Jerod Santo: I wrote Python for about six months a decade ago, and I had never written any before… And it took six to 12 hours of just like dinking around to where I could do some basic – having already know how to program, right? Like, I was a Ruby programmer. Moving from JavaScript over to Python is not a leap. Of course, Python has this huge breadth of things you can do, but we’re talking about just core functionality. I don’t think it’d be a big leap for somebody if they wanted to make that leap. But like I said, you probably don’t have to; there’s probably JavaScript tooling for this exact kind of thing. But to get into the Textual world, if you wanted to, Python is a very – Will was talking about it earlier on, it’s a very approachable programming language, which is kind of a weird, ironic thing with it not being approachable from a distribution perspective historically…

So I don’t think Python is going to hold them back… Yeah, getting people interested in text applications in general is probably a bigger selling point. How long do you have, Will? Do you have a runway calculated out with your current burn rate? You know, startup stuff… “We can work on this for N years, or N months before I need to have some sort of an income.”

John Spurlock: So I guess stats is a good jumping off point to talk about OP3… It’s a little esoteric, and so you kind of have to know all the pieces. But I think we’ve talked about the various pieces at this point. So it solves basically three problems. The first problem is the problem that Jerodd, you talked about… It’s that we actually have these nice third-party independent analytic services, but the incentive for them is to take the download data that they get, and immediately join it to other third-party IP address databases to enrich that data… Because they’re interested in not only showing how many downloads a particular show has for an advertiser, but what the demographics of that audience is, and what the income of that audience is, and breakdowns by gender, and by ethnicity, and by political affiliation.

[00:24:08.27] And I’m sure you guys know, it’s just a matter of how much you want to spend to how kind of creepy the level of detail you can get, down to the neighborhood level, down to what apps they have installed, down to what they were yelling about… And they have an incentive to do it, because that makes their product more attractive. But they also have an incentive – or even if they were not as interested in that, they also then become, if they’re popular, like Chartable and some other companies, they become a very attractive acquisition target. So they can be joined by a yet larger company with their information, for strategic reasons.

So that is kind of one problem that’s out there. So even in – you mentioned already that Chartable and Podsights, two of the largest of these analytic services got acquired in February. And you can imagine, I’m sure even some of your listeners might be thinking “Oh, this sounds like a great business, so I could probably write some Perl scripts to do this. I’ll just be the next Chartable.” And that doesn’t really solve the problem, in my estimation, long-term; it just repeats the pattern. So they’re gonna get popular, and they’re gonna have the same incentives, they’re gonna – it’s the same sort of pattern. So that’s problem one.

Problem two is that the hosting companies themselves, there’s all of these hosting companies now; they are very mature, and they all offer very similar features. As you say, table stakes features. They all now have to have some sort of stats ability, so they have to task someone as part of their sprint every month, like “Oh, make sure the stats are clean, that we’re filtering out all the bots, and making sure we’re performing the calculations properly.” And they’re basically all showing very similar charts, similar charts and graphs. And it’s even more ironic, because the larger shows then turn around and use a third-party service to they use their stats, because they don’t trust the host stats.

So it’s table stakes, and it’s a lot of work, and they’re all doing it separately, right? So it’s not really value-add anymore. It used to be. Maybe even five years ago, it was like a selling point; you could say, “Hey, we offer great stats.” But now, most offer stats, and they don’t really view – they would love to get on and work on other things. So they kind of view stats as a – I don’t want to say a commodity, but you know, you specialize; when an industry gets mature, you specialize into functions.

So that’s problem two… Problem three is a little more esoteric, but it’s the whole notion of – we talked a little bit about kind of open podcasting before we started here. One of the cool things… I was around – I’m old enough to know, like, before the internet was around at all. And when the internet came around, as somebody who likes to build things, you just have this limitless possibility of “Oh, how cool would it be if we did this? Or that?” and you’re only limited by your ideas and the building blocks that are available. Obviously, there’s some downsides that we’ve seen as well, but there’s still so much opportunity there. And podcasting is one of those sort of interesting places where you, anyone can publish an audio file and have it pretty much distributed automatically to all these different venues while they sleep. So it’s not that hard. However, there’s some big chunks of the system that are not like that, even in open podcasting.

