00:00it's still quite hard to describe things
00:01with words designers yeah when they go
00:04and work when they do work for clients
00:05like it's one of their pet peeves
00:07because clients are like they don't like
00:09it but they can't explain why with every
00:11new technology some jobs are lost While
00:13others are gained and while people often
00:14focus on the farmer in this episode
00:16we're highlighting the latter a highly
00:18creative role that emerges alongside AI
00:20The Prompt engineer until AI can close
00:23the loop of its own each tool still
00:25requires a set of props and just like a
00:27composer feeds an instrument a set of
00:29notes to play a prompt engineer feeds
00:31the AI a map of what to produce and if
00:34we know anything from music it's like
00:35composing great music takes great skill
00:37so in this episode we dive into the
00:39emerging importance of prompting the
00:41early learnings and how to do it
00:42effectively and also where this field
00:44might be heading and we do so with guy
00:46persons guy has been an early mover on
00:48the text image AI space having written
00:50the dolly to prompt book in July of last
00:52year so will the prompt engineer be more
00:54like the highly sought after devops
00:56engineer or a proficiency like Excel
00:58that you find on every resume listen in
01:00to hear guys take by the way we're
01:02thinking of running a prompt competition
01:04coming up so if you think you have what
01:05it takes email us at podpitches at
01:07a16c.com but the subject prompt engineer
01:11as a reminder the content here is for
01:13informational purposes only should not
01:15be taken as legal business tax or
01:17investment advice or be used to evaluate
01:19any investment or security and is not
01:21directed at any investors or potential
01:23investors in any ACC fund for more
01:26details please see acc.com disclosures
01:38guy welcome to the show thank you for
01:40having me I'm excited to be here let's
01:42take a step back when we originally
01:43reached out to you it was around 6
01:45months ago and you'd just written
01:47something called your prompt book why
01:49don't you give everyone a little bit of
01:51an idea of what that prompt book was
01:53what it is now and also what prompted
01:56you to want to write it in the first
01:57place this was in the initial Heyday of
02:01Gali 2 which was like open ai's text
02:04image model they rolled it out to a few
02:07test people at a time they were super
02:10cautious about how it might be misused
02:13how it kind of could end up kind of
02:15having a backlash all these kinds of
02:16things which then only increased the
02:19sense of people wanting to get their
02:20hands on this thing because at the time
02:22this was pretty things you might think
02:24of now a stable perfusion mid-journey
02:27kind of predated those by some small
02:29margin and seemed way ahead of anything
02:31people were kind of tried using before
02:33if you've used the text to image AI by
02:36now you know it's basically a text box
02:38and it all comes down to what you type
02:40in it doesn't have you know buttons and
02:42in all the kind of controls you might
02:44expect when you write logins like
02:46Photoshop so the question then becomes
02:48like a lot of people once mine goes
02:51blank or you don't actually know the
02:52name or the words of like what you're
02:54trying to type in right so if you've
02:57actually been to art school or you're up
02:59on your art history or in your design
03:00language then you've probably got a head
03:02start on everyone else but on places
03:04like Twitter and Reddit there are people
03:06posting like these amazing images but
03:07because of the nature of social media
03:09it's all lost so I started trying to
03:11like collect these cool examples and
03:13these cool kind of terms people are
03:14using to create these like amazing
03:15visual effects so I started putting
03:17everything in a essentially like a slide
03:19deck by the time I'd copied and pasted
03:21all these cool things I've seen there as
03:2380 100 slides long something like that
03:25so then I rather Grand me called it a
03:28book and shared it shared it online and
03:31it's just a jumping off point for people
03:32to realize the kind of stuff at the time
03:34that these were just about becoming
03:36capable of obviously now they're
03:41advanced definitely and we'll get into
03:43that because within six months it's
03:46crazy to see how these tools the way
03:48people are using these tools how that's
03:49all changed in a matter of again just
03:52six months it feels like yesterday when
03:54we didn't even have access to this but
03:56this idea that these are tools and just
03:59person a versus person B may not get the
04:03same result they may not have the same
04:04understanding of how to Leverage The
04:07Tool and so before we get into maybe the
04:09tips and tricks that you've learned I
04:11just want to give the audience a broad
04:13sense of how much time you've spent
04:15within the bowels of mid-journey Dolly
04:19stable diffusion like if you could give
04:21an estimate how much time do you think
04:23you've spent kind of mastering this idea
04:26of prompting I wouldn't say I'm a
04:29in in any sense it's like so engaging
04:33and interesting to like experiment with
04:35these tools so you know like in the last
04:37six months for sure like a couple of
04:39hundred hours well I really admire is
04:41people that are using these tools to
04:44create this kind of like real body of
04:46work where they kind of they really like
04:48pursue a direction to like discover
04:50what's possible I think I saw a thread
04:52where you know I think it's a
04:54mid-journey you can get it to tell you
04:55how many prompts you've ever done and
04:58there are people in the thousands
