Arseny Zarechnev

another thing that I noticed is that it began to reduce the amount of communication between us humans. Like humanity suffers a little bit from it, I think.

As an example, a few times I had a question, which in pre-AI times, I for sure would type into company Slack, like a question sort of like about best practices and your experience. How would you approach this? Right. I would, I would get some wisdom from, my peers and several times already I started typing this and then I thought, well,

would I distract the people from the work that they're doing?

And then I might just rather ask Opus and then you ask Opus and you get every detail outlined for you with the wisdom from the whole internet put together. can query it as much as you can, get all of the details out. yeah, productivity, huge, huge gain for productivity. Problem is solved, but at the same time, less interaction with

of humans, right?

Mm.

okay. Welcome to Taste Bench. I'm Bruno Marnett. My co-host is Philip Tan, and this is a podcast about AI and taste. In addition to our usual Wednesday episodes, we're releasing a bonus episode today, Friday, because we recorded too many of them. And our guest is my old friend Arseni Zarashnev. Arseni is a tech entrepreneur who previously co-founded a cryptocurrency exchange called Mothership, and he's also a very talented software engineer who has worked in many places, including Atlassian, Enki, and Ostrobock.

We're very excited to talk to him today because Arsene is one of the most AI peeled engineers I've ever met. He's a power user of all the modern AI coding tools. He has also built many AI agents himself. And this was particularly interesting for people who are not engineers to understand the extent to which AI is changing the lives of people who use it the most.

thank you Arsenal for joining us. I'd love to

I'd love to start with a question about what AI can do or cannot do. there's been a lot of experiments recently about people trying AI to run vending machines and other kind of shops with the only purpose to make money. what I'm curious about do you think an AI could run a shop?

I do have a couple accountants that I know of. Even my mother-in-law is an accountant and I once asked them and excited to this part by what I can do now. Hey, can you tell me if you had like this super cheap and super smart employee, what are the problems that it could solve for you? Both of them, Dave.

fought for a while and came back with a set of problems that really only humans could resolve. And it seems like that they have all of the boring things sort of figured out or they're used to them already. And either way, those small boring things turn out to be not the problems that really don't let them sleep at night.

So what what are some what are some examples things that like only human can do for them?

Examples such as calming down an angry client or figuring out some really custom, so to say, text optimization strategies, things where an AI probably would not be as helpful.

I see. So like the what are the people aspect? I I guess the Yeah.

More like the people aspect, yeah.

so that's a nice contrast with like the the famous kind of vending machine experiments, which don't really have people dimension. The the other dimension I was curious about is the I guess is aesthetics involved, right? Like if you are if you're running, say, a jewelry shop. do you think there is limitations there as well?

I think there is a universal limitation where that when people are searching for a unique piece of jewelry or let's say a unique piece of art, a lot of value comes from knowing that some craftsperson has put their time and efforts and struggle and artistic tears

Mm-hmm.

into making this piece of art, so this jewelry object.

And I think this fact would survive through the ages, survive the AI revolution and whatnot.

Can I ask you I before you talked a bit about jewellery, you said there are different production ways of doing it. More by hand, more into the indu industrialized way. Can you make can you explain a bit of the distinction? Because I think it's it's actually there's a parallel that is interesting to explore.

Yeah, there was a big shift recently when 3D printers became more generally available and also the resolution became very, very good to produce these tiny elements on the pieces of jewelry. so currently there are two main production techniques. the so-to-say industrial one is called casting, where you cast the metal.

You first 3D print the object, and then you make a mold for it, and you pour molten metal inside, typically with an assistance of a vacuum machine. So it pulls the metal in all of those little crevices and creates fine details. And this contrasted to a more traditional approach, where it's how you imagine jewelers work.

a guy sits at his desk and files away at a ring for hours and hours and then making everything by hand. so yeah, the casting is definitely the way that most jewelry items are made these days. The by hand method is still alive in the repair, I guess, and still alive as...

really useful to learn the craft. And even with casting, usually you cannot, it's not as simple as I described that you just squeeze some metal in and you have the final piece. In more complex pieces, you would have several details that you still need to clean up, you need to assemble them, and you need to put the stones in. So there's still a huge amount of work there.

