I Fixed My AI's Biggest Problem

Hello. All right. Oh, okay. Let me just close this stuff. Um, okay. Okay, so today we're this is the third stream uh for this recreating Alex Liberly script series and uh you know because my friend complained time that it was too slow, I made some decisions in advance. Uh I just actually started doing them now or like 30 minutes ago when I scheduled this.

Uh, but I've I've come up with some decisions that I think are okay at this tutorial at least. Oh, that's a that's another video I'm preparing. Okay, so let me close that. Let me close that. Let me close that. Okay, cool. So, here's what we're going to do today. goal is really just to mostly just to make the scripts better.

Like last yesterday we found out that it seemed a lot like the reference script. It was uh what was it doing? It was using a lot of the same words like it was using social the word social a lot which was like present in the reference that we used as a cheat or time hack uh to be able to come up with something quickly.

it was able to come up with like, you know, good stories about the topic and good points, but yeah, it just sounded too much like the style reference. And so, I think the uh all the scripts are going to be a lot like the style reference. So, that's why we're trying to improve it today. So, the uh there are many options to do that.

The first one I want to try is just like changing the prompts, right? Because if we can do that really quickly uh and just be done with it, then that's great because it's low cost, low time cost at least. We have other options as well. Obviously using new data at trying multiple prompts, multiple LLM calls, multiple agents, etc.

By splitting it up, usually you're able to get better performance because the LM is better able to focus and utilize the token window. And then, you know, to actually be able to evaluate whether the script is getting better because we need to do that otherwise like we don't know where we're going. Um, oh yeah, right here.

There are multiple ways to do it. Last time I just got like a feel for it just by reading it quickly. I was like, "Oh, there's a lot of the word social. I remember reading that style draft script." But that's prone error. It's not really scalable. I don't have that time to do that. Yeah. So, what I'm going to do is try to focus on some like warp frequency data and visualize it, etc.

We could obviously always like look at a um making another AI to do it, but I think it's I think it's critical enough that it warrants a human looking at the fine level data. um data we can actually focus on. An agent is only as good as we can make it, but data can be as good as the person interpreting it and the quality of the data.

And then, you know, I kind of have this as a one-off. Maybe I'll just move this aside and we can do it we can do it in a later stream if we don't have time today and actually just focus on the other stuff here. Let me just move this. Yeah. Okay. So, in improving the script, it would be nice if we were able to give it some feedback to make it better as opposed to just like manually changing ourselves every time, changing the prompt.

If that doesn't work, one way to do it is just like, you know, add comments. And so, that's how I'm going to do it. And then finally uh to have that to actually use that feedback. So this is just like how we store the tape feedback and type feedback. To actually use that feedback you have many really wide options.

Some of them are really expensive like changing the vertical. So you change the base model with supervised fine-tuning. It's very risky, very costly. Uh and then you have other options that are really cheap uh which is just modify the prompt by feeding it the feedback. It's the comments here. Previously we saw we saw this but we didn't like it.

Put that in the prompt. Second option that we have is reinforcement learning. That's going to be expensive. So I'm not going to do that. Another option that we have is rack integration. So retrieval augmented graph. um that is really good for like being able to get to reduce hallucinations, but that's not a problem that we have now.

We actually want to improve the quality of the scripts. So, we're not going to do that. And then like if you're going really out there, the other option that we have for improving the script or make making it improve itself continuously is using a genomic style tournament. So basically you have a tournament of agents where they're like um competing and the the best one wins.

And when you say genomic style is like you're basically allowing it to mutate like genes or a human is able to mutate through generation. And that's sort of like a evolutionary kind of thing. So I'm I'm just going to go with the lowest cost fastest option which is modify the prompt using the feedback.

All right. So that's that's the point of this stream. We'll start by improving the script output by prompt fine-tuning. So this was like the not user friendly interface that we had. We actually had a nice user interface on lovable. So we're just going to open that. Oh, it's opening the wrong window. Come on.

There we go. That's for another video. Okay. Medical scripts hub. That's the repository that we had last time. Okay. So, this is where the website was published. Uh, we just had publish on this lovable website. We're just going to use that. It's just easier. Also, sorry if I'm grunting a lot today. I kind of pulled my lower traps while running.