04:59hundreds of thousands yeah yeah and I
05:01appreciate how humble you are
05:04you are as we're again we're six months
05:08when there's a new coding language and
05:10then you see people write job
05:11descriptions for developers looking for
05:13someone with five years experience when
05:15that particular language has only been
05:16around for six months or a year and so
05:19yes I don't think anyone could
05:20definitively say they're an expert in
05:22prompt engineering partially because
05:24it's only been around for so long but I
05:26do think you've at least shared a lot
05:28more than you know the average person in
05:30this space and given your experience
05:32with these tools I'm curious if you see
05:35a parallel skill set where you can kind
05:38of compare prompt engineering to is it
05:40similar to learning to code is it
05:42similar to being able to storytell
05:44effectively is it similar to being able
05:46to process numbers in an Excel sheet
05:50like is there a parallel skill set where
05:52it reminds you of you know something
05:55you've done before I think that was an
05:57area I don't know if we're still in it
05:58where there's a certain category of
06:00person who could consider themselves
06:02like good at Googling stuff
06:04do you know that there's this thing
06:06where you can type in like file type
06:08this and find this thing there's this
06:10big debate over whether like especially
06:12in text to image you know is it really
06:13is there really like any Artistry to it
06:15and I for me like I'm not so sure
06:17because I'm I'm no artist but there's
06:19definitely something it's almost about
06:20it's like discovering an image that's
06:22already out there you've just got to
06:24find the words that summon it forth as
06:26if you're kind of navigating like an
06:28infinite Pinterest of things that
06:30haven't quite existed to be manifested
06:33them well I mean to that point like we
06:35have so much information online I feel
06:36like that is a skill set even before
06:38these AI tools like I used to work on a
06:41product called Trends and that really
06:42was about using the right tools like
06:44subreddit stats or ahrefs or other data
06:47sets online and learning to parse them
06:49and learning to surface what other
06:51people find interesting
06:53um but let's get into the nitty-gritty
06:55like you wrote this prompt book you've
06:56been playing around with these tools for
06:57quite some time are there certain
07:00learnings maybe the 80 20 approach of
07:02becoming a good prompt and engineer in
07:05terms of things that you think are are
07:07really valuable to understand maybe it's
07:09the prompt length maybe it's using
07:11certain modifiers within your prompt
07:13maybe it's just like a framework for
07:16thinking about prompting is there
07:17anything that's surface that you think
07:18would be really valuable to someone
07:20who's just starting out with prompting
07:22oh yeah like I think if you've never
07:23used one before like the best way to
07:27explain how they work is to always like
07:30describe something as if it already
07:31exists imagine that it's an image in
07:35some kind of downloadable clip art
07:37library or it's it's on it's on a
07:40photography gallery and you know
07:42someone's written underneath oh this is
07:45a fine example of a modern uh
07:48photography shot and those are the kind
07:50of descriptions that you're trying to
07:51kind of mimic to tell these kind of
07:54tools what you're looking for and it
07:56also gives it like a natural sense of
07:58why these tools are bad at some things
08:01and and the kind of prompts that don't
08:02really work because you never you know
08:04know if there's like a let's say some
08:06like an archive image of some women
08:09celebrating on the steps of a church in
08:131972 it will have that kind of caption
08:17but they never go the woman on the left
08:20is wearing a yellow hat the woman on the
08:23right is wearing you know they just
08:25don't go into that kind because you can
08:26see it so ironically they often describe
08:29that generally what the image is about
08:30but not like how you would draw it step
08:32by step and that's why these tools are
08:35less good at saying like I want this
08:37thing over here and then that thing next
08:38to it and then something on top and that
08:40thing should be much bigger because
08:41that's in the real life that's not how
08:43images are described in language so
08:45that's a really good point though
08:46because that's how these AIS were
08:48trained right so I think Dali trained on
08:49600 plus million images and they're
08:52using that alt text from online of
08:54saying again that descriptor and I've
08:56never thought about it that way but
08:57actually training yourself to become a
08:59good prompter by reviewing the inputs to
09:04the tool which I I've never done this
09:06before but I can imagine someone
09:08literally going online and reading the
09:09alt text on different images and going
09:11ah this is how this was described this
09:13is how an AI might interpret my future
09:15prompt and I think to your point also is
09:17something that I've learned from my very
09:18limited set of prompting is just the
09:20level of detail that you need with your
09:23prompt where when I first started I'm
09:25like you know monkey wearing a hat you
09:27don't even realize until you start
09:29prompting the many iterations that could
09:31come from that like you have one image
09:33in your head but then you get back all
09:36of these different results and then you
09:37you end up looking on different prompt
09:39search engines or libraries and seeing
09:42what other people are doing you're like
09:44this prompt is like 200 words like I
09:46would have never thought to do that for
09:49now yeah and I think there's something
