Andy Tyler

this is the second time basically you said that they don't seem to get it. What does it feel like for you?

I think.

It might not be something that you, that is really possible to understand in the first instance. I think it might be something that's learnt over time.

At an event, say, do, as Bruno said, I do a lot of events and there's certain people that I have huge trust in that help at these events and that's because the people get it. Like, you can say, okay, just, okay, can someone just figure out, you know, we've run out of drinks and...

everybody that helps me are like high performing elite startup operators you can kind of just shortcut to say can you figure out what we're doing with drinks I've got to sort out the system you know the there's

Hmm

no power for the for the mic because I've got to sort this out and you sort of trust someone to figure that out

That is when something is is got that is like somebody acting with with full intention and

Hmm.

and adding their

And agency.

own flavour on agency yeah

Welcome to Taste Bench. I'm Bruno Marnett. My co-host is Philip Zahn, and this is a podcast exploring how AI could help creative people do their best work. Our guest today is Andy Tyler, OpenAI's official Codex Ambassador for London. Many people in London's AI community would know Andy through the hackathons, Demo nights, and community events he helps organize. But Andy is not just an AI evangelist, he's also a startup founder and has built many products himself. He also worked for several years on product at the BBC, an old institution with high editorial standards.

Which one would expect to care deeply about taste?

let's imagine tomorrow a new startup comes out of stealth and say, today we've sold Taste. Let's say it's an AI lab that has invented a new capability for AI called Taste. What do you think that product would be doing differently? Like how would that feel different from today's Chat GPT or Claude?

Well, hello, Bruno and Philip. Thank you very much for having me. I think this is a really... This is almost like an acid test of a question, like what to define taste, because I think actually the... This is one of the problems with like AGI in general, right? The definition is...

really up for grabs still and everybody has a different definition. I think actually I saw Sam Holtman saying that was a little bit like when you ask people what their definition of AGI is, it actually says more about what they say as the answer than it does about the actual answer, doesn't really matter. I think that we can substitute taste for that as well. So for me...

I kind of like, when we've chatted before about taste Bruno I kind of see taste as like as a little bit more along the lines of

Less about, that looks good and you know, I like that design and more about something getting it. Like when

Mm-hmm.

somebody actually like understands what you mean on a deeper level. When you can kind of forego the preamble. For example when you're giving somebody an instruction.

and you say we need to do xyz and you can kind of leave them to make the right decisions and when needed ask the right questions that is kind of what AI is trying to do right now and you get a lot of this sort of back and forth of did you mean this did you not mean this but it still doesn't quite get it

in a lot of scenarios and it's difficult to define what like getting it is i

I think

I've just substituted the word taste for get it, but I think that,

Mm-hmm.

you know, you can... I think that that sort of shorthand of being able to just, like, the learned context we have as humans, we're able to say simple tasks and...

You know, like if you're going to get an ice cream, right? Ice cream's

Mm-hmm.

top of my mind right now because it's like 30 degrees in London.

You

But if you ask somebody to go and like, I want to try a few different ice cream flavours. A human would go and get you like a couple of, you know, a Magnum, a Solero, you know, maybe a Twister sort of thing. Whereas an AI, if you ask an AI to go and do that, you you might end up with an entire...

entire ice cream shop or you might end up with one box of vanilla ice cream and then a whole bunch of different toppings. It's very... it doesn't quite get it.

You said AGI is ill-defined, but there's, I think from an emotion perspective, there's something people often talk about, which is feeling the AGI, right? Like this kind of moment where you're

Mmm.

like really impressed by, or having an AGI moment is another expression people use. We see that a lot for technical tasks. Typically coders, they just ask Claude to do something and they're so impressed that it just did it, right? Did you...

Did you ever have a concrete kind of AGI feeling for something that was not a technical task, but something that's more in the nature of subjective taste?

I think this might not count to answer your question, but I think we both had a similar feeling after ChatGP2 Image 2 came out.

Mmm, yeah.

Yeah, Image 2 was a huge... that was a huge AGI moment for me. I mean, I still feel image generation is orders of magnitude more impressive than text generation.

If you just think of the token space, like a pixel can have like so many different values and then the combinations next to that, you know, there's only a few thousand tokens I think in the token space of normal LLMs.

Mm-hmm.

yeah, image generation where text is coherent and layout is clear.