And of course, traditional training helps in performing this kind of work.

Casting method which is the new method. Can you do actually kind of jewelry that was impossible to do before?

Yes, yes, definitely.

So we already had a state where machines in a way can be augmenting, I guess, true hours. is this is this pure kind of physics? Is this more about like, you know, machines have got like to a better precision or does this involve some the AI, basically?

I think it does not involve any AI and it's mostly physics, but it's also, think, maybe economical sense, right? Because of course, a dedicated person could reconstruct something made with custom techniques. It would just cost astronomical amount of

Mm.

money and like huge amount of...

What maybe just taking like a step back, like I'm I I'm curious what drove you to into jewelry making. So you've known each other for a while, you've started from a career more in engineering, and which is very digital, and then you you spent some time in the physical world of jewellery. Wha what what made you choose kind of this activity among the other choices?

the exact thing, how you phrased your question, Bruno, it's a thought occurred to me. Hey, I was moving beats all my life and it's kind of interesting to try and move some, atoms around. And I think I came to this as, entirely by accident. Honestly. we were, with my back then fiance, we were.

choosing rings for ourselves. And we didn't really like anything that was available in the stores here in Estonia. And half jokingly on the way back home in the car, I said, my God, what's there to do? Like, it's just a piece of metal, boom, boom, and you got the ring. And then when I came home, I watched some YouTube and found it's quite an interesting thing to try.

estimated how much would it take to just buy the very basic equipment from Amazon that I would need to make a couple of wedding bands. And decided to give it a shot. And somehow later, maybe four months later, I ended up in a three-month jewellery program in Rome, studying like actual jewellery making.

How does that work, this program? How is it structured? What do you learn?

It's very hands-on. There is maybe two hours of theory in the very beginning, and then you just learn and learn and learn. And in modern jewelry making, there are two branches of what you can do. You either train to set stones, or you train to make rings, necklaces, and like the objects themselves.

And that sorts of separate and like separate.

branches of the profession. But you can try to learn both and see which one you like more. And of course, there are shared understandings you need to have between them.

Now that you know a bit about I mean a bit you know quite a lot actually about the production of jewellery, how has your judgment or your evaluation of if you're looking at a piece of jewellery, how does how has your evaluation of that changed, given what you've learned?

It did change, in having more appreciation for how things are made. I definitely held jewelry in lower regard. So I never even thought about it, you know, cause for me it's, it was more like an outlet for working with atoms, working with metal, right? It's, I had a past time where after work.

I'd watch some blacksmithing videos on YouTube, just as you know, ASMR

I don't mean something.

sort of thing, turn your brain off and watch. So it's, guess, some more practical version of that in the modern world where it's certainly harder to sell a sword nowadays than to sell a diamond ring.

Your case. So I mean you what started basically was your then marriage. This is probably a likely one for other people to look at jewelry seriously the first time. so basically they have one shot at evaluating. How if you would give them a recommendation how to improve their choosing or generally think about the choice process that they undergo, is there anything that comes to mind?

I think what comes to mind is I'd recommend to search for your local jeweler and come and make a custom order. Because it would not be more expensive than the store bought thing. Sometimes it would be even less expensive, but it would be truly yours and the jeweler would be able to give you options and ideas and help you.

to make something cool and memorable and unique.

In the in in a way, mm many of the things you describe so we underst we understand this difference in the digital and the atoms, but you're essentially talking about craftsmanship, right? And the I guess in your career as a software engineer, you also had a very similar kind of concept of craftsmanship, right? For a long time, you must have learned how to do things the right way and so on. Obviously, the the world has changed a lot in the last few years, and now

Like the way we make software is is a lot more automated. do you see a parallel between like the way software is now auto-generated and the way jewellery is now often made by machines? In at least you you mentioned one type of it is made by machines.

think that transformation is still a little bit different because in jewelry, yeah, it's a machine, but it's still a tool, right? It still requires people to design the object, to control the process of casting, which is also a craft in itself. And it's not that easy as to just pour the metal in and then to assemble it. In general, it's still...

quite a lot of artistic and manual work where I think in software nowadays you can skip all that and you can go from an idea to a product within an hour, right? Without ever designing anything.