09:50to be said like I think the longer they
09:53are the the less there's definitely like
09:55diminishing returns as I went through
09:57your prompt book there were so many
09:58different ways that you could describe a
10:00shot you could say a different camera
10:01angle you could say a time period as you
10:05just spoke to you could say a specific
10:07type of artistry or or even a specific
10:09artist I know there's some controversy
10:11around using specific artists work to
10:15train your new images but let's look
10:17forward to today I feel like as we
10:19talked about six months later these
10:21tools have evolved a lot are there any
10:23different ways that you can prompt today
10:26or kind of Leverage these tools that
10:29didn't exist six months ago that are
10:31really important and maybe extending the
10:33way that you can use them 100 so the
10:37main one and these things are like
10:39changing all the time right but now
10:40there's increasingly tools where you can
10:42prompt like with an image again that's
10:45almost like an entire New Field of
10:48exploration because it's not combining
10:51the image with the other with like your
10:54words in the way you would expect
10:56something like Photoshop to do it like
10:58it's not collaging them together so the
11:01results can be really surprising really
11:03unexpected really quite difficult to
11:05control but then you potentially have
11:07interesting opportunities like okay I
11:09can make a load you're kind of abstract
11:11kind of stuff using like my brand colors
11:15or something that's important to me
11:16photos of me who knows and then yeah and
11:18then I'm gonna use that and kind of
11:20multiply that Baseline that kind of
11:22visual base with custom other prompts
11:25and then everything will have this kind
11:27of lightness and then of course like the
11:29big thing that happened since the days
11:31of the prompt book and so on was of
11:33course that huge spike in interest in um
11:36selfies right like the lenses and the
11:40profile picture.ai and there are like a
11:42dozen of them which was just prompting
11:43with your face basically
11:46um prompting with 20 selfies and being
11:48like yeah I want to see more this guy
11:49because it's me obviously and then
11:51within the image to image
11:53space you've now got other startups that
11:55are doing interesting things where like
11:57okay give us 10 core images and that
12:00will generate you like infinite versions
12:01of that based on like the modifiers that
12:03you want to see so there's all kinds so
12:05that's a really interesting space that's
12:06gonna probably power like the next
12:08generation of how people especially
12:10consumers like interact with these
12:12products yeah one way that maybe you
12:15could put it is that when we first got
12:17access to these tools you were really
12:18starting from scratch you didn't even
12:19have the prompt libraries available to
12:22you you were just like okay I have this
12:23image in my head but today you not only
12:25have those libraries you also have
12:27images that you can input so you're not
12:29starting from scratch you have a
12:30baseline of as you said maybe it's brand
12:32colors maybe it's a certain style and
12:34it's instead of having to articulate
12:36that yourself you can just say hey
12:38here's what I want but to your point
12:41sometimes it's hard to control right
12:43because you're you're trying to say
12:45something to the AI you're trying to say
12:47I want this output you don't always get
12:49it and so something I want to ask you
12:51about is how you've learned to Reign
12:54that in to really get a higher
12:56throughput of images that you want
12:58versus images you don't want over time
13:00because something that's interesting
13:02about these AIS is that they are a
13:04little bit of a black box you can't
13:06understand every little piece that went
13:10from your input to your output and so
13:12you can't like find two in it in the
13:14same ways as maybe some other skills
13:16that we've learned in the past and so
13:17how have you learned to actually become
13:19a better prompt engineer given that
13:21black box nature I mean I think another
13:23aspect is that you and I both type in
13:25the same thing it's not going to make
13:28the same picture because it kind of
13:30starts from this random cloud of noise
13:32and your cloud of noise is different to
13:34mine and then it's slowly turning these
13:37clouds more and more into something that
13:38looks like uh orangutan in a tuxedo so
13:42that's really frustrating when you're
13:44like testing things because was it good
13:47or did you just get lucky or
13:49alternatively if you're not seeing what
13:50you expected should you just hit it
13:52again and again like when you see
13:54someone else has made something really
13:55cool did they do something really clever
13:57or did they just you know is it like a
13:58persistence I have found myself in that
14:02I have an idea for what I want it's not
14:05something that is super important where
14:07I need to nail it so I'm just I just
14:09need it close enough and I'm getting
14:11these results and they're getting a
14:12little closer and closer and closer but
14:14I have found myself in that spot where
14:16I'm just like let's just generate it
14:18again like if I do this enough times
14:19I'll eventually get to something that's
14:22workable so do you have any thoughts
14:23there in terms of like
14:25how you don't end up in that spot where
14:27you're just like hoping
14:29for a better image you're kind of like
14:31pulling the AI slot machine if you will
14:33I no I mean I think unless you've seen
14:36unless you kind of have evidence and I
14:39there will be more and I think it's why
14:42some of these like other tools and