I think that is, that's quite phenomenal to me. Yeah, especially some of the images.

What's

interesting there is when you describe what impressed you, you make it sound kind of a technical progress, right? You're talking about text being

Mmm.

coherent and so on, right? So does this really count as an aesthetic or a more subjective experience, or is this really the technical aspect that impressed you?

I think...

What impressed me was the... So I asked it... I assume you guys have seen Stargate and SG-1 Atlantis, right? I

Yes.

am a huge fan. Recently rewatched the entirety of Stargate Atlantis from the beginning and had a fantastic time doing it. I asked Image 2 to generate a poster I could print out for...

Stargate and I'd never actually seen the words Nakwita which are the Nakwita generators they use to power the Stargate in the Stargate universe and

I'd never seen that written down. It got the spelling right. It laid out everything there is to know about how the Stargate operates. And as big fan, I like, I know, you know, how the chevrons should work and all these things, you know, very nerdy. And it just got all of it right immediately in this, like, maybe it took 60 seconds to make this image. And the image was laid out beautifully as well. It chose a design that looked very much like it had been created by the ancients. And the poster was

weathered in a really tasteful way and it's something that I would happily print out in 4k, like have a 4k print out and stick that on my wall.

Hmm

The taste part of that was not necessarily the visual components that I really got excited about. I got excited about the fact that it understood the detail to include of mentioning an aqua generator, how the chevrons work, the different,

Okay.

the ancient language that it referred to and the symbols it used. It felt like it was actually an artifact that had been created by

somebody on the Stargate production team as like a bit of promotional art. That was really tasteful.

Can I ask you so I I basically I mean you mentioned many things and I took a lot of notes. I'll try to make it brief and not talk too much, but ask questions instead. But what is interesting, what I noted about the things that you mentioned with respect to taste, I think are two dimensions. And for us, there we have multiple working hypotheses, basically, that there are different dimensions.

One you said, I think the last one you basically referred to was consistency. the initial one was also judgment, basically the experience or saying, yeah, this is good or this is bad.

Mmm.

Yeah.

Do you have maybe on both, any do any topic, any domain comes to mind where you feel like judgment has improved massively in the last two years? Where like two years you go back and you feel like you're feeding it something, it's not working,

.

but it's working now. It's gotten better.

I think in judgment wise, I think you can be a lot more lazy with your prompts now. I do remember like a couple years ago you

and

could basically like you had to be so prescriptive like you know I remember writing out like always in caps and then never in caps and like having a list when I was giving a system prompt to like when I was building an agent. Now, like...

you can just be so lazy with what you give it and it will just do it, especially with coding agents. I think that's one of the things that has changed is that the harnesses have changed and therefore the context that it pulls in when it comes to like having judgment, the context is already there or is able to get it so that you don't have to give

you had to do a lot of that judgement yourself before and like, you know, how do want it to respond and all these things and otherwise it would make an assumption that wasn't necessarily what you wanted. Whereas now you can just really can be so lazy.

Is it still true for the more like front end part of coding? Because I'm asking this because I know you're, you tend to like the product side of engineering, right? You like

Mm.

things to look nice. Is it still true even in the more like visual part of the work that you would trust a cloud to come up with nice colors and aesthetics for websites and so on? Or would you feel the need to be prescriptive?

and

I think the short answer is you can get to OK fast, but actual like... And there's a difference right between UI and UX. So the UI can look very nice now, know, gone are the days of everything being this purple blue gradient that...

that it seems to be the, you I think you've called it the average before, Bruno, of the internet. And I like that I can get different designs. I think now we've kind of gone to this sort of...

Actually, the default design is kind of just a nice default rather than I think that whenever you prompt it you still get the same prompts no matter who prompts it or whatever project you're doing you still get similar sort of vibed looking

Mm.

outputs but they are nicer now. But I think once you've gone from creating your first interaction to trying to add more

And I think that this is true generally in all codebase. When you're coding, I'm coming at this from like a coding perspective because it feels like that's all I do now is just generate little tools alongside my, you know, outside my day

Thank

job. I think that there is a steep drop-off in understanding of these models of how user experience should work.

And I think that I've seen this when I've done browser, you know, built browser models and stuff before and done a lot of browser use work. You know, step, doing one step at a time is very, very

Mm-hmm.

clean and it works nicely, but as soon as you try to get it to do a complex thing or something that it has to assume is the next step, like a drop down, clicking a drop down and then like the assumption of whether you could scroll in

in a drop-down that wasn't like a native Mac drop-down to select something. These things are still, they are much, much better nowadays.