So you were saying the in the case of software, AI is more than just a tool. in which sense it is more than a tool?

Cause we taught it to make decisions for us. that right? Cause imagine if you, if you ask an AI to build, build an app, why, why software was so expensive before AI, right? It's because essentially what the software developer does is he makes, and the software designer as well. They make lots and lots of tiny decisions, right?

Mm-hmm.

It's not just, let's make a calendar page. It's like, okay, now what elements do we have? How do we position them? Do we query the calendar for every user or which user should have access to this page and so on and so on. And there from like a one sentence task, you end up having to make hundreds, if not thousands of tiny, tiny decisions on design level.

on architectural level, on actual code level. And I think what modern AI tools do is they make all these decisions for us, hopefully using best practices and the wisdom of the crowds.

Is is the result that the decision is sometimes better or or is it I I'm I'm curious in a way that so in the case of jewelry you you probably care about you know making the the final result quite pretty and custom and personalized and so on. in the in the case of software, there could be a same

Imperative, right? Like if you care deeply about what the product looks like and so on, you you may want to still do many of those decisions yourself. but do I understand correctly your today that you would actually trust the machine more than yourself to do some of the small decisions?

Yeah, sometimes to my detriment because we learn to trust it so much. Claude does. Really not often it makes mistakes, but your trust grows and grows. And then when it makes a mistake, it's really painful sometimes because you lost your guard.

I think to expand on your question, I think it depends on the type of software, right? So if you need some throw away piece of software, like a calculator for one meeting or one presentation, is perfect, right? If you're trying to compete in lifestyle productivity apps for iOS, well,

Mm-hmm.

I think there is still quite a lot of space for...

creating unique designs, interesting approaches, and polishing your user experience and the usual things.

Ask you a question about in the in the overall space of AI, image creation, video creation, coding, whatever, pick pick your favorite. Where are where do you feel is the most kind of aha moment for where where was the biggest aha moment for you or the biggest kind of impact, improvement, whatever you want to call it, that you recognized in the last years?

Definitely engineering, it's definitely engineering and programming. And we have to see this propagate to other areas just as much because software engineering and machine learning as well, AI research, they transformed so much. It's just chatbots now. And rarely you do need to dive into the code itself. Sometimes to detriment, but most often it changed them.

equation a lot, like the costs of diving into the code. Now something like 10x slower, you become 10x slower when you start paying attention to what's actually going on. And sometimes it's just super, super hard to fit into team dynamics, market dynamics, everything is moving so fast. sometimes I guess it's just easier.

If you are able to throw away and start all over again, then to try and fix something that broke along the way, right?

So suppose you have person who's not really familiar with AI so far. Maybe they have used

Mmm.

ChatGPT superficially, maybe they got also a kind of a a wrong answer and they were like, faulty machines producing just slop. Is there anything on the creative where you would say, Okay, if I want to introduce and make this person clear there is actually something significant happening, what would you pick to show him?

I think coding again, I think coding again. have a couple friends who are not programmers and have achieved quite good results with modern AI with or without trying to learn more about the basics of coding. But yeah, one guy is a comedian. He managed to build his own ticketing solution.

because he was annoyed with the commissions, the ticket master charges. And he was like, Claude, please help me. And boom and run. He's selling hundreds of tickets now every month via his own platform. So I think still engineering is the most impactful. I think Bruno had quite a few things to add to this topic.