Believe it or not, I really like running, but maybe I should start warming up my back when I do running. Okay. All right. We want to be able to open it and be here. There we go. These are the scripts that we had. This is the last script that we published. Uh, I just remember we had this little formatting issue, but that's minor.

We bigger issue is that the script itself kind of sucks right now. It was nice and long, but yeah, have that issue with social, I think it was. Yeah. Nope. All right. Social. Yeah, it just keeps saying social, which was a lie. It was really impressive in the reference scripts. Okay, so what we're gonna do is Okay.

Yeah, we're just going to modify the prompt to tell it to not basically we're going to explain what the problem is and tell it to not do it. Not I had that problem. So avoid adding a really high frequency of a rare word. I would say that social is a rare word and it shows up a lot in this script. Actually, that prompt sounds problematic.

It might be better to go for something like Yeah, sorry. It's going to be It's going to be problematic because there are times where we want to use this rare word over and over again. We just don't want to use the same one from the same script. So, we'll just clarify that in the prompt. I'm going to put it right before the output this because that's going to be more heavily weighted than anything else here.

Uh, okay. Over here getting some text. Sorry about that. Okay, so Oh, wow. What the hell? Is it listening to me? How does it know that? Maybe it's reading my my clipboard. I might have copied it or something. How does it know social? That's crazy. Oh, maybe you can just like figure it out from the from the style task, which is also crazy.

That's pretty good. Okay, avoid. I'm not gonna say that because that's just way too specific and limiting. We can have the word social in some scripts. We just don't want that in all the scripts. So, avoid using uh an abnormally high distribution of a word in the style transcript. For example, don't say social too much just because it's in the style transcript.

Style reference transcript. I guess it's not really a word. Uh h I guess I could think of it as a distribution if it's like look thinking about it in its network and how it's going to generate all the script scripts. No, I think it might be better just to like think to to for it to treat every script as an independent because that is what it's going to do.

It's not going to look at the previous scripts as it's making a new script. This is really hot here. It's almost uh yeah, it's getting summer and we're still waiting on our AC to get fixed. Should be fixed on Monday. too high of a frequency word from the style transcript. Don't say social too much a lot.

Um, don't see social a lot just because it's in the I'm gonna actually be more coded here so that it knows what I'm talking about or that it can think of it as sort of like a variable. I have a very specific meaning when I say style reference transcript and it's that. Okay, so this is a rule. These are just rules here.

rules. One two. Yeah, actually that's not really a rule. That's just the output. Okay. Style reference transcript. Maybe I should make it even more clear here. Style reference transcript. There we go. I'm going to delete this old prompt because you know we have this working. I just never got around to deleting it.

Okay. Specializing medical and health YouTube videos. Write a 6,000 8,000word transcript for new video titles with the title. The script must be entirely thought out. Uh uh doesn't really add value telling it what the who the writer is and the basing of the following transcript entirely in the style tone and pacing of the following.

There we go. That's pretty good. Or it looks good. I don't know how it's going to perform. Let me just like save these scripts from last time. Python script to push YouTube scripts. Yeah. And then I think it was just like basically generate another script. I didn't already modify this before then, did I?

Yeah, I didn't. Okay. And now we're going to fine-tune the prompt. Basically, try avoid Yeah. Uh try avoid sounding so much like the reference. Yeah, this is not really consistent styling but actually I do have a problem with it. So, I'm going to fix it. There we go. I might actually change the commit once I know it's actually working or not.

Okay, so we've got that. Now we need to actually generate it or run it. Uh yeah, Python generate. It was what was called Dr. Alex script. Okay. And then we need to put in a title, I believe. What's the title? Oh, yeah. We're just We were just using the ones from um from YouTube account or the YouTube channel.

Oh, that's my video. Dr. Alex. There we go. Okay, let's do this one. We did I already do that one? Nine everyday foods. Yeah, I tried to do that one. That's surprising. I've never seen it not make a script that long. Like, it didn't make a script today. That's abnormal. Very strange. Maybe he knows we're on to it.