So let’s say you had a great idea for an app, you wanted to create the next big podcasting app. You can go out and scrape all the RSS feeds, you can get the show and episode level information, you put it all together… Hopefully, they have nice chapter information, and tags, and so forth. But think about what YouTube does - they have some things that you currently can do, like comments and monetization. But even a more core thing is recommendations. So what podcasts are people listening to? What podcasts are people listening to around you? People that subscribe to this, subscribe to that. Even if it’s a very popular app, they know within their app stats like that, but they don’t know across the whole industry, because it’s so distributed. There’s no place where that information resides. It resides basically in silos, at different levels; at hosts, and so forth.

[00:28:23.18] The services, these third-party analytic services - they have it, actually a fairly broad sample, but they don’t make it available. So that’s kind of a third aspect, is it would be great, it would unlock all kind of more features that we could build on top of the open podcasting system, keep it competitive, if that data was available. But then, as a listener being available in a safe way, right? Because more people listen to podcasts than make podcasts; you kind of need to satisfy both concerns. And as a listener, I’m not sure I love the fact that my IP address is going everywhere, right? That’s where these analytics services come in. And most apps don’t disclose that this kind of stuff is happening, right? So you want services to kind of do right by the listener, even without an explicit agreement there.

So OP3 is a system that I was putting together, I was like, “This is the internet, and I can build stuff… Let’s try to build an ideal system that solves a bunch of these problems at once.” So it is an analytic service in that is very similar to Chartable. You add a prefix, op3.dev/whatever to your episodes; it’s completely free. It runs on a CDN platform, so it’s up 100% of the time. It’ll never be the bottleneck in getting to your content… But it’s sort of radical in that it turns around and does all the minimization, and so forth. It throws away most of the requested information, and it stores it. But it turns around and makes all the participating shows data, the minimized data, the minimized request logs available to anyone. So it turns around and makes hashed IP addresses, what episode was downloaded, when it was downloaded, the user agent… Things that are not user-identifying necessarily - it turns around and makes that data available, so that now a startup can go and look at that data to say, “Oh, this is a signal of what podcasts are trending in Cleveland”, that sort of thing. So that is radical, because none of these other services will – you know, they are very against that. That’s kind of the core mission, is not to do that.

Leo Mutuku: Thanks so much, Joyce. I agree with you and Godliver on the data sources that you mentioned, that might be available in terms of supporting the building of technologies to support agriculture. In our context, we have tried to use mobile phones to basically collect groundtruth data which typically is not available, as you mentioned. Datasets out there might not be relevant for our context.

So we are trying to see what are the easiest tools and technologies we can use to collect this data directly from the farms themselves, from the farmers… And we’ve found that mobile phones, using tools such as the GeoODK app, really help in maybe creating proxies for some of more advanced data collection tools such as the Garmins, which might be expensive in our context… And we’re using this to collect data, for instance, on farm boundaries. It’s very important to calculate an area and the cultivation, and what is being cultivated within these different lots, so that we can try and estimate, let’s say, the yield production, the access to inputs required for bountiful harvests.

[15:51] So farm boundaries are one of the datasets we also try to collect. The other is the access to specific inputs. When we talk about climate-smart agriculture, in this case we would like our farmers to access hybrids of seeds, for instance, and fertilizers that work well in their context. So access to input data and what inputs are being applied in different parts of farming communities is a very important dataset to collect.

You mentioned the use of things such as photos to be able to identify pests and diseases as part of farm management practices, but I would also want to talk about another dataset that we find very important, which is commodity prices. So what are the market prices of these crops once they are harvested at different times throughout the year… Because if those are predictable in a way, then it’s easier even for farmers to estimate what their income would be once they take their crops to the marketplace.

So I think across the general ecosystem, different stakeholders have a role to play to make these datasets available. For instance, the market prices might come from government sources, and yield estimates, while a lot of projects are trying to see how we can make it easy to collect this groundtruth data from farm samples that can be used to train artificial intelligence algorithms and tools.

Adam Stacoviak: “…on a word calculator.”

Adam Stacoviak: The math being off is just like driving me crazy. I just did it today, I was like, it’s accurate 95% of the time, and then one time it’s like, it does terrible math. I’m like “I could never charge our customer that or our partner that, because that’s not the real number.” And so I’ve got to check it every single time. I’m like, “I just want the word calculator part of it.”