I have one of those AGI, feel the AGI moments when I used like codecs computer use the other day, because it is incredibly good. But all the models, all the browser use models, they're okay. And

.

this is where the UX part comes in. They still don't get it. There's still a bit of taste lack.

Can I ask question? Yeah, y this is the second time basically you said that they don't seem to get it. What does it feel like for you? What you know, what's the ring in your brain coming to the heart, whatever it is basically that makes you feel like, yeah, this is for yourself,

not necessarily the model getting it, but you yourself, if you see it, how does it feel like?

If I see something when someone gets it, do you mean? Or...

Yes.

For yourself,

Yeah.

if you get it basically, if you feel like, huh, this is it. And how is it different from model kind of doing a good job for you?

that's a really good question.

When, are you talking when I actually prompt something or if someone generally asks me?

Take it where you want to go. What what comes first?

I think.

It might not be something that you, that is really possible to understand in the first instance. I think it might be something that's learnt over time. Because I'm thinking about when I ask someone, you know...

At an event, say, do, as Bruno said, I do a lot of events and there's certain people that I have huge trust in that help at these events and honestly a lot of the time you kind of just feel like, okay, everybody else is doing stuff now, what am I here to do? And that's because the people get it. Like, you can say, okay, just, okay, can someone just figure out, you know, we've run out of drinks and...

everybody that helps me are like high performing elite startup operators and we kind of operate a lot we take that operation to the events right so I've worked with a lot of them and you can kind of just shortcut to say can you figure out what we're doing with drinks I've got to sort out the system you know the there's

Hmm

no power for the for the mic because I've got to sort this out and you sort of trust someone to figure that out and you know that they will get they they've got it

That sense is a sense of trust that I know that that person has handled much more

Mm.

complex things and more, you know, very, very capable of just figuring something out and I know that if it turns out the drinks are like...

there's no more drinks and there's not going to be any more drinks. I can trust that they have decided that that's the best course of action because for whatever reason, know, it's going to, there's no shops nearby or the delivery is going to not get here

Hmm.

until the event ends. So there's no point. I think it's that trust that you get over long time horizon.

once the task is complete and you sort of review you're like okay I fully understand that you did that that that was the best course of action that I would have done myself. That is when something is is got that is like somebody acting with with full intention and

Hmm.

and adding their

And agency.

own flavour on agency yeah agency and

All the things that you would hire for when you're hiring for a startup, right? For an operator in a startup. That... Yeah, it is, it is. No, no,

This is getting a bit astray. Sorry, continue. I didn't want to. But

I I just want to one more on top. So we know this feeling. There is someone basically, you know, you can you can hand over any task to this person, or you can ask

Mm.

them for anything. But how do you compensate that act in in your engineering practice? Basically, what's the equivalent that you're doing that you know will not get you to a human being, but basically helps you to get your interactions with the AI to work well?

I think this massively depends on what you're doing. think if you're writing production code, a lot of the time you just can't substitute yourself looking at the code and actually going through it the same as if the code had been written by another human and you were doing a PR rather than just an agent has written the code. If you're doing something that...

is throw away. I'm tending to...

spend a lot more time letting it write tests.

When you often are hacking on something, the last thing you're thinking about is tests. This is a tactic that I've taken from you, Bruno, because I'm allergic to tests in all shapes and sizes, although for anybody watching this in the future, tests are very valuable and important. You should

write them still. But when you're hacking for something for yourself, you like, you know, hackathon vibes or prototyping or just doing something for yourself,

You don't care about tests, but the AIs are kind of really want to write tests. Sometimes you have to like stop them from doing it when you see them go, you know, creating slash tests or installing a, know, Jest or something. I think letting, yeah, doing some of these things that add a validation and

I was building a little menu bar.

tool and one thing I found greatly helpful because it itself was assuming that things were going correctly. Sometimes the tests don't take the form of a test. Sometimes you need it to actually like screenshot something and show you so you can see whether the actual... you can add your judgment, your taste and just direct it closer to what you actually think. The solution should be so you can actually stick

it to be, allow your taste to be more prescriptive.

Can I ask you? This is exactly I mean, this is perfect. So let's assume, you know, I want to write an essay. So not not coding.