14:44communities are so important you know
14:45where you see lots of other people's
14:48is you know if you can see someone else
14:49has done it then you're like okay I I
14:52ideally you can also see the prompt they
14:54use and work out how they did it but
14:56even if not then you're like okay I can
14:57get there also you run into these things
14:59where you always think it's like the
15:02most simple thing and then you're like
15:03it doesn't know what a hot dog is like
15:06it just doesn't understand the rules
15:11you know physically what can and can't
15:13that look like and you're like trying
15:15and it's like now the sausage is a right
15:18angle and the the bun has ears because
15:21it's starting to throw in some like
15:22fashioned like you know Aesthetics and
15:26then you're like and minus minus no no
15:28dashing that's kind of the limitation of
15:31is at the moment which is it's amazing
15:34until you're trying to do something very
15:37specific and especially if you want to
15:39do something very specific just also to
15:41like a very high like or professional
15:44standard well I'm glad you even
15:45mentioned the negative queries that's
15:47something I think a lot of people don't
15:48know is that you can you can say hey AI
15:51I don't want this it doesn't always
15:52manage to still generate what you're
15:55looking for but there's also almost like
15:56these glitches one of them that is kind
15:59of Infamous now is hands right so you
16:01can generate these beautiful images of
16:03these Instagram looking models and you
16:07can put them in all these different
16:08backgrounds and you're like wow this is
16:10amazing and then it's always like well
16:12look at the hands which is kind of funny
16:13I feel like it's a it's like the perfect
16:15manifestation of how technology always
16:17is like much better in One Direction
16:19when it's invented but there's always
16:22like these things that need to be
16:23iterated on and so are there other
16:25things worth knowing about whether it's
16:27these negative props whether it's these
16:31glitches that are still in the Matrix
16:33what would you call out from your again
16:34many hours of being deep in in these
16:36tools I think it depends on the model
16:39one example was when dally came out it's
16:41not very good at understanding that it's
16:42drawing things in a square so if if
16:45you're drawing a person it's often going
16:47to have like its feet and its head cut
16:48off because it's it's seeing those
16:50important photos but one thing you could
16:52do with Nelly is you can actually upload
16:54like an image to like do variations of
16:57and if you upload an image that's just
16:58like a little white border
17:01um then it knows that nothing can go
17:02there and that that kind of encourages
17:05it forces it to kind of think inside the
17:07box if you will but then of course you
17:08have now tools like mid-journey you've
17:10been like it iterating on their text to
17:13image model like a lot more aggressively
17:17understandably I think maybe have some
17:18other things in there in the cooker you
17:21know which have now bring that into the
17:23model itself so when you type things in
17:25it knows it's a square and actually it
17:27will sometimes do quite clever things in
17:29order to fit it in that space so if you
17:31ask for kind of like a group selfie of
17:33three people you know on something like
17:35dally that's going to be cut off at the
17:37end because you're used to seeing
17:39someone taking like a disposable camera
17:42photo whereas imagine is clever enough
17:44to know that one of them kind of needs
17:46to be standing behind the other or like
17:47leaning in from the side so it's kind of
17:49clever how they've managed to like solve
17:51that composition problem
17:52um within the AI which then either you
17:55know the prompt engineering thing I
17:56think is just understanding the
17:58possibilities and the limitations of
17:59where you are at the moment meanwhile
18:01there's these other people who are doing
18:04some like very technically serious work
18:05to kind of make those limitations kind
18:08of no longer uh relevant yeah well I'm
18:11glad you brought up the differences
18:12between in these different tools so if
18:14we talk about just stable diffusion
18:16mid-journey and Dolly I feel like there
18:17are three that a lot of people are
18:19you liken the ability to prompt within
18:22each of these more like the difference
18:25between Excel and Google Sheets where if
18:28you know how to use Excel you really can
18:30drop right into Google Sheets and it's
18:32relatively straightforward you might
18:33have to switch up your shortcuts a
18:36little bit or learn one little thing
18:37here and there but for the most part you
18:40can again drop from one to the other or
18:42would you liken them more to learning to
18:45speak different languages it's not that
18:47different I think the principles are
18:49like very similar and then the nuances
18:52are slightly different so I think now if
18:56you went from like like daily to Mid
18:58Journey it would be like amazing and
19:00then if you went back into the direction
19:01you'd be like it doesn't do what I want
19:04but that's because mid journey is doing
19:05so much of the heavy lifting to help you
19:07make something really good if you are
19:09using the tools to create some like very
19:12specific effect kind of like imagine
19:14that I guess yeah like a very
19:15complicated Excel formula that would not
19:18have the exact same output in the oven
19:20at all because they're trained on like a
19:22different set of images stable diffusion
19:25I think some five billion and then like
19:28a smaller set of like 12 million for the
19:30what does nice look like fine tuning
19:32that's happened on the top and how