Yes. So the so obviously AI has impressed many people by generating code and you just hinted at the fact that it could expand to other domains, so we could see other domains where it makes kind of micro decisions, you know, better than us. would you imagine AI to be able to do micro decision in a space like I don't know, interior design? Let's say you want to decide furniture in your living room and you have to decide a lot of little details, you know, on colours and shapes and so on.

Would you today already start trusting Chat GPT to make some of those decisions, or do you think there is something missing?

I think I trust it because you know, would you trust Ikea? Yeah, you would trust Ikea, right?

It's all relative. Actually I would

Funnily enough,

I I AKEA actually has brought out or is bringing out a system where you can use AI to basically choose faster and kind of more wholesome.

But

actually to be honest, I I do trust that IKEA has spent a lot of time thinking about consistency. Like if you IKEA is about building a whole set, right? Of things that have been co-designed together to kind of match, right? so it's it's obviously lacking originality if everyone has the same kind of IKEA design living room, but

I I guess the question is more like w when there is a choice. If you're not like buying an entire set that has already been like pre designed for you. Like if you need to just add something in the living room that already has its own like constraints, something. Do you think you you would trust Chat GPT to choose like the color of your sofa or or will you still go to a person?

would trust my wife. I would trust my wife. But

Okay.

huh.

otherwise, think to try to actually answer the question, I think in these kinds of problems, the solution would come from a vendor, right? Because yeah, it would draw you if you ask like plain chat GPT, not a

Mm-hmm.

dedicated system built by a vendor or furniture.

that would suggest you like actual real furniture. But if you ask ChatGPT, you would get some generic couch. Maybe it would be, maybe not generic, maybe it would be an awesome couch. But then how would you source it, right? It's a hallucinated couch. And on one

Yeah.

hand, you're saving time and money on hiring a human designer. On the other hand, you now have to spend 10 times more on making custom piece of furniture. it's...

Not really a reasonable solution, right? It's not saving anything.

I I'd see the the practicality of yeah sourcing some attayo. But but again, imagine maybe five years in the future, you know, ChatGPT has all the tools, he can, you know, call suppliers, it can, you know, arrange everything for you. I I guess I'm trying to understand. Like, do you see a limit where you start needing some sort of like aesthetic judgment or something more out of the the core kind of expertise or like the core skills of Chat GPT? Or do you think everything can be somehow rationalized as

you know, training a machine to make good recommendations on sofas.

Hmm, I think that with time, humans would continue to push maybe in collaboration with AI, maybe by just our own human ingenuity, we would continue to push design aspects and interesting solutions further. And as for the AI, it would continue to pull up the floor, right? Because...

It would be, you mentioned originality, right?

Mm-hmm.

I don't doubt that AI is able to produce good quality design decisions, but it's taste and design, right? It's an area where when everybody starts doing the same, it's not cool anymore, right?

Mm-hmm.

And that's, think, a forever problem of the generation of AIs that we have now.

with the websites, right, the frontend skill for Claude.

Mm-hmm.

I was amazed when it came out how my prototypes started looking like, the designs that it outputs. And then a week later, I see half of the internet covered in the same very designs. You're not as impressed anymore. It's like, okay, I need to be more creative again.

Interesting. The I mean some people are worrying that so web design is one example, you see that, but you still see it in in writings, right? Like lots of LinkedIn posts these days just all sound the same. Like you can tell a half of the people are just generating their content with cloud. do you think we're kind of losing something? Like just I at a more like global scale, you know, like at the as a culture or something, when people are all starting to use the same generators.

Like i is is there a bit of a concern that life is gonna get dull, basically.

I, I yet did not dig too deep into this, but I did notice a couple, a couple things, how it affects my own writing from chatting so much with this agent at work. I see that sometimes when I'm messaging humans, like I'm writing my thoughts out, I see the writing style of Opus like imprinted into my brain. My honest take is, no, no, no, no, I'm not typing that.

I had quite a few moments like that where it's just like the structure of thought, the structure of your paragraphs starts resembling AI output slightly more. And this is a bit scary, I guess. And another thing that I noticed is that it began to reduce the amount of communication between us humans. Like humanity suffers a little bit from it, I think.