Adam Stacoviak: …is Docker, Ollama, Open Web UI, either a really powerful CPU, or an IGPU. I don’t think Ollama actually recognizes IGPUs, or GPU. And GPUs are scarce, expensive… You can find decent ones on eBay… They’re actually – if you can eBay well, you can find a decent one. Maybe just buy the M4 Mac Pro, and call it done. And maybe you’ll have a limitation, right? That’s essentially an IGPU. It’s an integrated GPU. I think Ollama runs well there, to my knowledge. But if you want to go crazy and do multiple GPUs, you’re going to have to motherboard that thing. That means you’re going to choose between Linux and Windows. And in my case, I was like “Okay, I got better mileage or easier mileage on my shoes when they were stamped with Windows.” And so that’s the route I went. But basically, Docker, Ollama, Open Web UI… Pick your model from there. Capable hardware… And that’s how you can get to local LLM today. You can play with DeepSeek, you can even do a smaller parameter model… LLaMA 3.2 is probably the one that most people can run, because it’s small enough. LLaMA 3.3 is pretty big. It’s 70 billion parameters. It’s sizable. 70 gigs, I believe, which is larger than most people can run on a single machine, unless you’ve got multiple GPUs.

I wanted to get to that point where I could run those things, and what I would love now is once I’ve sort of gotten that mastered and running stable, to begin to – and I don’t even know what the terminology is, Jerod, so forgive me if I’m stupid in this regard, having done what we’ve done for so long; now I’m actually learning it… It’s I want to train one of these models, or teach it a particular trick. The particular trick is like “Okay, here’s all of our agreements, here’s how we price, here’s how we quantify… Make that job easier.” Can it make that job easier? Can it make me faster in those ways, and be just as accurate, and sweat the details just as much as I do, in the real time, doing those exercises myself? That’s what I’m really trying to get to. I would love it if I can get to certain things I do that are deep think things, that are crucial in business to get right… I don’t want to give people the wrong price, give them the wrong contract, or the wrong terms… I want to give this thing certain parameters, certain business parameters to get to certain results, so I can kind of like word-calculator my way to faster, easier, better quotes for people, and stuff like that. I don’t want it to write my novels… I just want it to help me make those kinds of things faster, and be my word calculator, and be very accurate. So that’s my next journey.

Jerod Santo: That’s a funny workaround, because if you’re a calculator for words, you’re not necessarily a calculator for numbers…

And when you can’t do those reliably, then you could just write a program that does it and returns the same thing every time.

Adam Stacoviak: And that’s what I mean about word calculator. You could probably gush about AI for forever. Let’s not do that. Let’s talk about two more things… Especially if you have time. Do you have more time?

Adam Stacoviak: And I was like “Man, I don’t even own this film.” So I bought it on BluRay, because it’s not on 4k… And for whatever reason, whenever I go in to make MKV to rip it, it was naming all the files a dot. And so all the files it was ripping, each one of them was a dot, and it had an increment, but all the files were hidden. And I didn’t notice it until I went back after it was done… And I was like “Where the heck is all the – I see that it successfully ripped, but where’s all the files?” And I’m like “Oh yeah, these are hidden files. Let me go and fix this.” And there’s like 25 of them in this directory. I’m not going to go one by one inside a terminal and rename these things. And I’m also not going to think about a script to automate renaming these things. So what did I do? I went to my good friend, ChatGPT, and I said – I copied the directory and I said “I’ve got, on a Mac, I need to rename all these files so they’re not hidden. Ideally, they’d be in sequential order of sorts.” And it gave me a script. I put that script in, and boom, all my files were there. I mean, that is the beauty of this word calculator, Jerod. Like, I don’t have to think about any of that stuff. Sure, I could have, it would have been a fun exercise, but it would had been a – what do you call it? A squirrel, a yak to shave. I’d have been there for 20 minutes thinking about this, versus one. Not even one. Off to the races, Fight Club is renamed, and boom, it’s on Plex, and I’m watching it, and I’m happy. That’s the way life should be.

Mat Ryer: Yeah, I get it. I get why people – it was like with calculators in my… I remember my cousin wasn’t allowed to use a calculator in one of his exams… But I was younger, and we were allowed to use calculators, and he was outraged. I mean, it was an English exam, so it didn’t help, but still…

[singing] Sometimes your jokes don’t go down so well… Don’t worry. Don’t worry. It’s alright…

Just a little one for me there…

Mat Ryer: Yeah, no, and I think it’s quite convincing. I think you might convince some people… It’s just – we did a hackathon, and I was writing React code, which I’ve never really written… TypeScript, React. I was dealing with that stuff, and probably prototypy quality level, if we’re honest. I would want someone who really knows it to probably rewrite it, frankly, at this moment. But yeah, there’s times where that autocomplete just really does accelerate it.