19:33they've optimized it you know they they
19:34have made different like creative
19:36decisions there it's maybe like driving
19:37a different car like if you if you like
19:40floor the accelerator in different and
19:42then do this in various different cars
19:44some are gonna take off so I'm gonna
19:46trundle along so good analogy do you
19:49also find that I mean we've talked
19:50already about this idea where sometimes
19:54but then but that final 20 the real
19:57refinement to get to exactly what you
19:59pictured in your head or exactly what
20:01you want and didn't picture in your head
20:03sometimes requires another tool and so
20:06have you found I've heard some people
20:07are using facetune or different AIS to
20:10take it to the final level or I guess
20:13you could also use in painting and out
20:14painting a little more discreetly so how
20:16have you found the relationship of maybe
20:18one tool to the suite of other tools
20:21that exist out there I think it has lots
20:24exciting crossovers but actually I kind
20:27of think it's a big opportunity for like
20:29the photoshops of this world because
20:30those are tools that presuppose you have
20:32some kind of original image to be to
20:34like manipulate whereas now there's at
20:36that there's a huge amount of that raw
20:38but maybe not perfect material there for
20:40people to work with there's lots of
20:41things also that I've been trying to
20:42like do imprompting that are actually
20:45like more easily achieved in other tools
20:47so you can you know spend ages trying to
20:49get this kind of Vintage Film look but
20:51if you're like an Instagram influencer
20:52you'll know which I'm sure you are um
20:55who isn't um but there's loads of
20:57there's loads of iPhone apps right that
20:59are out there just to like give all your
21:00photos that kind of like dreamy
21:03you know Vintage Film look yeah I mean I
21:06think back in July when you first wrote
21:10quested featureless for dolly two and I
21:13some of those have been implemented I
21:15haven't cross-referenced them but are
21:17there things that are on your new list
21:20of hey these tools are great but they're
21:24missing XYZ or they're lacking in these
21:27areas this would be top of my list to
21:30see improved on you know I think we're
21:33going to see like more models come out
21:35and I mean the fact that stable
21:36diffusion is kind of Open Source means
21:39that lots of other things are going to
21:41be built on top of that and I think it's
21:42gonna be really exciting to see some of
21:45people take that in either kind of on a
21:48individual sort of proceamer level
21:50people building their own models to
21:52create their own stuff more likely some
21:54bigger organizations you know training
21:57it for like specific purposes the whole
21:59challenge the whole opportunity I think
22:01was like how do you go beyond the text
22:03box how do you go beyond this like just
22:05blank rectangle to create something
22:09that is more user friendly that's more
22:12inspiring that's more how people think
22:15because on the one hand if you're if
22:18you're not an artist the ability to
22:20describe things with words is definitely
22:23but if you think about the next layer is
22:26it still quite hard to describe words
22:28designers yeah when they go and work
22:30when they do work for clients like it's
22:31one of their pet peeves because clients
22:33are like they don't like it but they
22:35can't explain why or what they want
22:37different they're like oh I want it to
22:40do you know what I mean like more and
22:42they're like I don't know I don't know
22:43what that means which is basically the
22:45position these you know AI models are in
22:47so I think you're going to see could you
22:49see like a conversational interface can
22:51can you do the generations fast enough
22:52that you're always showing people
22:54multiple options possible new directions
22:56almost like in a sort of
22:59multi-dimensional space where it's like
23:01do you want to take it more this way or
23:02more this way part of the prompt book is
23:04I didn't know what you know metaphysical
23:08Kodachrome all these things were but
23:10those at least have names you know once
23:12you know the name you're like oh that
23:13that thing okay it's cool I didn't know
23:15it was called that but there's probably
23:16other Aesthetics right other styles that
23:19we don't have were actually words for
23:21it's like you know that kind of gritty
23:23but like modern gritty like almost like
23:25shiny gritty like the grit has a shine
23:27on it and probably I could make you a
23:29mood board of that and you'd be like oh
23:31yeah like that's a thing but there's no
23:32word for it if you can create ways of
23:36unleashing the inexplicable the
23:39undefinable uh um which is what in the
23:43day like visual art that's the exciting
23:45thing about visual art is to express
23:46things or moods or things that you can't
23:49quite put into words you totally have my
23:51mind spinning thinking of different
23:53ideas a couple of them that came to mind
23:55one of them is just a better onboarding
23:57experience but one where you're guiding
24:00the new prompter to understand how all
24:03these things might fit together to your
24:04point like try this oh look at what you
24:06got here oh did you notice how when you
24:08use these two prompts together this one
24:10kind of overshadows the other maybe
24:12there's a third word that's a synonym of
24:14this and I think you've kind of done
24:16this on your own by just going through
24:19and prompting like crazy going through
24:20these different prompt libraries and
24:22trying to sort through the signal from
24:24the noise but I do think any one of
24:27these models or maybe something built on
24:28top of them could have just a much