As an example, a few times I had a question, which in pre-AI times, I for sure would type into company Slack, like a question sort of like about best practices and your experience. How would you approach this? Right. I would, I would get some wisdom from, my peers and several times already I started typing this and then I thought, well,

would I distract the people from the work that they're doing?

Yeah.

And then I might just rather ask Opus and then you ask Opus and you get every detail outlined for you with the wisdom from the whole internet put together. can query it as much as you can, get all of the details out. yeah, productivity, huge, huge gain for productivity. Problem is solved, but at the same time, less interaction with

of humans, right?

Mm.

Yeah, that's kind of sad. I have the same frustration as well.

Do you have any what do you do when you find yourself sounding yours basically your own writing sounding more like like the AI? Is there anything you're intentionally doing to counter that?

I think intentionally I definitely am not using any tools like Grammarly or tools that correct my writing. just keep it as is because in the age of AI writing and like this perfect grammar correction tool, I hope that in my writing there is at least some style, know, how I like to communicate. But I guess practically nothing.

It's just something that I observed proactively well. The thought is communicated, the agents move forward or teammates are updated. So it's fine sometimes. Yeah, you just go back, correct yourself, try to phrase it differently.

You also engaged in other domains like music, as far as I know, electronic music. Is there any role for AI there? How do you see the lo maybe broadly, the music landscape being shifting due to AI? What's your take on that?

I

think the take is applicable to every artistic domain out there is that humans will keep pushing. AI will keep catching up and every now and then inventing something cool.

But then again, it's...

It was easy enough already to produce like pre-AI slope, right? A simple jingle, a simple beat.

Yeah.

And it's insanely hard without an AI to create something, create a hit, right? Create something that would resonate with a lot of people, touch your heart and so on.

So far, yes, we have quite a few of AI records climbing the charts, but it still doesn't touch your heart as much.

Why is that?

I don't know. I don't know why.

So,

Maybe it's my human bias. Maybe...

Do you think the

the the ones that are successful on Spotify, like the the country music generated song and so on, do you think they're they succeeded because they somehow hacked the algorithm or because there's like something good in them?

I I guess first question is like have you have you listened to some of them? Have you have you or have you just

no, not really. That's not, my genre. Yeah. That's not what I typically listen to.

I I'm asking the one one impression I had is that it was very you know a very safe music. It was really like very very generic sounding. what possibly made it succeed is that it's very clean. It's like that AI did a really good job at making sure like it's easy to consume. Just like, you know, Autotune when it came out, you know, make some made something successful compared to the ones who are not using Autotune. But but I wonder if there is maybe an illusion of

You know, we say that AI is kind of creating music, but in a way it's just making music an easier to digest. W do you think that's a possible reason why it's

Maybe,

maybe if it sounds about right, sounds about right maybe.

Do do you use it yourself in like when you do electronic music now or?

No,

no, no, no, it takes all the joy out. Like what's the, what's the point then?

interesting.

What about movies? I mean I guess you know, before that, yeah y y you know, most people do not or were not engaged in doing music sorry, doing making mu movies, or clips. Is it something you use? How do you think about that?

I'm not using it, but I'm really looking forward to talented people creating like whole TV shows, YouTube shows or movies with the help of AI. And again, I guess that would be at this point for sure a combination of human creativity, ingenuity and AI used as a tool.

that now instead of millions of dollars, huge camera rig and so on, you need a couple thousand bucks of tokens and there you go. Of course, still it's 15, 20 second generations, right? So there is a lot of human efforts in keeping things consistent, editing, writing the script and just generally making something interesting. So yeah, in this area,

I'm really looking forward to it.

You said in in the case of electronic music they would take all the joy out of it. would you say it would be the same if you were like generating your your videos with it?

Would it also take some of the joy, or in this case that would just be a cost saver?