Okay. Now now let's not be too speculative. All right. I wonder if he's working on the burstiness of the video. If you watched my last video, you know what I'm talking about. It's so warm. Okay, so it's generating the script. Let's go back to our little or I guess our big plan. We're prompt fine-tuning.

This doesn't work. We could, you know, move on to the next steps. Actually, we're going to need to do this anyways. Okay, sorry. I got to I got to answer this text. Just one sec. Okay, I'm back. Oh god, I meant to sound a lot more excited. We're back right here. It's okay. Now, we said we were we could I mean, we're going to need this anyways, the word frequency visualizer.

So, I'm just going to store that while I uh or before I like actually look at the script. Actually, let me just push it. Python uh was it push script? Yeah. And then it's like a file name. We need we need to refer to it by name. It looks like it didn't save it. So, something went wrong. Did it log what the error was?

invalid. I wonder to be honest. Like I could try to fix it, but we could just think of it as a oneoff and not spend too much time on it because we're going to review the transcripts anyways. If they don't push, they don't push. Uh so yeah, I'm going to have to try that again. Then I'm going to do what I told you, which is create that visualization.

Okay. Create uh functions uh create um create a model that I can use to analyze the um generated script. JSON files most recently. JSON with fields transcript topic title. Um then use that module to visualize the Yeah, for now the module only needs to analyze the frequency of words per transcript. use pandas.

Uh, I'll put that in the beginning. Then use a Oh [ __ ] Okay, there we go. Then use a model to visualize the output of in a Python notebook. Do not put too much logic in the Python notebook itself. Only enough to visualize just the function call and One, two lines per cell. There we go. Oh yeah, I put it to planning.

That's pretty good. Now, where are we with the script generation? Still generating it. Okay, I'm just giving it I'll just be back. I need to use a washroom. Okay, we back. I hope we aren't waiting too long. Okay cool. Let's see what anti-gravity is. Oh, wow. It's still generating the script. Is it saying what's going on here?

I hope it's not thinking too much. It's probably thinking too much. That could only be it. Unless the EPI is really slow, which is also possible. Yeah. Okay. I guess that's an object-oriented I'm I'm not sure that's really an object-oriented approach. Uh but doesn't matter. Okay. Script analyzer. Oh, I guess cuz I said module.

It's crazy. Okay. It's a Python notebook. Sensing it by phrases instead of actual words. It's got this. Okay. I like this. Right? That's what I like to do in in notebook because mostly because these agents are not able to edit notebooks well but notebooks are still really good for visualizing things fast and so this is the balance that I have to maintain or strategy.

It's got like a nice little function here. And then it's already updated my get recommendation text. That's pretty good. It's verified the plan. Okay. Sure. Okay. What was the plan? So this is the option that it chose object oriented approach. It could have done that. It's lightweight, easy to test. Why did it do it?

Okay, not do it because you can't put state, I guess. Or it's not usual to put state in functions. You could just keep them as floating variables in notebook. But I feel like that would actually be probably better for notebook. But this decision is so unimportant what we're actually trying to achieve that I'm probably not going to fight it.

I'm going to try to modify the instructions though because this has been like I just don't want to fight it and uh right now it's just, you know, giving me too many things that I want to fight it on. Okay, I'm just going to hide this because I can't remember if I have any like specific or sensitive settings. Okay.

All right. Not that. Okay. Yeah. I don't know when I've changed it to always proceed. I guess it has done that. I'm going to put Unfortunately, it doesn't have an agency sides for the terminal because that's the one I would like it to decide on. I guess that's not very secure, but like the having an always proceeded option is not great.

But having that off is just way too slow. Okay. So, I changed that. Let's see if that takes effect in the next one. I might have it get out of the object oriented approach. Next one. Oh yeah, I definitely did not want to do that. That's way too overkill for this small project. Okay, so it chose chose option one.

Yeah, I really wish it would chose option two. Whatever. Yep. Yep. So now let's actually look at this script. Might actually be readable. What kind of state is it keeping? I guess just the data. Yeah, it's really unnecessary data frame as well. You said the data what does the data have is using okay so it's using self it's using the path it's always loading it on initialization oh this is like really ugly like it's not even passing so you have to have this in this order otherwise it won't work that's like not obvious at all which is a problem.