Mm.

Basically, what's the test equivalent? What's the kind of engineering equivalent that you would tell someone novice basically using these tools and trying to also get his own judgment, his own taste into this into its into his work, which is not coding, which is not math, which is another domain where we see a lot of progress and which is more objective. So really like let's start with writing an essay that

has I I really want to have it some kind of quality from an aesthetic perspective.

find that when you're doing writing, although I tend not to use... I use AI for writing less and less now because I find that generally it

doesn't get it. But I mean, the part of AI that I do use for writing is absolutely dictation. Dictation has gotten to the point where it punctuates everything so beautifully that like...

you can just stream of consciousness. When I do need to do something, if I don't really, I'm not that fussed about it, I will tend to do multi-shot examples. I think that's really useful when you, like one thing that I'm actually really surprised at recently is the performance of the Gemini that they've got going on inside of Gmail.

You know, because I assume they've got so many examples that they can use

Hmm

that it's just really good and it actually does get what you want to say. And times when I've seen that work well when you're writing emails is you copy the thread and obviously in that thread you've got how you've responded before.

.

and it will pump something out that's like a good enough baseline for you to work from. I find that when you're prompting writing, the prompt is basically just as long as the thing that you would have written anyway,

Okay.

so you may as just write.

Can I follow up with this? So in the context of writing, there are different strategies how people approach it. But some writers basically approach it in the sense of okay, I I want to get everything on on one on one page, and then basically the next day I go back to it and very carefully kind of redact it, critique it. you can make the same distinction in the support of AI for you in that specific thing, right? You can ask it to write me this page, or you can use it on the other side, like, hey, I've written this. Help me critique it.

What's your view on these two capabilities? Do they are do you feel like they are very different? Or do you also feel like one is more safer to use? Maybe one is better to use? Any takes on this?

I

The critiquing, 100 % critiquing it and like pushing back on some things, I think it can often explore things that you've not thought about and particularly like not for when you're writing like prose or any sort of you know blog writing or anything.

But for engineering reviews, think AI is just phenomenal for that kind of reviewing and critiquing and exploring all possibilities at once.

For writing, I think it's...

Hmm.

It's good in some scenarios, but often if you ask it to... you put a whole bunch of jumbled thoughts and you ask it to rewrite it, it still won't actually get what you're going for. Yeah, and it still won't...

Get it.

even if you ask it specific parts, like, how can I rewrite this? You know, all the versions it'll come up with will be great, but...

often won't actually hit the point that you're trying to make and

Mm-hmm.

you kind of get to this sort of... When I have used it before to write long form text it will try and blunt some things that you want to be impactful and like...

be spicy as the sort of thing to grab the attention of somebody and it will try and sort of like dumb that down sometimes and be like, you sure you want to say that? When in actual fact that's the juicy part of like what makes the writing that you're saying the takeaway that the person has from having read that writing. there's a little bit too much of the averaging and the sort of blunting going on.

when you ask it to review these things.

Can I make one thing I've noticed so far is that you've mostly focused on the question of understanding. Well, I get it in the sense of, you know, it's going to read my mind, it's going to predict what I'm going to do next. For instance, in the case of a Gmail, that's very clear that in a way it's a prediction engine, right? It just knows that after this email, the polite answer usually looks like this, right? It feels

Yeah.

a little like a prediction. likewise, when it's averaging something in a way, it's just predicting the average behavior. But I guess one...

when possible new capabilities, which we seem to be getting closer to is AI actually making choices, all right? Obviously

Mm-hmm.

the technical implementation of it, maybe the choice is kind of an illusion, right? But at least it might replace choices that human might do, right? There's a lot of studies now about Gen Z using a child GPT to advise on like big life decision, like, should I break up with my girlfriend or not? You know, this sort of a big event, right? Or personally that I know sometime I would love AI to be

good at recommending furniture or things that I just know I don't have yet a lot of taste in and I would like some support and I would like it to cost less than an interior designer. Do you feel ready already to get an AI to make choices for you in a way or you just don't want to? What's your take on that?

I mean, I think we have to remember that while this technology is phenomenal,

Mm-hmm.

you probably shouldn't be making life decisions without

Yeah, I think so.

another human involved. I think probably the best thing to do there is like, you know, we have AI, you have a hammer, everything's a nail.