24:30better onboarding experience and then I
24:33also in terms of these visual Styles I
24:35mean it reminds me of you mentioned
24:36Instagram a lot of Instagram influencers
24:39for a period of time we're selling these
24:42filters because they had figured out the
24:44precise tuning of every little variable
24:47which sounds easy but I had tried to do
24:50it myself I I never managed to create
24:53good Lightroom filters but people had
24:55and they they would sell them and so I
24:56wonder if you'll see the same thing
24:57where maybe someone creates a zip file
25:00of a mood board and then they train the
25:03AI in some way that does make it I guess
25:07play nice with that particular concept
25:09that you can't distill necessarily into
25:11a single term someone did a paper on
25:13that I think it's almost kind of what
25:15led to that selfie craze which was you
25:17don't need to put photos of Staff in
25:19that original 600 million
25:21training data or wait for the next time
25:23we do that again for it to teach it what
25:26you look like there's this kind of um
25:28embedding trick where you can show it
25:30like a bunch of photos of you and then
25:32you can refer to you as and it knows how
25:35to kind of recreate that and there was
25:36also an interesting thing in the same
25:38paper that hasn't really been used or
25:40like commercialized in the same way
25:41which is to do that with style so rather
25:44than show it yeah like this is what this
25:46person looks like it's like this is what
25:47the style of blah blah is called here it
25:49is and then off you go which obviously
25:51has all kinds of potentially Shady legal
25:54um but let's assume this is a lovely art
25:56we've made um ourselves to the idea of
25:59honing in a style I do wish there was a
26:02version of the product where I could go
26:05and like we've talked about maybe upload
26:07certain brand images or certain brand
26:09colors and then have it iterate with me
26:11where it shows me a bunch of images and
26:13I say um it's okay but I want a little
26:15more of this color I want it X Y or Z
26:17and then we keep doing that to the point
26:20where I get get a bunch of images where
26:22I'm like yes this is the style you can
26:24lock that in you lock it into a variable
26:25that you can then plug into future
26:27prompts there's some people out there
26:29that have like some people that have
26:31been like sharing interesting themselves
26:32they've like managed to like lock in
26:33like a particular look and now like
26:36every blog post they have on their
26:37lovely you know content marketing thing
26:39is like oh there's always the same kind
26:41of thing and that's like pretty cool but
26:43we haven't seen that always built into
26:45the like Foundation models yeah as like
26:48a way of interacting with it and then
26:49there are some a couple like a few
26:51startups like um scenario which is doing
26:54it for late game assets and then like
26:55Leonardo which is like more
26:57multi-purpose I think or is just
26:59positioning itself that way which yeah
27:01is again all about can you like control
27:03things down to like consistent look what
27:06we've talked about so far is this idea
27:08of controlling the AI but I also like to
27:10think about the ways that when you work
27:12with these different models you learn
27:14more about whether it's art learn more
27:16about your own creativity or how you
27:19think about these these images so the
27:23example that it reminds me of is in
27:25chess when we finally built the Bots
27:27that were better than humans and chess
27:29not only were we surprised by the fact
27:31that that could happen but we were also
27:33surprised by all of the different
27:35openings or moves that humans in their
27:38thousands of years playing chess had
27:40never considered that were better than
27:43some of the moves that even the best
27:45chess players in the world had used and
27:47so have you seen any of that even from a
27:49personal experience level like where
27:52you're in these tools and you're playing
27:53around and you again you're learning
27:56with the model if that makes sense it's
27:58almost surfacing things that you had
28:00never considered before I like that I
28:02think I think what you're attaching is
28:03whenever you're using these tools you
28:05have these two modes right where you're
28:09either like waiting to see what it shows
28:11you or you kind of are visualizing it in
28:13your mind and you're like no not that
28:15not that but if you just let it take you
28:17where it wants to go then you're
28:18suddenly like I have no idea what I'm
28:21looking at but apparently I'm here you
28:23can do all kinds of within Dali there's
28:25like this variations tool so you just
28:28let's show an image and it'll be like
28:30here's four more that are kind of the
28:31same but obviously over time it starts
28:34to if you leap and leap and leap and
28:35leap you end up on this like completely
28:38bizarre visual Journey it's very it's
28:40like it's like yeah like a psychedelic
28:42dream it's fun to play around in these
28:44tools but ultimately there well there is
28:47a market for just interesting art in the
28:51world a lot of this will need to ladder
28:54back into you know whether it's blog
28:55post sharing images whether it's
28:57creating you know the next sneaker
28:59design that you end up selling are there
29:02areas that you've seen really emerge
29:04from this where people are using these
29:07tools today and applying them to again
29:09what someone might call a practical use
29:11case and in addition to maybe what
29:14you've seen so far are there other areas
29:15where you're excited to see this be
29:17applied sure it's interesting though
29:19isn't it because I think especially
29:21given the tenor of the conversation