[01:15:57.05] There are other times where its suggestions lead you down the wrong path. And I have the experience enough maybe to know that, and others might not. And so I do think we have to be cautious with this. But you have to know what to press on a calculator, too.

Johannes Schickling: When you say “How do we build an API?”, I’d love to hear a clarification of that, since one idea around local-first apps is that you actually – you should ask yourself, “Do we still need an API?” It’s all about the client-side experience at the end of the day, or some other things that you need to build. Maybe you want to send an email if a certain thing happens etc. and you might want to do that on some server-side thing. But one of the ideas of local-first is trying to nudge you away from that very API-centric way of thinking of an architecture. If you’re building, let’s say, a calculator app that just happens to synchronize between your phone and your tablet, then you could try to think about that as a “Oh, we’re going to have an API here”, etc. But if you just start out working on that in a single-player experience, you would never think about “Do we need an API for a calculator? No, we just built that thing client-side.” And now, if you have that as a starting point and you want to synchronize data between your different clients, you don’t necessarily need an API for that. You just need a transfer mechanism between the clients, and that’s rather where the local-first idea comes from.

That being said, you can still interact and integrate with APIs, but I think with local-first, you rather think about “How do I synchronize the data?” and less “How do I request-response with an API?”

Russ Cox: Sure, sure. So we’ve been trying to figure out what we should do with AI; what sort of new capabilities the recent AI advances in the particular large language models unlock. And one of the things that I think really helps – so there’s a lot of, I think, unwarranted optimism about exactly these amazing things that LLMs will be able to do. And maybe they will, maybe they won’t. But at the moment, I think that the main strength is - I saw an article, maybe a couple of years ago now, that was sort of framing them as kind of a word calculator. So computers are number calculators, but LLMs are really kind of word calculators.

So in the situations where you have a lot of words and you want to deal with them, having a word calculator sounds like a good thing. And so one of those times is when you’re doing software maintenance and you’re dealing with other people, right? They’re speaking English, and so having some sort of word calculator that can do things with the English text sounds like that would be a win.

And so in particular, we’re looking at what we can do to help open source developers with running their own open source projects, and we’re using Go as a test bed for that. And the goal is really to try to help automate away the stuff that no one really wants to do, like the basic triage of issues, or figuring out where the duplicate issues are.

Right now we have a bot that in the Go repo, when you post a new issue, it looks up using LLMs and vector databases other issues that are very closely related, and it posts a list of at most the top six or seven, maybe ten; I forget exactly what the cutoff is, but there’s a score cutoff of how related they have to be. So sometimes it’ll post nothing. And sometimes it’ll say “Hey, these three issues look very related to this issue.” And originally, we were looking at that for automated duplicate detection and just closing a duplicate, but it’s very hard to tell the difference between “This is a duplicate of the issue” or “Yes, this looks like exactly the same thing, but you thought you fixed it, and now it’s happening again, o maybe it’s different.” You can’t tell the difference between those from the reports.

But just even pointing out “Hey, look, these other issues might be related” has been incredibly helpful, because you get an issue that comes in and you’re looking at it and you didn’t even know about this other issue that someone else took care of. And you say “Oh wait, that does look like exactly the same thing.” In fact, the bug fix for that one has a subtle mistake in it, and that would cause this new one. Seeing those connections is really, really helpful. And in particular, having this kind of database of the context for the project, and having the computers take care of saying “Hey, this looks like related context that you should know about”, that turns out to be incredibly helpful. Because once the project is bigger that you can’t keep it in your head - none of us can keep Go in our heads anymore, all the stuff in the Go project. Having that kind of automated retrieval is actually just incredibly helpful.

So we’re looking for how can we use LLMs and sort of recent advances in AI to help with that kind of stuff, the stuff that people aren’t good at, and that honestly, isn’t that much fun. Groveling through all the issues to try to find the related issues is not something I want to spend my time doing. Whereas, you know, we’re not that interested in having the AIs write all the code, because that’s the fun part, right? Like, why would we take away the fun part? Let’s take away the not fun part.

