Put the Steering Wheel Back On

0:00 Paul Swift

for anything consequential, always have an air gap. meaning nothing

0:03 Sam Rogers

Yes.

0:03 Paul Swift

leaves nothing is sent by an agent or anyone if it's serious. it can only be generating and producing things.

But it but don't give it the send button unless there's a reverse button

0:25 Sam Rogers

Welcome to Signals and Subtractions. I'm your host, Sam Rogers. Every week, everyone at this table brings two things. One signal worth watching, one subtraction worth making. That's the whole show. What to watch, what to drop. And today's question When AI does the making, who's the client? Somebody has to say what the work is before it starts. Somebody has to answer for what comes back.

And most of us assume that that's the same person, but often it isn't. And sometimes these days it isn't even a person at all. Joining me today is Paul Swift, a dietitian by day but under cover of night. He transforms into his alter ego as the AI Chemyst He writes articles and rights wrongs under distill AI tion on Substack. We'll link it in the show notes.

Recently he's been working out what's left for the human once the model can do the making. he's building his own answer and giving us a behind-the-scenes look to that question, as we'll explore today. So welcome, Paul.

1:32 Paul Swift

Hey Sam, how's it going?

1:33 Sam Rogers

It's going great. So glad to have you here and anxious to get into the builderness of this episode. Paul is the perfect guest for today because he's got a regular job and family and life, but rather than trying to just predict the future of AI in our lives, he's actually building it into the present

and you've got no formal training in AI or coding or or anything

1:57 Paul Swift

no engineering background. just been wildly curious and open to experimenting and I love R and D. that's my bag. So this is like just pure R and D right now, right?

2:08 Sam Rogers

Yeah. So this is a great example of a passion for the process, a curious and scientific mind, and a little free time every now and again.

so yeah, let's get to it. Paul, guests go first. So what's the signal that you're watching right now in your own work this

week?

2:25 Paul Swift

One shot generation. I was flicking through X nearly every day. and there's just so many people now posting entire games, levels, it's like

2:34 Sam Rogers

Yeah. Yeah.

2:36 Paul Swift

3 D environments, people building

multiplayer games and doing it within four hours and sending it to the friends and just so many things now that it's just quite mind blowing to look at. from a signal point of view, that really links into what I've been that w that we're gonna get into today, which is kind of really thinking about one shot generation. it's become so capable that I'm finding it

increasingly valuable now to treat that as not the only thing you do but the artifact of what it generates turning that into an ongoing process.

3:09 Sam Rogers

you built a whole studio, like basically from your phone, to make one six and a half minute video like last I saw. so maybe

you could say a little bit about that. Like why not just make the video? Like what is it that you're trying to create here?

3:23 Paul Swift

Yeah, so it's the first thing that I did was do that one shot and then just to really see what the capabilities are. And I think that's really

3:30 Sam Rogers

Mm-hmm.

3:31 Paul Swift

important. is that we you know, I'm what I'm not saying is that we should stop one shotting things. And by one shot for anyone I guess who's listening, who's wondering what I'm talking about there is kinda just try and do a whole video in

one prompt and let the model try and build six minute video. so in this instance it was I gave it one of my older articles and I said within a very short prompt I want you to turn this into a video explainer using you know motion graphics and XYZ. and it really did just build out a six minute video which kind of blew me away. I was pretty jaw dropped. And it wasn't too far off in quality.

and content than the eventual video that I made however

4:15 Sam Rogers

Mm-hmm.

4:15 Paul Swift

the big thing that we'll get into is that was considered pilot one as a proof of concept right because what I found was it puts you in this situation that if it can make a whole six minute video at once and you just drop an article in what I found myself doing was saying

Hmm at two minutes I didn't like this bit. At four minutes I don't like the video.

4:35 Sam Rogers

Mm-hmm. Yeah.

4:36 Paul Swift

I don't yeah, so you end up kind of like bug fixing after the fact. and that kind of starts to get a little bit tedious in a way. and you kind of putting in your judgment later. So like that's kind of where my mind went with the concept of the studio

is

Hey, can I break this down now and reverse engineer that? so that I am actually involved from the start in step

4:59 Sam Rogers

Cool.