Let's not forget that sometimes advice from picking up the phone to a friend or a relative is probably the best thing to do. Because I think, and also like on your furniture example,

Hehe.

that's also probably a scenario where you probably don't want to hire a designer to pick a sofa up for you.

shortcuts you can probably take to get someone else, you know, a design thinking friend, you know, you can send a photo of your living room to maybe.

Well, what would make the design thinking friend conceptually better than CHPT today?

I mean, I think there's a big difference of should I move to Australia or should I buy this specific sofa? Like,

Mm-hmm.

I think the consequences are probably quite different. So, maybe...

With this stuff I can cost more, but it depends.

Yeah, yeah, some sofas are outrageously expensive.

Yeah.

That depends on your taste. Do you have expensive taste or not?

Maybe, yeah.

I think... Maybe

that's not the most...

In reality, you probably end up just picking a sofa, rather than asking your friends for their opinion. Maybe if you've got two that you go between. But I think AI in general is just still lacking a massive amount of context. Let's say, know, should I go and take a job? You've got two job offers, one in Australia, one in London, and you live in London. Should you take it?

there's what are your general life ambitions, you If you say, you can prompt AI to give you an answer in any which direction you want, like you can just preload that at the start and basically say, you know, be very pro London or be very pro whatever, moving abroad. Or even indirectly, you might have accidentally given it the context that like,

you say you've already you've got some flight tickets to Australia in your your gmail and it's already looked at them and thought you're going already maybe you you want to go and live there so i think that when it comes to like personal very personal things the human experience the human lived experience still has a tremendous amount of value

Where do you s sorry, where do you

Yeah.

draw the line? So I give you an example. If you're reading something if you're reading a book from the whatever, the Middle Ages, or f even from the Romans, from Greek literature.

All the models basically have massive context compared to normal human readers, right? And I think it's the same. I mean if you take Bruno's example and you would ask, help me with this furniture. And I think there there is some there there is this kind of trade off where clearly it there was personal input, but there's also the other side where it can you can query it basically and and get get information and maybe help yourself forming tastes about certain things or informing at least yourself about this. Any thoughts on this?

I

think, yeah, I think you've absolutely nailed, like, that's the line, to be honest. Understanding, anything that helps you understand and inform you to make an opinion is where AI is just so incredibly

.

good. Like, especially deep, complex, like, spaces that you...

want to have an opinion on, but you don't have the time to become an expert. AI is not going to make you an expert, but it's certainly able to help bring the relevant information up to the top that would have taken hours of research, a whole evening of research, you can get in 30 seconds now. That is my most favourite use case for AI.

just being able to be more informed when you're making decisions, especially on more scientific topics, let's say, more fact-based topics rather than

Okay. you

opinion-based topics, and even like being able to...

understand what the pros and cons are of both sides. You know, there's been a lot of work to make sure that AI is not recommending, you know, that you go off and commit crimes and all these things. So you often do get a very balanced objective view. You know, obviously, you know, AI can make mistakes, as it says at the bottom of all of the chatbots.

but you do get quite a good baseline to be able to make your own opinions on. That, I think is, when you come at it from that angle, it is so, powerful. So powerful.

So it's basically the analytical capabilities that in whatever taste related or judgment related things you need, there's a capability basically that you can always lean on, which is not always right, but it's it has so much capability compared to the the state like five years ago.

Yeah, absolutely. Being able to reason over a bunch of information at speed, it's something that previously only humans were able to do and it took a long time. Yeah, yeah.

And reading a lot of books.

Maybe just to push the thought experiment a bit further. So imagine now an AI is infinitely smart, in a AGI sense of it has been trained on all the math, the coding in the world, like every kind of formal domains.

all information possible about you. Let's imagine you've recorded all your life and it's all in the context of the AI. Would you say this AI is now going to be more helpful in a situation that requires taste compared to a real human being, or would it still be missing something?

I have some thought about what

This is it.

could be missing, I'm curious what like infinite intelligence plus infinite

This is very, very interesting.

context does this cover taste or is this something else?

I it's difficult to know because there's no AI that has the long enough context to be able to understand your entire life, right? I think it depends on what you're asking it about, right? If we're talking about taste of...

website

or the UX of a site, that's

Yeah.

one thing. Life decisions, big life decisions, you know, like should I break up with this person and such. I've tried to have conversations with a lot of different AI about personal topics. There's no substitute for a human conversation. I think there's something inherent to the lived human experience that

Even somebody that you like, even other humans sometimes give really bad advice. And I don't know if we actually want an AI that's going to be able to make decisions for us. You know, what is the scenario where we would actually want an AI to make decisions for us that had such huge consequence for us?