So that’s the basic idea of the project… It’s just “Let’s figure out how we can point AI at the stuff we don’t want to do”, and also, just learn a bit about what AI can do. To me, it’s sort of still a bit of an experiment to just see what are LLMs actually good for, because none of us really know.

Ryan Worl: Yeah, at WarpStream we know that that’s like us. A very important part of what we do. But it’s always easier to walk into a sales conversation with the hard facts numbers, and not the – a lot of vendors use those exact attributes to describe, to attribute a lot of savings to their product… Which is probably true. But they feel a little bit more wishy-washy compared to the hard facts numbers. So that’s why we lead with those in our pricing calculator.

And obviously, those are still things that we highlight when we’re talking to potential customers, to help them understand the value of the product. But we like to think of that as more like the icing on the cake stuff, and the cost savings is what we’re promising them, basically. Everything else is just icing on the cake.

Joseph Jacks: …and in order to upgrade the software, you would have to order, and they would have to ship you a new series of floppies, from which you would need to kind of like physically install, and you would own those physical like pieces of technology that would give you an upgrade. And then the pricing model was all built around that, too. It was perpetual licensing, where you would like basically buy that version, and then you would have maintenance that you’d pay for the version, and in order to upgrade a new version of the software you’d have to physically spend a new upfront amount for the new version, and then do that over and over again.

[34:18] Most of the – I would say probably 99% with some interpolation synthesis, mixing and matching of ideas… Most of the categories in software, including SaaS, which have become a little bit more like fractal domain-specific… Cal.com has almost like one feature, and wow, it’s a billion-dollar company. That just happens because of economies of scale, and humans using more and more technology. But most of these ideas were implemented in the seventies, eighties and nineties, even in the mainframe days. Scheduling, booking things… I mean, this is like calendar functionality.

I was recently watching the oral history of Mike Markkula, who was the first CEO of Apple. He was a guy that Sequoia founder, Don Valentine, introduced to Steve Wozniak and Steve Jobs to basically help them with the business plan, and like run the company. And Mike was also a really brilliant programmer, and he wrote Fortran and basic programs that would actually run on the early Apple computers. And the collection of those programs came with a credit, sort of a copyright credit at the bottom, which was Johnny Appleseed. You might remember this from the early Apple computers, this Johnny Appleseed name. So Johnny Appleseed was Mike Markkula. And Mike in the computer history museum describes this – it’s on YouTube, you can watch it. It’s a really great, multi-hour oral history. Mike actually personally wrote the programs for the applications that would ship in the early Apple computers. So including the calculator, calendar app, bookkeeping app… A lot of these basic – a very basic, primitive word processor…

And so I just think back to like the most fundamental, basic things that humans do as part of their work or their lives have been re-implemented and re-implemented so many times in the history of technology, going back to literally the fifties and sixties, that when we have new ways of distributing and innovating in the core technology, what happens is you just reinvent and you re-implement what happened, what came before, with shinier and newer, sexier, more sort of societally-relevant and interesting abstractions. And then those things end up taking off. And oftentimes people are given a choice. There’s sort of like the free version of the thing, maybe it’s open source and it’s like less sophisticated, but you can totally control it… Then there’s the proprietary version of the thing, which is much more polished and sexy and packaged. This is so fascinatingly – for the first time, and sort of skipping to the AI stuff, we’ve seen this incredibly shockingly surprising almost acceleration of equivalency. This actually happened this week - I mean, the latest LLaMA models are state-of-the-art. And by all accounts, the benchmarks are pretty much indistinguishable, and the capital required to produce that is similar. Mark Zuckerberg literally did an interview a couple of days ago saying they’re spending tens of billions of dollars building LLaMA, and they’re giving it away completely. And in fact, they’re making – Yann LeCun was just tweeting about this the other day… They’re making their licensing even more permissive. You can do distillation, quantization… You can use the outputs from LLaMA models to train other models… It’s almost like an Apache or MIT-level permissiveness. The artifact is very different for a neural network that has massive amount of information compression. Not in the Pied Piper sense, by the way, but in the neural network, AI, tokenization of –