I think for people

who are in love with the movie industry and like the processes that make a movie, a movie, right? The set, the props,

Mm.

the aesthetics of Hollywood movie making. Yeah, it takes the whole joy out of it, right? For sure. So same for me with music. just love to tinker with synthesizers and not necessarily like...

turn it into Finnish tracks even, rather like jam, jam, jam with friends, having a beer and just playing some electronic music together and getting really excited when something cool and unexpected happens, right? So I don't think sitting there and prompting the AI would ever replace that.

Did he take some of the joy encoding?

Yes and no. Yes and no. I think that I did go through some sorts of psychosis periods, which starts on like, I'm doing jewelry and music. Yeah, but these are like hobby level compared to software development, which has been my bread and butter for like more than 15 years now. So of course, when AIs came out, I did succumb to...

AI psychosis and trying to figure out what's my value. But I think enough time passed to see that there is still value in an expert human guiding these AIs, right? Knowing where to look at. And I don't think I miss the code encoding itself sometimes, but still there is enough...

Enough things to get your brain busy with on like a higher level. There are still so, so many decisions you have to make how to allocate resources, which are also finite, right? How to spend those tokens, where to spend those tokens and what problems. So I wouldn't say the joy is taken out.

But for many people it is...

Yeah.

Yeah you when you talk about when you're talking about micro decision, you're thinking, you know, you gave example more on the like front end, like small changes, right? one trend right now with AI is that we can do longer and longer job, right? We can do bigger tasks at once, which is kind of a different level of autonomy, right? Like you can get an AI to think for an entire day about some sort of research problem, right? Like you try different ideas and so on.

In in that kind of context, in the kind of the bigger task when there's more autonomy, do you do you feel we are still limited by how kind of creative AI can get? Or like I I'm curious that like would you trust AI, for instance, to to come up with good research ideas or to invent new I don't know, architecture or new solutions? Or do you still feel like you have to do all the kind of creative work yourself when it comes to the bigger ideas?

I think that's somewhere in the middle. think we're at a point where AI can combine ideas well. The human should sort of kickstart the AI's journey by providing the right starting materials, maybe seeding it with an idea, seeding it with, in case of research, seeding it with a few papers, such as...

some suggestions for initial research direction. But then, from what I observed, there is enough creativity to at least combine things, right? Maybe not invent new things, but at least try to combine useful approaches from two different papers into one thing and get some sort of result.

There is also, they are also pretty good at methodically applying scientific methods. If you task the AI to come up with a theory, set of experiments to confirm or reject it, then run it and reflect, analyze the results,

Unit.

come up with the same and go on in the loop, come up with the next theory and so on. They do get places.

They do get places.

What you describe sounds f from the for the science part describes sounds to me like that the thing where it seems to affect be effective is to

mimic or replicate processes and not just, you know, basic decisions or single things, but more like we are almost like a pipeline. Yeah, we're doing data science that you pu you plug things together. it's like create different hypotheses, evaluate them, etc. Would you agree with that? And what would be the thing that would make it more powerful or maybe even also more creative if if that is available at all?

Sorry, could you rephrase the first part of your question?

Yes. Okay, let me let me try to reformulate it.

Yeah.

what you described as coming up with experiments evaluating it sounds to me like what it seems to be good at is kind of automating parts of the processes that were a kind of a mixture between humans doing something, then going back to data, maybe doing something theoretical and then plugging again data. So it seems like it can merge these step by step things which are separate into automated workflows.

So the innovation there seems to be mostly on the workflow side. Would you agree with that?

no, not really because in the process that I described after performing an experiment and looking at the results, we asked the AI to come up with the next theory. It has some sort of a goal where the metric should go. And as the context grows and it makes more theories and experiments and writes down some notes and it's like research journal,

continues to come up with more and more interesting theories. So it's not purely, it's not just iterating hyperparameters, right? It's just like changing the variable and doing the sweeps, two X, four X, eight X, whatever. For me, one of my favorite things to ask the AI to do is to reflect, like deeply reflect and think back to the conversation and...

and then come up with theories or move forward. Because it does deeply reflect and think about what we've done so far.