4:59 Paul Swift

through each step, rather than coming back in at the end like a bookend.

So yeah, and I guess that's where I started taking that into the idea of okay, well we've got pilot one, it's a proof of concept, it's a one shot. I can see visually now that it's capable of doing what I want it to do. And then taking that video into a brand new session, calling it pilot two, and then saying to the session, we're gonna break this video down into chunks and

we're gonna rebuild it chapter by chapter as a review desk, as a studio, with surfaces that I can add my judgment in and we can build it step by step together, rather than hey, do it all at once.

5:44 Sam Rogers

Yeah. So what does the review desk let you see that a finished video would hide?

5:50 Paul Swift

we can talk about it, but it's it's better to show it, right?

okay, so for everyone who's never seen this before, so this is just Claude on the browser and I've got it in code mode with some of my other sessions here. this pane right here is

the end of a very long conversation where we've built building the entire video. we got up to chapter seven so I've kind of already loaded it up. when I'm talking about studio I'm actually asking it to essentially build a visual dashboard, a visual place

that I can also put my judgment and things into. the thing that I like about it is there'll be some people who are looking at this and thinking that this looks kind of similar to Claude design. I don't know if you've used that before, Sam.

6:35 Sam Rogers

Yeah, I have.

6:36 Paul Swift

Claude Design yeah. in a way that is a studio, right? It's a design studio. you prompt and then it creates something and then you've got those little toggles.

And you can change and tweak things, right? So it's very much the same concept as that, except it's all built in the command line interface here. And I think the beauty of it for me is I can make this completely bespoke. I can change anything that you can see here if I don't like it.

getting back to your question what does this do that a one shot, you know, a six minute video can't do with some additional prompting where you say, I didn't like that and I didn't like this is I think that it's really hard to articulate and describe and define what you want when it's a visual or a video thing, like an image. It's always

a lot easier to show it

7:25 Sam Rogers

Yeah.

7:25 Paul Swift

or to point it out.

one of the other advantages of how I've built the studio is not just the toggles but it talks you through what's been done and what decisions have been made.

It actually presents you with decision cards. So did anything go wrong in your device? No. You can put comments in. you know, the two dials that we recently changed from the skyscraper. So that was from the last decision I'd made. so

7:48 Sam Rogers

Mm-hmm.

7:48 Paul Swift

you can see on here this is chapter seven, round four. So it builds on each decision until I've said I've approved chapter seven.

so for instance in here, two dials, yeah, it's nearly there. Is the whole film approved? No rework first. So this is where kind of the magic happens. So it's facilitating that decision making. sometimes

8:06 Sam Rogers

Yeah.

8:06 Paul Swift

it's hard to sit there and just write a massive prompt to say, Oh, at this point I didn't like this, at this point I didn't like that. Again, it kind of makes it

really difficult but actually the decision cards in here really facilitate that for me and it kind of goes back to to the title of this episode which is kind of you know thinking of myself like a bit of an active client and then I've got

8:29 Sam Rogers

Mm-hmm.

8:30 Paul Swift

the model is a bit more of a consultant and then also it's a coder and it's a bit of a team at the same time. It's got many hats on and you can ask it that so I'm kind of approaching it like a client

'Cause I don't have any formal training in videographics or anything, but I know what I like when I see it, right? And I think most of us are like that. the cool part about it is at the end of every review desk is the prompt it automatically compiles the prompt for me. if I type in um

Yes, it's great. In any of the comments it actually live creates the prompt itself. So by the time I get to the end, after I've put all my decisions and all my comments about certain parts, it actually has already compiled the prompt and you can send it to session in Claude. so this one's already been sent, it says here

So it really kind of was a very frictionless way to collaborate with the model on making the video.

9:28 Sam Rogers

Yeah, and I of course love that language. Like that's what I'm all about is helping facilitate AI collaboration, whether that's with people and AI or a series of agents, different LLMs, multi model families, like what does that collaboration look like? What does it look like from AI to AI, person to AI, all the different permutations?

the great part about what Paul's doing is he's not trying to sell it to you.