Okay. Uh, whatever. Some things aren't always great about object-oriented programming. Okay. So, what's it doing for calculating the word frequency? Oh, hey, this is done. It's got so many lines. Did it just like spend I wonder 100,000 tokens at the end all just with new lines. Let's see how far this output goes.

So we have 300 kilobytes. The last one was 15. What's uh the one that we did yesterday that was like 5,000 words. Oh god. So this is like 300 time or 30 times more 5,000. I don't know what happened. I can't even read this. Well, let's try to read it cuz I am curious what happened here. We are trying to improve the script after all.

So this is worth our time at this point. Started doing it. Okay. So, what happened here? Maybe we can prompt it to close this up, right? Uh okay. Your mitochondrial. Thank you for the attention. And then new line, new line, new line, new line, new line, new line, new line, new line, new line, new line.

Why did it do that? Let's see what the thinking says. Looks like it didn't think too much. Let's begin. Uh okay. This is unexpected again. Maybe it's because we've given it too many words to work with. So, what I'll do is I'll just add this rule anyways. We're getting stuck repeating the same tokens at the end of the transcript.

This should not really I should not really have to do that. should be part of the the model being used. We might have to switch the model eventually. We just decided to use Kimmy. Okay, we're going to we're going to change the max tokens. What was the output number for this one? This one was a good script.

The one that we had yesterday. Oh, wait. I guess it's this one. No. Yeah, this one. Do these eight movements, right? Yeah. And that one has the date time stamp 2211. Okay. So it's this total tokens that it used about 12,000. So let's limit it to 30,000. So this is 3,000. This is 30. I don't know why I have it at a as a high temperature.

That was the default, but that's not really something that I expect to get a lot of positive change from changing or positive effect. I'm just going to focus on the transcript or the prompt. Okay, so we edited this. I guess we don't need that rule. Nah, that's still useful. I'm gonna change the prompt like this so it doesn't have a new line at the beginning.

That's probably what it did. It actually told it to have a new line at the end. I guess just noticed that. But basically that's what's interpreting because it had a new line here. Okay. Let's try that. Sometimes a little change will fix it. It's always hard to tell though because once you have one change and you only test it once, you can't really make a conclusion.

Yeah, you've already proceeded with that. Are we still looking at that script analyzer? Okay. Lower case. Yeah, you want to normalize it when we're counting. Removing extra letters. Filtered stopwards. Yeah, we don't care about the frequencies of the stopwards. Yeah, these are fair words that we don't care about.

It's already done. That's great. Okay, let's just finish reading this so that we're not context switching too much. W frequency and we get this frame at the end. Okay, this is dynamically calculated. Oh, this is okay. Yeah, this is uh brutal. It's not even this I don't like this is what I don't like about like some object-oriented architectures is that we're using this variable here and just modified it here.

But like you can't really see that just because like it's attached to the state. I guess I gota get I just got I guess I just got to get used to that. when I'm looking at auditoriated code. Okay, let's run all that. Oh, wait. Let me just accept that first. Exit to here. Oh, accept it. Oh, this keeps happening.

Eh, we lose the interpreter. Oh, you can tell by the aid that I grew up in Canada. Okay, I'm just going to close this. Start it up again. Looks like there's a problem with this one as well. What happened here? Should say I feel like this is way too long. the way that it's logging these errors. Well, at least I can see it.

It's invalid JSON. Okay. Where is this? Failed to parse. We'll see if we can uh yeah print f line that failed to parse was print specifically. Is that going to do it? I feel like this is going to fail now. Can't believe it's using this exit. I guess we're just scripting. Okay, let's just restart this because I want to be able to get my Python interpreter on.

Yeah. Okay. Selected the right environment. That's great. Now let's run that again. I wonder if it has to do with the title. Okay. rather look at this one. Never push the metform in one. Okay. So the title of the transcript is movements like it includes the word movement. So we'd expect movement in there a lot minutes walk five day one.

Okay. Okay. So, we're going to do we're going to also like do this same thing for the um for the reference Oh, I guess we don't really have the reference like that in that format. So, what I'm going to do is I'm just going to make it that format so we can analyze it. Okay. What was the file name style transcript?