29:26um around these tools and the ethical
29:30all and legal aspects they're in I
29:33suspect that when you to an extent when
29:36you see these things used especially in
29:39prominent context they might not be
29:40advertised as such much as like green
29:42screen right like when green screen is
29:43used and films you shouldn't be like
29:45that is an amazing use of green screen
29:47you should just be like oh my God like
29:49he's dangling off uh a thing
29:51this must have cost Millions so I think
29:55you know when we see AI tools used in
29:58lots of context not this is covered up
29:59but you know they might also be just a
30:01narrow part of the creative process like
30:03raised this point online I think that
30:05you were making which is like well
30:06where's this all going like will it ever
30:07make images good enough and like do with
30:10other people like want to look at them
30:12because it's not like we have this huge
30:13history of like logging in to social
30:15media and looking at just like abstract
30:17pictures like oh horse owner
30:20um on a surfboard I mean things tend to
30:23have like a grounding in reality right
30:25like that's what makes them viral or
30:26interesting it's like a real thing that
30:28happened but then someone was like no
30:29like maybe this it won't be that it's
30:31going to make content so good that it's
30:34like better than Netflix or like better
30:36than Instagram or better than it's like
30:38it's the hobby of doing it like that's
30:40that that's the entertainment well I
30:41mean there are skills out there to your
30:43point where writing as an example some
30:45people just like the right to write and
30:47whether other people read it doesn't
30:49matter they actually enjoy the process
30:51and so I definitely could see an
30:52entertainment angle but I do think that
30:56to an extent you know even with writing
30:58as an example A lot of people really
30:59hate writing and a lot of people find
31:01Value in the money that they get paid to
31:04um or the writing is used within a
31:06script which then is published on
31:08um and so it's like how is this stuff
31:10used in The Wider World whether it's you
31:13know on an e-commerce website whether
31:15it's actually you know one day
31:17integrating with 3D printing and like
31:19the stuff that you generate in
31:20mid-journey then can actually be printed
31:22into like a real life product that you
31:24sell this isn't just a gimmick this
31:26isn't just a toy there's this very high
31:29level debate around kind of artistry I
31:32suppose and as if everything is either
31:35going to be like in the Louvre or saying
31:39that right in the Tate I'm from London
31:41or you know or or in the bin But
31:44ultimately if you look around just this
31:47any space that you're in and look at
31:49everything that has like a visual
31:51component or like a design comparison
31:53there's so many different levels at
31:57which we engage with
31:59are you know like the pattern on a
32:01cushion the the warning label on the
32:05coffee maker the sausage dog on a on a
32:09they're all different things and so
32:11there's some things where human the
32:13human touch is like literally the point
32:16but other things it's like a soothing
32:18pattern to look at to say that your wall
32:20isn't just gray and so there's all kinds
32:22of layers in between and I think we'll
32:23see them used more and more so the final
32:25thing I want to ask you about is how
32:27this all fits into you know the wider
32:29skill set that people might have so on
32:33one hand I can see how there might be an
32:35argument that this idea of the prompt
32:37engineer is going to be one that only
32:39few can do really well right so people
32:41are really going to master this skill
32:43set and they're going to be much more
32:44valuable than the people who don't know
32:46how to prompt well but then I can also
32:48see an argument where as you said maybe
32:50this gets abstracted and we have great
32:52uis where truly it becomes the type of
32:54thing where basically anyone can do it
32:56and anyone can do it pretty reasonably
32:58well and it just becomes you know
33:00similar to being able to write and read
33:03these are just kind of fundamental
33:05Elemental skills that are in everyone's
33:07skill sets they're taught in schools
33:09where do you sit with that in terms of
33:11how you see this progressing like is it
33:12worthwhile you could also position the
33:15question as to become an excellent top
33:18one percent prompt engineer or is it
33:20like oh everyone should kind of have
33:22this in their toolbox well that depends
33:24I think on the one hand there's
33:27obviously every incentive for the people
33:31that make these foundational tools to
33:34make from engineering for instance not a
33:38thing because they want everyone to be
33:39able to do it right again if you look at
33:41the most recent model of mid Journey
33:42like version four if you like stuff that
33:46would not have been even possible six
33:47months ago you can literally do the
33:49thing where you type in I can't remember
33:50your example but like I remember because
33:52I posted one that someone was arguing
33:54about it and I was like look at this
33:56space stuff I just typed in space duck
33:58and it's like this amazing astronaut
33:59Doug and he's like there's no way you
34:01just type that in so I went back and
34:02checked and I was like no that I lied I
34:04actually typed in a really cool space
34:06duck but at the same time like with any
34:08material like artistic or otherwise if
34:11you push things to the boundary there's
34:12always going to be someone that like
34:13explores everything that's possible or
34:16like just iterates and iterates iterates
34:18on something you know they're obviously
34:20gonna explore further on the map of