This is what he made for his taste, for his way of approaching things, his way of thinking about it as a director or you call yourself a tastemaker sometimes in some of your posts, which I think is it's not the kind of language that we would normally think of, but I think it's very

apropos.

10:11 Paul Swift

Yeah, yeah. I call myself a lot of things. That that is defin

10:14 Sam Rogers

Ha ha ha.

10:15 Paul Swift

some of them I can't say on air, but yeah.

10:17 Sam Rogers

Yeah.

10:17 Paul Swift

I think in a way that's where we are all going is being the humans that provide the steering, the direction and really the why.

And I think that's probably what's missing a little bit in some of the discourse is

10:32 Sam Rogers

Mm-hmm.

10:32 Paul Swift

I think everyone's worried that the doing and the how to do it is all being hollowed out by AI. And I think, you know, to a degree it is. these two words that I keep in mind, which is automation and augmentation.

if we look at the review desk that I've created, it's a sort of nice mixture of both where the model is automating all of the sort of decisions that it needs to create and code the visual experience, the video. But the review desk puts the steering wheel back on, I guess is a good way to describe it. And now I've got

11:02 Sam Rogers

Mm-hmm. Mm-hmm.

11:05 Paul Swift

the dashboard a little bit like

a brand new electric car, like let's say in a Tesla or something, it's got all of these, you know, dials and dashes and a few pedals and a steering wheel, but the actual complexity is all being run by the machine, by the computer, by the model underneath, right?

We just need to make sure we've got a handle of the steering wheel to say, this is the direction we want to go. This is the direction we want to I want you to stop now. you bring the purpose, you bring the destination, you say, Right, I need to go over to this shop over here, you know, because I need groceries. You're bringing the purpose into the vehicle in a way, and the machine's just allowing you to

enact that purpose.

11:46 Sam Rogers

I want to get to my signal here real quick It kind of compliments yours, which actually just comes from lunch today. I spent it with a room full of lawyers on four checks that are important.

To consider before any kind of AI answer leaves your desk So how is it that we're expressing not just our taste, kind of that upper bound, but the lower bound

a lot of times AI will hallucinate some things that may or may not really exist. So as the human, taking the ownership, taking that steering wheel, does it exist? And then reading it, does it actually say what the summary claims? Because oftentimes AI will take some shortcuts, no different than a human would. We've

Talked about that on this show before. Validating that the summary actually matches the content and how to do that, And then the date of it is a critical piece. It's so easy to overlook because, especially when things are changing frequently, such as new litigation, new cases, things like that. AI may be giving an answer that was accurate.

From 2022, and there was a new case just filed a month ago. And then the last part of the four, which is super important, especially for people who

hold some sort of certification. doctors, lawyers, accountants, all those kind of people, maintaining that human ownership. AI can do a lot of things for us, but what it can't do is own things for us. we have to say.

I'm signing it. I'm putting my name behind this.

the rules on this haven't changed. They're still the same. If you're putting your name on it, it's your name. And AI doesn't absolve you of responsibility or accountability. You're steering the wheel, you're deciding to put that automation engagement of like the full self-driving mode or whatever you want to say.

Like you're making the choice to do that. So it's really critical that that skill

Be exercised by all of us. It's not just in the realm of doctors and lawyers and accountants. It's really something that we all get to do, is how I look at it, but that we all really need to do to maintain that floor of how we're working with AI, how we're creating and collaborating, where beneath this nothing will pass, right? and we have some.

human ownership and accountability there. And then we get to play with stuff like you're creating, Paul, which is that ceiling how well can I do this? what inspiring things can I create? what am I going for here? What's the goal? how much value can I deliver for the lawyers? How much risk can I reduce in that process for the client, for the firm, for the, you know, the person signing their name.

There's no one right answer for all this stuff. it's all very nuanced. especially when it comes to stuff like you're creating, Paul, I'd say it's all very fun.

14:37 Paul Swift

Yeah, yeah. I mean I love I was just nodding away all of those things you were saying because AI the large line large language models, you know, it's a distributed network, right, on your computer. There's no accountability that the model can have.

14:50 Sam Rogers

Yeah.

Yeah.