Generate script. Style transcript. Transcript path. Transcript path is this right here. You see right here? Oh god. Okay. We did have a words counter that kind of helped us with this. Let's do that. At least got the words. Have the words. What was this? Baby spill something through there. Okay. Yeah, we just want to be able to get the transcript.

Might as well just tell it to do a different one because again, we're using this like scripting. Yeah, this is kind of important. generate a uh phone script that will convert into a JSON with one field transcript that has all the text concatenated. This could be a flash job. It honestly doesn't need to plan it.

Yeah, this should be fine. Okay. Okay. Yeah, that's great. It's actually converting the script um that I want convert it. I don't want it in transcripts. I want it in the same directory as uh these guys which I currently have in the roots. I'll clean that up later, but I don't want to have to mess around with this with the path logic here.

So, where did that go? So, we'll accept it. And then we're just going to make that change. It's not this one. Okay. Oh boy. Okay. So, Python concatenate trans. There you go. Just like that. Okay, great. So, now we'll just do we'll take that and visualize that. Excuse me. Where is so is social here? It's not.

Maybe it's because it's not top um I don't really want the top end. I kind of want all words that have a frequency over five. I don't really want to change logic. That will be unnecessary. Or we can just do this. I expect social to be here. Are these searchable? Not really. These are not bad. Okay. You can see in this visualizer that we don't actually see social.

Do we? I don't think we do. Okay. It's interesting that it still says that this is new still. All right. So, where is the script analyzer? We want to be able to get arrange and visualize that one at a time. I am really feeling that epoch today. I ran for 65 minutes track of running 4,000 calories today and I've eaten how many have I eaten?

3400. So I'm going to be about I'm going to be about 100 short today. That's okay. I can't eat anymore. eat a lot, especially for breakfast. So, how's that going to work? Start end top end. Okay, that works with me. So, basically just allows us to admit some of them. That's good enough. So now say now we have top 100, right?

Remember that was uh might have to reload this. Hold on. Start. And yeah. Okay. So let's reload that. That's pretty cool. Animated backwards compatible as well. Okay. Where is that word social? I was expecting it to be in here. Isn't it in here? Social. I guess it's in here just five times. We're not even scratching number five.

Okay. So, top. Let's say 200. And we'll No, we'll start at Yeah. 150. Okay. We finally got social. listening to like it's like 150 basically. I feel like that's an unsafe range. So, here's 125 and then we'll do 50 uh 175. It's not going to look social. We should have social right here. Okay. This is This seems like a fair range or like a work a range that works for this.

I'm not sure how how we'll do all kinds of anomalies. We'll have to like Yeah, I guess this is the kind of thing that we have to give specific feedback on, but this is a good start. doesn't have a really high high frequency of words. It's pretty evenly split. Uh where does it say? What if I do? Because this was the one that was problematic with the social.

Come on. Come on. Okay. I guess I want to be able to see dog 50 really. Okay. Yeah, let's just do that for now. Okay. Yeah, cuz like I can see that's probably like a good thing or a better rule is to like focus on words that are just like have a frequency of about five or higher because they don't really know I didn't really notice social till it was at about five.

Okay, that's still a problem. Let's say top 100. Do we get down to four? Okay, we'll get down to four. But this is pretty unreadable. Oh, okay. Let's see if we can make this vertical loud. Um, bigger currently block words. Use a flash. It's hard to read and there are too many words being viewed. Allow the figure to be longer to fit to the number of words being observed.

Yeah. vertically. What's it What's it doing, I wonder? Okay, it's going to ask me to do it. I don't care about the exterior at all. I hope it went with option one. Yeah, This is way better. See, we didn't even need the uh We didn't even need this. Look at that. See? I guess it's harder to see the X-axis.

Whatever. I can just like remember where it is. Oh yeah, but it's harder to see. I guess this is about five right here. This [ __ ] right here. Is that five something? Breakfast. Coffee. Yeah. So this is the reference transcript. talks about coffee, breakfast, cortisol, inflammation, microbiome. These are just like trend words now.