34:22what's possible than someone that isn't
34:24so I don't think it will become like
34:25this necessary skill that everyone needs
34:27to have but I do think it will become
34:29you know like some people that expert
34:31wood whitlers or you know really good at
34:34animating hair or whatever you know the
34:36people that develop a real like passion
34:39for whether that's just you know
34:40creatively or out of curiosity to see
34:43what's possible you know we'll we'll be
34:45there we'll like do some of the most
34:47amazing things and then there's also the
34:50the secret prompting I guess like a
34:52copywriting thing would be like the
34:53obvious example at the moment you think
34:55you're typing something into a ux but
34:58there's something else wrapping that in
35:00a prompt and then sending it to like a
35:02foundational model so there's probably
35:04going to be some people whose job is to
35:06like come up with that layer of thing
35:09that the consumer or the average person
35:11is never seeing and they think they're
35:14just talking to the AI but really
35:15they're talking to this thing that then
35:18adds a little bit of juice to it and
35:21then tells the AI that this is going to
35:22be a tangent but it reminds me of I just
35:24listened to a reply L episode where
35:27they're trying to someone had remembered
35:29this song from his childhood and they
35:30were trying to figure out what it was
35:32you've heard this episode of people
35:33having it it's the only one but it was
35:39it reminds me of do you remember in the
35:42episode they find this lady who is a
35:45music producer but she is a music
35:46producer for specifically people who
35:49want to create music like The Barenaked
35:51Ladies yes it's like you know people
35:54have jobs like this when you grow up and
35:56you're in school and they tell you you
35:57know you could be a doctor one day you
35:58could be a teacher one day they don't
36:00tell you you could be a music producer
36:01for musicians that want to sound like
36:03the Bare Naked Ladies and it makes me
36:05wonder or think about you know what
36:07specific niches are people going to go
36:09into within this realm of problem
36:11engineering right like maybe you
36:12specialize as you said in in hair maybe
36:14in hands maybe in something for like
36:17Enterprise SAS companies I don't know um
36:19it's kind of hard to predict at this
36:21point since we're so early but yeah I
36:23think you're right that
36:26I guess kind of a bimodal nature to it
36:29it does seem like the kind of tool
36:30that's going to be on everyone's desktop
36:33but it does also seem like there is this
36:35opportunity to become as someone might
36:37say like a 10x prompt engineer I think
36:39that's interesting isn't it because
36:40that's such a tech World metaphor like
36:43the notion of 10x and like like because
36:45it even implies there's a scale where
36:46you can have one and therefore you can
36:48have ten of it which like in the in the
36:51record industry do people talk about
36:53being like a 10x Mac recording engineer
36:56like obviously some recording Engineers
36:58are like famous and like better than
37:00others and there's all this kind of
37:01talent but I don't know if people are
37:03like yeah like I'm a 10x but yeah just
37:06like producers and all the kind of
37:07people that go into making I think music
37:09or film you know that huge list of
37:13people you see at the end of every movie
37:14and where you discover a whole new world
37:16of careers that you might have had I'll
37:19unfortunately never be a best boy but
37:20I'm still hoping to be a gaffer then you
37:22know there'll be all those kinds of jobs
37:23I think in the AI the creative AI
37:26industry you know your point on the
37:28spectrum of like what is 1x and what is
37:3010x it reminds me of what is the most
37:33popular piece of you could say art or
37:37imagery that is shared online
37:39like what comes to mind for you there
37:43I don't know I feel like do you I feel
37:45you said that with like as if you you
37:46know the answer well I have an answer I
37:51want photos of parties so
37:58for me at least someone who spends a lot
38:00of time on Twitter is memes and memes
38:04are like the most basic kind of imagery
38:07ever it's like literally an image with
38:09like some some capitalized text on it
38:12and so your point just reminded me of
38:14this idea where art especially is
38:18subjective and what people like and
38:20resonate with is not necessarily the
38:22most refined or extravagant precise type
38:26of imagery which you can generate in
38:28some of these text to image tools but it
38:30doesn't necessarily mean that people are
38:32going to resonate with it exactly
38:34could you be attempt I mean until they
38:37invent an AI that can do 10x memes which
38:39is the last thing we need
38:44I love hearing about how
38:49definitely share the prompt book Link in
38:51the show notes because I think people
38:52can benefit from seeing the different
38:54types of modifiers that you can include
38:56in a prompt and also a link to your
38:59social because you're constantly sharing
39:01new hacks new things that you're
39:02discovering but yeah any any other
39:04places that people should look to find
39:06you or your work uh you can find me on
39:08Twitter at guy P guip
39:12um and you can find my substance when I
39:15finally post and promptresponse Dot
39:17substack.com awesome well thanks for
39:23thanks for listening to the a16z podcast
39:25if you like this episode don't forget to
39:27subscribe here on YouTube to get our
39:28exclusive video content we'll see you
![How To Succeed As A NEW & YOUNG Realtor [Deals Every Month + Luxury Listings]](/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FrtrRghgTayc%2Fhqdefault.jpg&w=3840&q=75)