Tejas Kumar: Yeah, that’s a good question. I just follow a number of experts’ definitions of this thing. I tend not to try and coin terms myself, mainly because I’m just not very credentialed, if we’re being honest… So how do I see agents? I summarize it, I summate it - I’m trying to find the right word - I deduce it from definitions from industry experts who have done it before me. So people like Andrew Ng, the founder of Coursera, and now the founder of DeepLearning.ai - I think he’s got some great content about this, where he defines agentic workflows as workflows that have LLMs perform three tasks, either all three or a subset of them. And those are reflection, meaning generate some output and reflect on it, “Is it good, is it not?”, and then iteratively work on it until it cannot be improved further. So there’s reflection. There’s tool calling, as I mentioned, with RAG, where the large language model will, sort of like a human being, recognize – for example, if you ask me to do a complex calculation, like 324 divided by 9 times 7, I’ll just be like “It’s time to get a calculator.” I’ll recognize that this is the sort of boundary of my capabilities, and go use a tool. So number two is tool calling. And number three - I think it was agent collaboration, where you have – yeah, it’s LLM as judge. It’s this model where a capable model (pun intended) coordinates lesser capable models towards an outcome you want. So it’s like GPT 4.0 being the most capable of Open AI’s models would orchestrate like three or four different 3.5 turbo models that are doing various tasks or generations. And so those three, either one of them or all of them, make up according to Andrew Ng an gent workflow.

[05:47] According to David Khourshid, AI agents are an implementation of the actor model, which is just a programming model where you have an entity called an actor, that sort of acts in response to observing its environment. So the classic implementation of the actor model is PacMan. Actually, a great example of AI, but rule-based AI, where the rules are known ahead of time, is PacMan, where you have PacMan, the little yellow pizza thing, and it’s observing the environment: where are the ghosts, where are the cherries, where are the dots… And you as the player take on the role of the actor. But there’s also demo mode, where the actor model is in play. And according to David Khourshid, this implements agentic workflows. However, it’s rule-based, it’s not generative, but it’s still an agentic workflow, where PacMan is an agent.

So I just take a mishmash of those two - these are the preeminent leaders in the space in my mind - and marry them, and that’s the working definition that I have for an agent. So it’s not a sentence, it’s not a nutshell, but I’m trying to give you more sort of a broad framework of how I see agent workflows. I have seen this term abused, where people will build – maybe abused is too strong… But people will build a custom GPT; this is a feature you can use from Open AI’s GPT 4. They’ll just build a custom GPT, add a system prompt, add some knowledge that GPT 4 can do RAG on, and call this an agent. I disagree; I don’t think that’s an agent, that’s just a RAG application. It doesn’t really do any of the things like we talked about: reflection, tool calling, collaboration, or observing an environment and responding accordingly. So I’d say those four tenets make an agent an agent.

Jerod Santo: [00:15:37.08] For sure. For sure, because it demos so well. And it just – it’s just so impressive. I mean, it’s the stuff out of the future, right? I mean, you basically take your story problem, that your teacher gives you, and you just write it down with your pencil, and draw some little diagrams and lay it out, and then boom goes the dynamite. But again, I don’t think – I mean, I don’t even have an iPad, but I think some of that is in… Math notes have come to also to the Notes app, and I’m not sure if that’s in there, or that’s just the iPad Calculator… But regardless, I will probably never use this feature myself.

Jessica Sachs: You mean like flights and stuff, and like calculator, all those things?

Jared Zoneraich: Yeah, 100%. And just to be on the cutting edge, and know that nobody really knows what they’re doing here. Some people do, but there’s very few of them. So regarding the APIs, ChatGPT, the way I think about it, and the way I usually explain it is it’s basically a very thin wrapper on top of the API. Every time you talk to ChatGPT, behind the scenes it’s using this LLM technology, it’s sending your message, and a little bit of a preamble before your message. And the preamble is basically - we can call it the prompt, and it’s basically saying “Hey, you’re an AI assistant. Make sure to be helpful to the user. Maybe don’t be controversial. You can use a calculator if you need”, stuff like that, and then giving the user messages. And that’s all it is. And the process of prompt engineering is “How do you tweak that preamble? How do you get it to respond in the way you want to respond by telling it what to do?”

When I talk about the API, I’m talking about the things Open AI has exposed to let you build your own track up and let your build your own products on top of it. And I just think Open API has done – if you want to get started and you haven’t touched these APIs today, the best, best thing to do is just go to Open AI’s docs and just read the Getting Started tutorial. It’s really well done, and I think - as someone who’s running a developer tools company, I’ll tell you how hard it is to write good docs. I think our docs – from my perspective, I think they should be much better. Other people say they’re great, but… Well, there’s always room for improvement, and it’s just a very hard thing to do. So that’s where you should go.