Okay. Collect my composure. So, we're have about five. Where was social? Social should be social wasist here. Come on. That's Michael Jackson. Okay. What do we have? Top 100 social. Okay, it's right there. Movements make muscle uh make sense. Muscles longevity morning maybe indexing on breakfast from earlier.

Yeah, this is not it's mostly the social that tipped me off and I feel like most people won't even notice it, but I do I do want it to be good enough quality. Um, so let's see what happened now with the with the most recent press. Hey, it actually did it. That's convenient. Let's just commit this visualizer.

We didn't get to use this part, so I'm not going to commit it. I would like to see it action, though. I should have committed that one here. Oh, I staged everything. It was all staged. I've been staged. Look at commit. There's a lot of crime terminology with Git now that I'm noticing. It's been staged.

It's committed. What else is there? That's criminal. Okay, that's distraction. Oh, that's also part of this. Okay, these are all part of the same thing because I was just trying to fix this metformin script issue. Find prompt to make it not crash. Okay, wait. Avoid crashing because silly leather. Uh on I say always reduce Yeah, guess I didn't really address the JSON there.

So, I just removed that from the commit. Okay, so now we have to analyze the script. Let's look at the how long it was or like the frequencies. Uh this was the file name. Come on. Shimone. Where we at? Is it this baby? This time today we're on the fourth. Yeah, this is the one. It's the only script we've had today.

Okay. Where do we go? Where do we go? Let's visualize it. I had the other one. Well, it doesn't matter. We wanna add another one. Let's go. Okay. Add another one. Sure. Sure. Okay. Let's close that back. Open it up again. Okay. just printing it wrong. I guess we'll say that this is the uh before prompt modification for uh what was it award frequency and we'll copy that duplicate it.

And we probably need to rename this. Uh yeah, I'll I'll finish doing that and then I'll start making an experiments folder. Okay. Rename. Well, we just want to talk to the name. Where's five at? It's about here. I Well, I guess let's look at the long ones. We expect metformin to be there. B12 drug longevity risk exercise taking lifestyle metabolic let's just collapse this and be able to compare it to uh come on collapse.

Okay, there it goes. So, this is the the one we don't want to copy. And then the one that we have is at the end. This is collapsed. See, there's a lot not a lot of B12 in this one. So, that's good. Is metaphorin specifically related to B12? Is that why? Okay. B12 deficiency drug longevity risk exercise taking training I don't see much that's in common which is good we just want we don't want it to like you're repeating too many things insulin is probably here the reference style I feel like it's in every health transcript ever or YouTube health transport ever.

It's just common topic. It feels so monotone. It's probably just like Yeah, that's whatever. I'm just trying to keep talking, I guess. Okay. Consistently. Kidney. Okay. Yeah. See insulin is there. Diabetic fasting mice. This is a mice study supplements. almost sec. Gosh, I feel good about this to be honest.

I think this option one might have worked and that's great because that's the shortest one. This is the beauty about LLMs is that you can just basically modify their output just by modifying the prompt, right? It's so convenient, so low cost. Okay, let's begin organizing the uh experiments. A experiments.

Yeah. I think this will do it. Oh, I've experienced this before. Okay, I'm just going to do this offline. We'll keep it the same. That's not a fun problem to tackle on stream. Okay, we don't want to modify that. We're done here. We want to be able to push this so that we can read it in the new thing.

So, Python push script file name. Great says it was pushed. There it is. This uh yeah, this topic is way too long. It's probably long. It's longer than the title. So, we wanted to have something more general than the title. This is nice and long. Not sure if this is like Alex Wly style the channel to uh to avoid liability.

I guess that's why They would do it. Gosh, I could I really want to be able to fetch the original transcript so that we can compare it. I could do that and that could be a form of feedback that we use. So we did these two We're now in the feedback stage. This is actually be nice good idea. So, let's add it.

I guess it's just an option here. We'll do both those. Which one might be more interesting? I think this one will be more interesting because we can, you know, we can judge whether or not it's actually the same script. And if it was actually the same script or you know if they talked about the same thing like this AI and that AI that's pretty good evidence that it is AI.