Okay. Okay.

I think consequence as well is quite tied to that. So I think to answer your

question in short, no, I don't think we would ever even, I don't think we would want that if that was capable. I think that maybe taking that as at the same weight as another human's, like thoughts about a topic, yeah, okay, but...

then I think you have to have a deeper conversation then about what are the motives of that AI? How is that being trained and prompted to, you know, is it's core desires to just output

Hmm.

text or, you know, does it have other motives underneath of wanting to continue the conversation? you know, how has it been trained? Ultimately.

Maybe short of taking decisions for you. Let's actually maybe take one step back. Let's say we're looking at artifacts produced by AI and not just code.

you know, maybe literature, music. What is your personal if if you're up for it, what is your personal subjective

Take on that. Do you make a difference between music being produced by human or by AI?

I mean this in itself is so hot right now because you know you have a lot of models that can output some quite high quality images, music, videos, know the sort of stuff that would take millions of pounds to create even with especially like these you know Sora. I was playing with Sora the other day trying to get it to generate goblins and...

Phenomenal. I could know it would have taken me weeks months years to be able to do that I think actually though that with any art you can take two pieces of art. Let's look at Banksy, right? Banksy's are worth millions of dollars Like anybody can actually recreate them though They're not that technically difficult to do But what gives art is the meaning behind them and I think that's the same with all

.

all forms of creativity is, you know, one person can release a song and it not do that great and another person can re-record it and they give it a whole new take and it takes on a world of its own and that's a completely different set of meaning, it might have more value, the same with any piece of art. So, I think...

It's kind of irrelevant whether something was generated with AI or not, because I think that the ultimate thing is what is the meaning behind it? And...

Maybe that was a bit too harsh to say it's irrelevant whether it was created with AI or not, but I think it's what's more important actually than whether something was created with AI was what is the human behind that AI? What did they want to create? If we take it as like a piece of digital art that someone's created in Illustrator or something.

that's a tool that they've used to create that and I think we'd all, you know, just because it's a charcoal style SVG vector doesn't mean that it, that you know, that choice of it to be charcoal has not got meaning. AI I think might, and least one root of it could be that it becomes another paintbrush, another tool in the artist's vocabulary.

I'm certainly not offended by...

creations created. I mean it's my favorite thing to do is to sit and generate loads of images like you

Mm-hmm.

know especially like stickers. I generate loads of stickers like I had to get designers to do that in the past. I can generate them myself now. I love being, I love doing that. It's so much fun, so much fun.

So

it's actually there is an aspect of democratization, right? That

Mm.

you get access to instruments that as you said before, certain things would have costed millions of dollars for certain people to do. It would have o needed organizing a whole team where now it's basically you're as an individual have the authorship

on certain things that you didn't have before.

But I think that by democratizing it, there's also an element of devaluing it as well. And I think that that puts more emphasis on the meaning behind things. There's a lot more

you

like disposability with creativity now. So if someone's taken a lot of time over something...

.

it has a lot more meaning. Like, for example,

my sister's an incredible, incredibly gifted artist. She does all sorts of amazing things. And the sheer fact that she creates something...

has more value, whether she's used AI or not, you know, she's given me many gifts over the years of things that she's created. Those are treasured because they're from her, not how she created them is, it adds layers of flavor.

But it's still the meaning as to why they were created and how they were created. It is more part of the story that AI being involved might just be one component of that.

So do you think we're gonna land on AI becoming a tool like a calculator or a paintbrush? Or do you think we will eventually think of this technology as something that is different because paintbrush doesn't make decisions, right? Whereas when you ask...

to add a baseline to a song you wrote, it's going to decide what notes to on the baseline. I think there's already a really qualitative difference between what AI does versus the paintbrush. Do you think it's just going to get normalized, which is going to redefine what tools mean to include that more intelligent form of tooling, or the line will be blurry?

I think you could argue this any which way and I think that it's way too early to know what the future, what the ultimate future for artists is going to be. I think you could maybe say that having a template in a music generation, something that people genuinely use to make music, know, garage band or something, is the inbuilt template for something or the backing track that's been generated.