14:51 Paul Swift

it doesn't own the consequences is another thing that I tend to say as well. it can tell you to do something, or suggest a simple example is Black Friday deals. it can suggest the different types of dishwashers that you could buy, that it's a lot cheaper or a good deal. but you're the one who has to spend the money and

You're the one who owns the consequences of being in the house it's really important

that we realize that we own the consequences, like you said, I often I'll say to people, for anything consequential, always have an air gap. An air gap meaning nothing

15:23 Sam Rogers

Yes.

15:24 Paul Swift

leaves the desk, nothing is sent by an agent or anyone if it's serious. it can only be generating and producing things.

But it but don't give it the send button unless there's a reverse button or whatever.

15:36 Sam Rogers

Yeah, yeah. That's one of the things that we were talking about today. and especially for me as somebody who started with AI before it was online, like just working locally, has been my default. So I'm always like thinking that way. that's not the common way that everybody thinks about it. But it's so important that like if it's an irreversible decision, it's a different threshold,

especially from like a, you know, a privacy perspective, working with all these lawyers, you know, it's information that can be trained off of. how do we manage that? and one way is to create that kind of air gapped model that the firm can use, but that's not connected to the internet. Like there's no way that it can leak.

As we've said before on this show, either it's architecture or it's broken. when it comes to security, it's not a polite request, please don't share anything. Like the more that we can make that really a permissioned infrastructure, which we talked about a bit last week. with Tim McAllister from Digicert.

it really does make a difference, and that's going to become much more important over the next year or two to come.

well, let's

16:43 Paul Swift

Yeah,

16:43 Sam Rogers

get to the subtraction portion. as we're talking about what things shouldn't happen or what restrictions we want to put. So given everything that we've said and what you've been working with, Paul, what has stopped making sense and what did you actually stop doing as a result?

17:01 Paul Swift

it's stopped making sense treating the the first artifact that it produces as the product itself. so, you know, very much relating to stopping the one

17:08 Sam Rogers

Stopping the one shot. Yeah. Yeah.

17:10 Paul Swift

it's I still one shot a lot and I think I probably I just wanna be super clear. I used one shot as the proof of concept.

AI's good at is tricking our minds to saying this looks polished. I could have actually sent

17:21 Sam Rogers

Right.

17:22 Paul Swift

away sent that six minute video and just plonked it on because it looked really good and it looked just as good. But the actual end video that I had after I'd run through the studio did two things. One, it had all of my decisions and judgments and everything in it.

it was just the way that I wanted it. And then that led me to say, I'm happy to put my name on this, stamp it and put it in the internet. Now, it goes back to your point about the consequences and having the accountability for that. in this instance it's a fun video but

you can transfer that same thing to something more serious. if we go back to this the subtraction, I've stopped treating the first thing as ready, as it as the impressive artifact. Let's say it's an essay. you can read that essay and because the prose are so good and so fluent

it's hard to have a comment about it oftentimes. what I've stopped doing now is treating that artifact as the final product. and then what I'll do is I'll take it to pilot stage two and actually build that as we go. where I'm putting my judgment, whether it's an essay, whether it's a piece of research,

whether it's a video or an article, it transfers that attitude, that approach to it transfers.

18:36 Sam Rogers

Yeah. this reminds me of a comment from our mutual Substack friend, Limited Edition Jonathan, who's been on the podcast a time or two. just had something that he'd posted like yesterday about how salespeople will like pitch something ninety times and maybe

18:53 Paul Swift

Saw that

18:53 Sam Rogers

make 10 sales

the expectation is that, like a salesperson, like you're gonna pitch a bunch and you'll sell a little and you'll be better for that experience.

19:04 Paul Swift

Yeah, it's the

it that kind of eighty twenty principle as well, you know, it's like you

19:09 Sam Rogers

Mm-hmm.

19:09 Paul Swift

could do a one shot and then you've got your eighty percent already. Or you can kind of flip it and say, We

19:16 Sam Rogers

Hm.

19:17 Paul Swift

create a studio and I'm focusing on the twenty percent in the studio

19:21 Sam Rogers

Yes.

19:21 Paul Swift

and then the output then is my eighty percent. It's kind of flipping it around and i again it's just how you approach things.