So, I'm just gonna uncommit this, push it so that uh lovable can it's lovable is going to be ignoring it because it's not in the medical scripts hub. So, what I want to do is basically I want to be able to compare add the generated scripts against the reference script. reference scripts. Uh we'll probably want to have to a many to many relationship with the reference script and the ones we add because uh we might want to generate the same script for this multiple topics or multiple scripts for the same topic, same title so that we can compare which one's the best.

Well um what should we do? Should we key it based on title? The title could change, unfortunately. We'll have to key it based on Yeah, based on the ID within YouTube. Okay. could put it in the same table. We could put it in a different table. Put it in the same table then. Uh ah I don't see why we just wouldn't put it in a different table.

I'll just put a different table. It's a different relationship. That's a good enough reason to create a different table. Uh create a new table. How are we doing by the way here? Nice. Create a new table. Original scripts similar to scripts for both tables. Add. I want to make sure I get the uh the right name for this.

I'm going to pull it out from the uh the repository that I published the most recent thingy on. So, I have to hide just for a sec. Okay. Go to death. Go to What's it called? It's called scripts. Was it the doctor scripts? No, it was in the papers repository bearing doctor scripts. I'll just open this.

Okay. What did we want here? Wanted the actual scripts. Oh yeah, we wanted an ID of these scripts. What was all the metadata that we're getting? So, the videos had IDs, but how were we getting these IDs? It's a video ID. attribute on the results from the transcript API that I used. Okay. So, what we'll do is Oh [ __ ] I just realized I need to change the light for my for our gecko here cuz it's night time now.

8:30 is going to do that. You want a sec? We back, fam. Oh god, I can't believe my back hurts me so much from running of all things. And then usually the day after running now I have deadlift day because I like to rest from running on that day. It's going to be well I was actually able to work out with the pain last time.

I guess it doesn't use the lore traps that much but whatever. Hopefully I don't get injured. Okay cool. So, we're sharing this here. This is a mess. You can ignore it. But yeah, we just basically want to do it by vid ID. Did I close the other one? Oh, no. It's just here. Okay, that's kind of fun. So, here's what we got to tell it to do.

Oh, yeah. We're talking to lovable lovable. Add a new field API vid ID. And then M2M relationship between those tables. What should we do it based on? So, we're just basically creating scripts and they're all should be just related. I guess we just need the ID. We don't really need any more information.

So, what's our query pattern going to be? It's going to be we have might be able to fetch out either one. Um, make this Yeah. Do not optimize querying for this yet. Yeah, cuz sometimes loable over over optimizes for both. Yeah. where its existence means that scripts is an attempt at producing something like original scripts.

Okay cool. See how many tokens this will take. We started with five, right? Yeah. All right. We're watching you, man. Oh no. you use this equal format like that is it is good generally to use snake case but which is this type of thing but I wanted it to be vid ID the same way that vid ID is producing it yeah Sure.

The field should be No, this is kind of annoying, right? That it did that by itself. Even though it's pretty clear. Yep. just wasted five credits on this mistake last time. Okay, let's just check that sequel. Sure. Sure. Sure. That's all we need. Okay, that's good enough. Maybe while waiting for that.

Oh, it's already done. Okay, I guess there's nothing to see. I have to address that modified the database until we get an error. It's just a, you know, it's not a production thing. So, we're just going to we're just going to trust it until proven wrong. Okay. Yeah. Not do that for a real application.

All right. So now we have we have that. We want to be able to just like Yeah, we want to be able to preview it. So I'm going to tell it when viewing a script. Currently on the front end we have a script detail page on which we can view scripts. When that script page is opened, run another run a request.

I don't know. Check in the database to see if there is an original script with the same. If so, uh, show it on the right side. Otherwise, um just uh adds Yeah. Add it to host at the top right. I add a toast at the right side. I just add a toast saying that 5second toast. No. Three second saying that the script did not have a uh okay a uh script a Dr.

Alex Wberly sketch with the same title was not found. Yellow toast morning toast. Yellow morning toast. Okay. Bam. This should be pretty interesting. It's interesting because I didn't read the code here. And so I'm curious to see if I was able to make a good enough prompt without reading that information.