Adam Stacoviak: Yeah, for sure. Can we talk about some subtle human leveling up that’s practical for ChatGPT? I mean, I know it’s not Cody. Do you mind riffing a little bit? So last night my wife and I, we were hanging up pictures of our beautiful children. We took pictures of them when they were less than one week old, and then we have pictures of them in the same kind of frame at like their current ages, and one’s seven and one’s three. So it doesn’t really matter about the age. They’re just not one week old anymore. So you have this sort of brand new version of them, and then like current version, to some degree. And it’s four pictures, because we have two sons, and we want to hang them on the wall. And my wife was trying to do the math - and we can obviously do math; it’s not much. It’s like an eight foot wide wall; we want to put them in a [unintelligible 00:15:29.05] with even spacing, all that good stuff. I’m like “Babe, we should ask–” And like, I’m more versed in this than she is; not so much that she doesn’t use it often, she just doesn’t think to. And I think that might be the case of why it’s not being more widely used, is they don’t think to use this. And I’m like “I don’t want to use it. It’s a word calculator.” Or “I want to use it, it’s a word calculator.” I don’t want to think about this problem myself. I don’t want to do the math. I could just tell it the problem; my space, my requirements, and it will just – it will tell me probably too much, but it will give me a pretty accurate answer. I’m like “Let’s just try it”, and she’s like “Okay.” And so I tell it “Hey, ChatGPT, I have this eight foot, five inch wall in width, and I want to have these pictures laid out in a grid. They’re 26 inches squared, 26 wide, 26 tall, and I want to have them evenly distributed on this wall in a four grid.”

It gave me the exact answer, told me exactly where to put them at. We did it in five minutes, rather than like doing the math, and making a template, and writing all these things on the wall. It was so easy, because it gave us the exact right answer. That’s cool.

Justin Garrison: Yeah. Maybe Excel is too difficult. If I go look at Excel and I’m like “Actually, there’s too many functions here. I don’t need all of them”, I can hide most of them. But at some point, I just need like maybe a calculator. Because actually, what I needed – I didn’t need Excel to do that thing.

Jerod Santo: The way Simon Willison described it really resonated with me. He called it a calculator for words. And so it’s going to be very good at taking words - you can put a lot of words into it and have it summarize those down, compile them down into less words, and like Adam just said with the smart diagnostics, “Tell me what this means” or “Highlight the difficult parts”, or whatever. It can also take a small amount of words and expand them into much more words. And those two use cases - and there’s many permutations of those - are hugely valuable, beyond just Q&A.

Kris Brandow: Yeah. I’ve ran a few little tiny plugins to do things like converting between binary and decimal, or hexadecimal and decimal, stuff like that, which is just like – I hate googling for that, because there’s not a Google calculator that just does it.

Kate Bradley Chernis: Yeah. I mean, I think my first advice would be just calm down, everybody. Hold on. So think about humans - we’re the only mammals that when we come out of the womb, we’re completely helpless. We can’t feed ourselves, can’t defend ourselves, can’t stand up; can’t even hold our heads up. Nothing. If AI was a human, it’d be about three months old. So I think that’s just one important thing to remember, is that human guidance is required.

Certainly, as we talked about at the top, the prompting expertise is going to be something that people are looking for as well, because in order to get the robot to do what you want, it has to be asked just the exact right questions; what are those questions?

With technology, as you guys already know - I mean, technology happens; it’ll continue to happen, and it’ll continue to improve lives and replace jobs… And jobs will evolve. And we’re in the same boat here. This strategy piece is where humans are still really absolutely necessary. AI is not sentient. Unfortunately, Hollywood really has given us a grave misunderstanding of what the definition means. I think it maybe it was Paul Roetzer from Marketing AI Institute, but they said “AI still lacks emotional intelligence.” Which is also true, you know.

[38:22] And then another quote, and I don’t know who said this, but it’s “If you’re not using AI, you’re going to be falling behind.” But the people who are using AI in tandem with the work they’re doing - just like that calculator we talked about before - they are the ones that are going to be getting ahead of the game. So you have to figure out a way to obviously embrace it. And I think it’s so funny that companies are abandoning generative AI, because remember when everybody banned Facebook? I think there’s still government entities that can’t use Google Docs. And you’re like “Are you kidding me? Get with the program.”