would look at your screen and they’re just using a computer. But all of these things are just layers on top. They add a new way of doing an old thing, but the old way will still remain. We’ll still need, for example, Amazon to be accessible.

Their website will still need color contrast. Their website will still need to make sense if you’re using it on a tablet or a phone or for a screen reader or whatever you are using, that will remain. It’s just we’ll have newer, easier ways of doing it, but probably very, very diverse ways of doing it, which if anything, makes it more complicated.

We’ll see. [Jess] The current format of websites, lots of words and menus to navigate already seems overly complex and outdated. How do you see websites evolving to become more user friendly, or will websites even exist in ten years? [Oliver] I absolutely will quite literally bet the farm— well, not so much the farm —my company, on the fact that they will exist.

We are all about the web, but that’s not just coming from a position of self interest. I think if you look at, so… For example, once upon a time we had radio, right? Radio was incredible. And then someone thought up TV and people were like, TV makes radio redundant. It didn’t, right? Radio is still a thing. And then TV’s replaced by the Internet.

Well, no, not really. It turns out we’re not going to lose existing paradigms like using a tablet or a phone or a computer, because I don’t care how good voice assistance is, sometimes I want to be able to sit on a train and view a list of products with my eyesight, if I have it, and be able to tap on things and explore them in my own time.

Right? That is a modality that people would prefer to just having a conversation where I’m like, Yeah, can you buy me a sofa? I mean, I want to see the products if I can. I want to be able to have that interaction. So we’re not going to lose those things. We’re going to add to them, we’re going to enrich them, we’re going to make them better, and we’re going to have better alternatives to them.

But I still think websites make a lot of sense. Sitting and doomscrolling through the web, I hate to say it, is probably still going to be a pastime in 20 years.

[Jess] Someone says, Can you give any insight into what Silktide is building right now using A.I. and what challenges have you faced?

[Oliver] I can, although I have to be careful.

What do I say? Right. Well, yes, we’re doing an awful lot of experimentation with A.I. I can certainly say that. We have done quite a bit of exploration with things like automated remediation or code assistant. And I can say that at the moment A.I. is not quite there for doing things like, for example, this is the exact code change you should make to your website that will fix this problem.

It’s actually kind of reaching at the limits of what current generative models can do. There’s definitely a hint that it will be possible, but it’s not there yet. That doesn’t mean that we wouldn’t introduce something in that space, but it would have to be caveated as, Use with caution. We’re also exploring a lot of synthetic user behavior analysis, which is basically to say, simulating real people behavior with A.I.

So, for example, can we have someone use a screen reader and test your site, or I should say not someone but a synthetic user do that. And in doing so, could you experience and discover problems that would normally only be apparent to a real person using a screen reader, but can you do it programmatically? I personally believe that is possible, but certainly at the moment it’s pushing the fringes of what is possible.

I would imagine functionality like that, so high quality testing, much greater assistance with remediating problems and the ability to have a conversation and interaction with Silktide about those problems is basically inevitable.

[Jess] Someone sneaking in an extra question. What stocks can you invest in to ride the A.I. wave?

[Oliver] Okay, I’ll give you the boring— Okay. So firstly, I’m sure there’s a lawyer out there somewhere who’ll tell me, Please note, this is not professional investor advice, etc., etc. I can tell you the most boring answer. Goodness. I’ve gotta be careful here, haven’t I? The most boring answer is the larger tech companies and even though they are already absurdly valued in part because of this, the truth is I suspect the majority of the value captured in this landscape is going to go to a handful of the existing largest players, if you are a cloud provider like Google or Amazon.

If you are a personal computer manufacturer like Apple or something, you are likely to have the biggest, best, strongest A.I. models. You are likely to distribute them to billions of people, you are likely to capture money from that. Microsoft in particular is very strong. Chip manufacturers like NVIDIA who are killing it again, are doing incredibly well. Avoid Intel like the plague, unfortunately. They’re the ones that aren’t.

When you get beyond that, you’re getting more into speculative territory. And I’m really wary of advising anyone to say, invest in Tesla right now, even though personally I think that’s a good shot. The one thing about this kind of A.I. technology is it tends to require very consolidated investment like billions, tens of billions into a technology that most small companies can’t afford to compete with.

And so I suspect that favors larger players more than it favors new disruptors. And that’s kind of unusual. We haven’t seen that before. But yeah, that’s that’s the TLDR. Big companies and I expect the big companies to become even more unimaginably big.

[Jess] Okay, quickly so we can wrap this up for everyone. What about smartwatches? Do you think A.I. features, besides features like Siri, will be adopted?

[Oliver] There’s a broader category here of wearables, I think. So. I’m currently wearing an Aura ring, which is a ring, I don’t know if you can see this on the camera, but it’s like a computer on a ring around my finger. Likewise, we’ve got things like AirPods and there’s new technology now, like the Rabbit M1 and the A.I. Humane which is like a wearable lapel.

I think you’re going to see a lot of technology like that that is incredibly small. And because of A.I., it can listen to voice, for example, or project voice even into your ears wirelessly. If you haven’t seen that, that’s crazy tech. And assist you in a way that previously a smartphone could not. So we’re probably going to move away from, for a lot of people, from having a phone that is like this disruptive, annoying social thing that gets in the way of your life.

I think going to see a lot more like, lighter technology that is there in case you need to make a phone call or set a reminder, but it’s kind of out the way and a watch is a very good form factor for that. It’s certainly not going to be the only one though. I think I would—

I’d like to see AirPods that have a full blown computer in them that you could leave on all day. And then they would just like listen to you and you could say, you know, can you remind me to do this? But they could also tell you about that meeting you had earlier today and say, Hey, I listened to your meeting, I’ve sent the notes to your email or something.

I think we’ll see that in the next handful of years.

[Jess] That sounds awesome. Right. Thank you, everyone.

[Oliver] Yeah. Yeah, that’s been very helpful, guys. Thank you for taking the time to join. If you’d like to follow along with more, you can check them out at Silktide dot come slash events. We’ll also include a recording of this talk for you to share with your friends.

And that’s it for us today. Thank you very much for attending.