Right.

And it can notice that we have tried a certain approach like a dozen times now, probably we should declare it a dead end and come back to the scratching board, right? And try some new approach or...

see some sort of thread that we missed throughout all of the experiments and find something that was missing, or maybe decide to go and look for relevant literature and maybe there is a solution out there waiting to be like strapped on top of the problem.

In what you said you assume there is or basically you said there is a overall goal that the AI follows, in which sense is this goal also revised in that process?

It's said by a human, right? You want AI to achieve something in your research project.

Yes. But I you could give it a stupid goal. In the process you described, is there any way of the AI basically pushing you towards changing your goal or acknowledging that it's not well defined or maybe too broad? Whatever.

In the latest in Fable, which I only had access to for a couple of days, it did push back quite funny to me a couple of times. I think in general, we reinforcement learned those AIs into being like really, really submissive.

They rarely push back. I think in a general setup, if you don't tell the AI explicitly that, hey, at some point you may conclude that the goal is down and just reject the whole premise without explicitly saying that that is allowed, I don't think this would be an expected result. It would just keep trying and trying and trying. What do you think, Bruno?

Yeah, I think the I I think a big question there is all is how clear the goals are, right? Like in if you're using AI to do AI research and you have like very clear benchmark, very clear metrics and everything, it's very kind of objective. it's true that it's it's really impressive how modern, you know, the the most recent models are able to break down the goal into sub goals, iterate, get the best assumption and so on, right? It's very methodic methodical from that point.

I I I guess really what I'm what I'm kind of curious about in terms of like what's coming next is are we gonna are we going to have something as impressive when we move to more like subjective rewards, you know, like if you say, hey, just make a good book about this topic, you know, or something more make a nice song, you know. Is this because this is not really what these kind of long running agents are are really designed to do, right? That at the moment they're not all the real reinforcement learning is mostly on verifiable rewards.

do do you think there will relatively soon be a next generation of AI which will be able to do the same long running task on maybe more subjective goals, or do you think this is there's kind of a natural limit to what we can do there?

I

don't see any practical barriers to, for example, write a good book. Maybe it still applies that you would need to see that with something interesting. Cause if your prompt is just write a good book, then you would have exactly the same book as the person next to you with the same prompt. So you probably would want to put in some...

some entropy, some ideas, maybe it would be enough to just put in just one paragraph of how you want sci-fi unicorns on Mars or something. And there

Mm-hmm.

you go. Would it be a good book? I don't know, but technically, technically I don't see why it would not be possible to try like a methodical approach of writing a book.

Right? Outlining characters and interesting plots and like rendering it into pages of the book, reconciling that everything's cohesive, like changing the tone and so on. Would it be a good book in the end? I don't know. I don't know. What do you think?

Maybe connecting it to the point before about the goals, if the goal or the subjective input is so important in the overall assessment of then what comes what comes out of it, how could you use AI to basically get better at posing the right goal or providing the right subjective injection that is needed in order to make this good?

I think that's one of the toughest and hardest like unsolved things in AI right now, how to make the AI set himself in a goal, right? Because so far it's it's the aspect which is, I guess, the most dangerous aspect of AI, but it's also the most powerful aspect of AI, right? For example, the Arc-AGI 3 challenge where the agent should play

games like it's think of Tetris

Mm-hmm.

Tetris like this block games in really smaller resolution and there's like arcades arcade style games puzzles like sell games and so on and the goal is not explained and it's really hard to figure out what the goal is

And so far, as far as I know, nobody had solved it properly.

Interesting. I I guess my my experience I had with this benchmark is that humans would women would guess the goal because there's just something they find interesting, right? Like there's some natural desire to want to make I don't know the game look more like whatever the final solution is. whereas maybe maybe you could just you could argue machines don't care. They're not like interested, right? Like in either possible configuration.

Do do you think that's part of the part of the thing that's missing? Is this some sort of like theory of what humans would find interesting?