19:28 Sam Rogers

I love that, the inverse Pareto principle. Yes.

19:30 Paul Swift

The inverse pareto, yeah.

we could coin that.

19:33 Sam Rogers

Yeah, go register the dot com quick before anybody else does.

well to throw my subtraction out there, one thing I was just saying, at this lawyer meeting, was where people will ask the AI, is it sure? Or

Yeah, ask the AI to stop hallucinating or something like that. And

19:49 Paul Swift

Yeah, and we're gonna

19:52 Sam Rogers

like you, I'm building a ton of stuff and I forget that like those basics that are basic to you and me, they're not necessarily common to the people that you would think, you know, or in a technical profession. but whenever you're asking the AI if it's sure, it's almost always a yes.

Because you can't tell the good answers from the bad answers, right? It doesn't blink red when it's making stuff up. so it really doesn't tell us anything. And it's just the confidence that the model always has. So instead replacing that with an exact quote and the pinpoint of, like in that legal example, like show me the exact line, show me where it lives.

Open it for yourself, read it for yourself, all that stuff that was, you know, part of the signal. when it's a quote that is like set up for you to research, set up for your normal workflow, which you can have in the prompt, you're setting it up for this workflow. You're setting it up to be vetted, to be not just rejected, no, that's not it, but.

Just like working with anybody in any creative capacity, you're gonna wanna affirm some of the things and get really specific. Like with the sliders you're displaying in your

21:07 Paul Swift

Mm.

21:08 Sam Rogers

studio. I wanna move this one. I'm not moving that one. Here's what we're focusing on.

21:13 Paul Swift

Yeah,

going back to what you said before, about the hallucinations,

and it 'cause I've been working a lot and this is kinda a tangent from the studios, but I've been working a lot around how you've got type one sort of hallucinations, which is the source does not exist.

And I think th in a way those are the least worrisome ones because if the link doesn't work, it's not there. And you could say you've made that up, it doesn't exist. what I've found is the more difficult and troublesome ones that I've been trying to deal with are the what we call the warrant. I don't know if you've heard of Toolman framework, but essentially where it gets like three different

sources, it cherry picks whatever parts of that help to answer the question you're asking and puts them into a it generates a new paragraph, cites those

22:02 Sam Rogers

Mm.

22:03 Paul Swift

three sources, but then when you actually read those final sources, the it's taken some of those out of context. So as written, it's quite indefensible. And that's what I'm trying to work a lot more on now is

How can I make the model better at providing a more defensible warranted claim based on multiple sources? So I'm working on a l on a few things that I can share with people at some point, but a lot of it comes down to having a generator agent or your Claude or whoever is your generator, and then having

22:39 Sam Rogers

Mm-hmm.

22:39 Paul Swift

two

sub agents or two extra you could even use a separate session and saying your job is to verify what it's saying is defensible and then you have a third one whether it's an agent or a new session is to validate everything that's been changed as a consequence of the middle checker. So you've got a generator, a verifier and a validator. And I'm when I'm working now

on anything serious, I tend to outline that and ask you know ask Claude to always spin up every time you've done something spin up a verifier agent to check those claims exist from what the generators made. And oftentimes it's just finding lots of mistakes and little things. And it's almost like the first pass really. even if it does a great job

I'll still in the end still want to click on the link to make sure that it is real and I'll still want to read the source. it's not automating you away, your expertise and your judgment, but it is adding a second pair of eyes, is probably what I say. So sometimes I'll do something myself and then I'll say, Hey, can you check what I've written right?

You know, so it's like it's nice to have a second pair of eyes on something, right?

23:53 Sam Rogers

Or a second pair of AIs, as the case may be.

23:55 Paul Swift

I second

pair of AIs, I love it.

23:57 Sam Rogers

Yeah, that's I'm not familiar with that framework but I've done something very similar kind of from the beginning with PAICE.work but also with Harnessie which allows you to use competing model families and even local models to dispense those kinds of roles. They're not quite

The same. I actually have four. but similarly, like it boils down to a certain set of jobs that make a team, right? Like on any team,

24:24 Paul Swift

Mm-hmm.