Does that have the same sort of effect? I think ultimately there's going to be camps on both sides throughout time in the same way that a lot of people, you know, you'll find people in the world who would say that, you know, art can only be, you know, oil paint on canvas and anything else is just some, know...

it is irrelevant. There's also a lot of people that would strongly disagree and say that, you know, things created, can come in lots of different forms and shapes, etc. I'd like to think that it's not going to affect a lot of... There's some scenarios where it's not going to actually affect...

art in a way, in the first instance you know, but I'm thinking about pottery and clay

Mm-hmm.

and physical forms of art, although I do say that I did go to an event last year where they'd hooked up a robotic arm to a paintbrush and it was painting art throughout the event,

Mm-hmm.

so maybe that's just a matter of time but I think that no matter

what the ultimate end is for this, the story will still matter more than the art itself. Why has this been created by, you why did that person prompt something to be created? Is going to be more part of the story.

Maybe to,

I find it interesting you use the word person, right? Because we're moving towards the world where things sometimes I get so automated that maybe there's no single person behind things, right? In a way you can think of, you know, the Facebook and social media as like a machine itself that's already like pushing things at people and there's not like a single person responsible for what, you know, ends up in front of your eyes. I guess maybe just pushing that, I think we're getting close to...

Kind of the end, maybe pushing that question of choice again. I think we discussed this before the video of Stephen Fry recently, I got viral, where he talked about his hate list. Someone recognized for having strong taste, someone extremely prolific in terms of how many books, how many plays he's written and so on. And yet it's not neutral at all. Obviously he has that...

strong opinions about many things, a sudden courage of opinion. Do you think in the future we will only have kind of like neutral AI again, like think of them as tools, they're only here to reflect the taste of the prompt engineer behind them? Or do you think there's a place for a more autonomous AI that actually have their own hate list?

I think that that is just pushing the buck back. You're just moving... I think it just moves the goalposts a little bit because actually

Mmm.

what you've done is if you're saying the AI itself, yes, you remove the prompter from it, the person who's prompted the system prompt, and you go back a level and you look at the actual AI or the LLM...

you then, the taste is already baked into it. Well, that's baked in by the people who trained it and the decisions made by, you know, what training data

yes.

was used. And, you know, when they were doing, if they were doing any alignment post-training deciding on the personality, it comes a lot back to the taste you have when somebody creates a product.

It's very much the opinion of a piece of software that helps you write a to-do list

Mm-hmm.

is often a reflection of the opinions of the person who created it or the team who created it rather than the thing itself.

It doesn't have opinions when it's telling you you need to input certain bits of information. That's almost from the creator flowing through it.

Hello there.

Can I ask a last question? sorry, we have been taking on you for quite a while already. this is gonna be a short one.

Now this is fun.

say you're you know, there's there's an artist say listening to this, not really experienced with you know trying out AI a bit, maybe prompting a bit, maybe chatting a bit. what would be the one thing you would like them to experiment on? Let's say Codecs. What's the thing they should try that?

basically opens up a new dimension for them that most people that are not technical are really not seeing that much. Like this aha moment that you and Bruno basically experienced before. What would you think is the thing that might open it up for them?

there are so many things, because in the past week alone, not to shell codex, but the releases have just been incredible. Like, I tried Computer UC today and it's mind-blowing. But I think actually just image generation, full stop. Like, the more creative parts of AI.

image, video, I would challenge anybody to try it and not find it so addictive to just create, create, create. I have found it very difficult and I've actually made tools for myself to be able to create more things in parallel to then actually figure out what is it that I want to create.

Mm.

That can be so...

know, refreshing to be able to have a hate list like Stephen Freymiller. I absolutely hate that he did that. And figure out what you are trying to create yourself. It can be just so, so freeing to just generate and have an opinion on something. The colors, the style.

Yeah, I find that an incredible use of time when you're doing something creative.

I think there is no better way to stop than this. Honestly. Thanks, Andy. Thanks for being part of this.

Thank you.

Brilliant.

Yeah, thanks so much. Thank

No, thank you. Thank you both. Thanks.

you. Bye bye.

Thanks for watching. We've got more episodes in the oven, so please like and subscribe to get notified when they're out. And please check out taste-bench.com to read about the taste benchmark and the creative engine we're building towards.