Maybe it's interest, maybe somewhere those puzzles certainly rely on some prior knowledge of how the world works, like gravity

Mm-hmm.

or mechanics of an arcade game. Some of the puzzles are really arcane and it takes you a few minutes to figure out what's going on, what do they want from you. You have to explore, click around. I guess in the process, what...

I was observing like what my mind does, trying to understand what the AI should do. And it's sort of also like a scientific loop you have from your priors, you have an assumption of what each element is. Then you also, when you look at it, you do like visual segmentation and you instantly assume, okay, this is probably the character. This is probably the door and I should move the character there.

Then you do an experiment,

Mm.

you press, you have very limited set of controls, you see you can press arrows, so you press an arrow and sometimes you surprise, boom, you were wrong. It's actually the thing that looked like a door is the character. Why you know this? Because you pressed right, it moved right, you made the connections with the results of experiment and also with your prior knowledge of 2D games.

From there, you readjust your hypothesis about what the goal is and what the end state is. Then you try to explore towards there. Notice how by just pressing one button, I instantly know what three other buttons do. If arrow

Mm-hmm.

right took me right, it's 95 % confidence I know about the three other arrows.

And it's surprisingly hard to build an AI which can do all of that in a loop.

Interesting. Yeah, it sounds a bit like the AI needs to understand the world in which the the the the the game evolved. Kind of need to yeah understand the laws of physics every time whereas it has possibly like too many assumptions about what they should be. So it doesn't deal well with out of distribution scenarios. Maybe that's the related imper related things. Maybe the related to this and to games actually.

In the in the in the context of the game industry, obviously AI is becoming very you know very powerful for the coding side of games. do you see AI becoming also a source of fun for video games? Do you do you can do you think AI is going to be able to generate interesting worlds, interesting puzzles that are unpredictable and surprising? Or so far would you say AI is still kind of the level where it

mostly would produce slop if you were using it on a creative side.

I think depends on how you use it, because there are definitely so many creative ways how you can apply AI here or there. then, AI in game development, very often people associate when talking about games and AI in games, not developers, but consumers often associated with asset creation, like 3D models, images for 2D games and things like that. And the community is just...

hating it, it's just not having it. Whatever game jams you try to join, it's always no AI, no AI assets, please only humans and so on. And everybody's so concerned about the wellbeing of artists and 3D modelers and how AI puts them out of the job. And for some reason, yeah, there is no such concern for engineers and

Mm-hmm.

you can use AI for coding your game as much as you want.

Mm-hmm.

but not to generate a picture of a button or a mushroom or something. And sometimes it gets quite ridiculous. For example, in a huge game last year, Arc Raiders, like a multiplayer online game, they used AI. And now on Steam, you have to disclose that you're using AI. So they disclose that they're using AI. And...

There was a huge backlash, like, how come you're taking jobs and so on? The way they used the AI is very smart. First of all, they hired voice actors, and they recorded a lot of lines from those voice actors. And then they trade the model, which would, in real time, snippets of speech when players...

ping some objects within the game world, like there is a rifle behind there, behind the corner, and you would get like the consistent voice acting, but then also this very specific, you you just cannot record voice snippets for

Yeah.

all of those situations. So to me, it's one of the more awesome and creative applications of the AI, where every time you ping anything in the world, you get this like custom

voice snippets for all players around. But yeah, it still got so much backlash and they were like, hey, we even paid, we even paid the voice actors and they agreed to this kind of deal.

Danger thing.

Maybe given the time, we already took up a lot of your time. just the last quick question. So you moved from the bit from the the world of atoms back to the world of bits. Are you still

Yeah.

also doing jewellery? Is this still something you do?

Yeah, as a hobby thing, I guess, as something to do when you need to relax and take a break from the computer because everything moves slower, I guess, at the jewelry table than at the computer where AI is pinging you left and right and everything has to move so, so, so fast.

Alsa Near, thanks for being part of it.

Examination.

Thank

you. Thank you for inviting me.

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'll be building towards.