24:26 Sam Rogers

you're gonna want people with different sets of skills that are working together and spotting each other. And we know how to do that well enough as humans, but we're still building the skill to do that with AI because

We're good enough at catching the mistakes that we make as humans with a fellow human, because we make similar kinds of mistakes. AI makes very different kinds of mistakes. And we're not as well tuned to that. that's one of those things we need some reps on, right? and in many cases, like just with the speed and scale of AI, it kind of takes another AI to be able to work that fast.

Because it exceeds the human context window.

the three kind of roles that you were describing, the fourth one that I have is exactly this, which is a watcher and logger that is always air gapped, basically,

I'm using a local model for this. It doesn't need to be frontier level intelligence. It's better if it's not. But something that can see what happens and state what happens without summarizing it and introducing risk in that summary. But basically just catching it all and logging it and making it referenciable for humans to be able to do audit, but also.

so that the other AIs can see that they can't cheat stuff because they do respond differently when they see that their activity is being logged. And they're very well informed about it using Harnessie they act differently. No, no different than like I'm gonna act different when I know that someone's watching me, you know? Like it's just one of those things that

26:08 Paul Swift

Kind of eerie, isn't

it? Yeah.

26:10 Sam Rogers

It is, yeah,

so yeah, having that watcher element.

And having it be not a secret, but public enough that the other AIs can see that log, but they cannot write to it.

26:23 Paul Swift

Love that.

it sounds like you've implemented that much better than I have, but funny enough, probably this week or two, I have started creating Watcher agents. So it's

26:32 Sam Rogers

Okay.

26:32 Paul Swift

literally called it the Watcher, which is kind of funny that you said that. Slightly different implementation, but

same sort of thing.

You need something to be keeping an eye on the bigger picture to see what's going on you make an interesting point there that I'd not thought about is how agents then change their inference or the way that they chain of thought if they know that something's else is outside of them is logging and watching. it's kind of fascinating, like the psychology of of a large language model.

It's

27:02 Sam Rogers

Yeah, it's

one of those things that shouldn't make any difference. And yet it does. And why wouldn't it? Because it does. Like for us, you know, every time we assume that they are so completely

27:09 Paul Swift

Yeah.

27:12 Sam Rogers

different and alien, we're usually wrong about that.

27:13 Paul Swift

Mm. Mm-hmm.

27:16 Sam Rogers

not to over-anthropomorphize them or anything like that, but there's certain things that I think are emergent properties of complexity that we are coming to understand better now that we have another.

kind of alien complexity to compare ourselves to. And we can see like this other form of intelligence developing. it's interesting how many similarities there are and where the differences are. and it's my hope with all of this that we will come to understand ourselves better because we are just as jagged as the jagged intelligence of AI. we're just jagged differently, you know?

And so putting them together, giving the parameters, building the studio, like giving that interface that works both ways, so that we can keep that steering wheel, you know, we can keep driving where we want to go, but also so that on the other end, can read our intent in ways that are simple for us, that it knows where to go with that, you know, how to respond to the controls.

that's really important too. anyway, thanks so much, Paul, for sharing what you've been building with Studio. Very cool, very inspiring. and I'll be chasing on your heels with that. I'd love to share it with you and see what you think. And

28:26 Paul Swift

I would love that.

28:27 Sam Rogers

yeah, I'll be sharing it with everybody else as a takeaway from this episode. but Paul, if people

Who want to follow you where do you send them?

28:36 Paul Swift

So I've got a sub stack, under the AI chemyst so C H E M Y S T and I think there's a link somewhere

28:44 Sam Rogers

Yep. I will make sure it's in the show notes.

thanks so much, Mr. AI Chemist aka Paul Swift for joining us today. we do this live.

Every Wednesday, a new episode drops on Friday, as well as all of the podcasting places. Sunday will be issue 76 of the newsletter. And if you're listening on audio right now, please click follow so that the next one just magically shows up. And if you're watching, I hear that liking and subscribing will give you good luck and good fortune for seven days. how can we find out?

wherever you're getting this signal right now, just make sure that you click the thing that gives you the next one next week. Everything about this show lives at sigsub.show. One signal worth watching, one subtraction worth making. Now go find yours.