I would not do that on a real project. Okay, we're going to have to pull up the changes once they're done. Actually, I might as well pull these changes. It just means like bring them from the remote repository. Repository. These are the SQL changes where we added those things to the database. So, we're pulling it as it comes.

Oh, okay. It's done. Wrong repo. Wrong repo. Which repo do we go? Cuz it's not Yeah, that's okay. Cool. I just want to check that it's pulling it right uh in the pattern that I expect. So, we've got the ID. Okay. It's created a hook. Okay. Sizing it based on whether there is a inside call original. I guess that's fair.

The table's called original. It's uh repeated some of the stuff here. I guess that's fine. We don't know if they're going to differentiate. If they're going to differentiate, then it doesn't make sense to just refactor them. Okay. So, for this one, I'd expect no alert. I mean, I'd expect an alert and no original.

So, what's going on? Did it just not download the repo? Preview. Should be able to preview it here. Hold on. Okay, we're not getting the toast and we're also not getting an original. So, let's see what happens if we run it locally. Be honest, I haven't run it locally. So, I have the keys set up right.

I guess it's all public. Okay. NPM run. I can't tell. I guess it's using V. Let's Let's look at the read me. The [ __ ] Okay. Package. There you go. invite build. Okay. MPM run build defic. Okay, that should be right. Oh, I didn't have installed. Okay, install. Okay. Oh god. Fall back. Don't run too hard, kids.

Is it not going to run it? Okay. I guess it is building. It's not actually developing. Stick it out. What's it say? All right. I don't see anything in the that log. So the question is now is it using the stuff script page console war I guess we can actually just log I'm just used to writing war script page uh use effect Okay.

Wait. Is actually like to have the variable names. What was it doing? Original. I hope it figures out this is how I want to log it quick. It's got the script. No original is original fetched I guess is original fetch is might not be evaluated to what we expect. Yeah. or what this function expects. Yeah, it's expecting it to be true or like tried, right?

Is it not actually fetching it? When does it change that use of the use query? Is it just a react query hook? Yeah. Okay. What other keys we get here? This error is loading error is refetching. Okay, that's nice. Any idea why is our original fetched is always false? Do you expect it to be false if it tried to load but failed?

Wait a minute. Do we not have a transcript? Oh, yeah. This is null. That's what we expected. That could be it. No, stop it. Let's see. Yeah. Okay, so this will have some more formative errors and that should be for the case that we're experiencing here. No comparable scripts. That's exactly what we expected.

It's doing an error and that's what it looks like. What's the warning look like? What is this library? Okay. Is it at least 3 seconds? Okay. I'm not sure why it's like Do I need to download a color library? Maybe it looks better on the thing. I usually put a different commit, but again, I'm only looking at this repo.

Oh, I got to remove that console log. Let's show alerts when there is no grass API vid ID. Okay cool. Done with that. Now what? Oh yeah. Now we want to be able to attach transcript API v ID scripts. And we want to do that as we're generating right? So right now we're doing it based on actually this is the ID right here.

So we can just Oh no this is for the reference one only. Uh we don't really have the ID here. Maybe we can pass it in. That would make things easy. Yeah. So, we're going to add the original one art add argument. Oh god. Yeah. Wow, this is great. Okay, cool. So, how do we get the the orgs? Have the transcript path.

I'm going to have the same format. Load transcript. Okay. Generate the script. automatically push both of these. No no. Oh, and we basically just want to pull it from brow script API. Make sure that it's like the same in the same format format that can be spat out. Okay. Where was that script that I This is the stream.

Yeah, this is the stream. Oh wait. No, this one. Oh, man. How many do I have open? This the one. Okay, now I could just duplicate this. Yeah, I need to refactor it. Um, I'm going to use that as context. Actually, that's too much. I can just help it out with reading a little bit. So, we get the video IDs and then this is how you fetch it.

Where were you putting the uh I guess in the headers. Yeah. Okay, let's go back. Not this one. Yeah. Okay. Fetch the transcript using to get these Yeah, we don't need to sleep. I mean, it doesn't need to sleep. Uh, okay. So, we get the transcript JSON. Okay. modify fetch transcript so that it outputs a JSON in the same format.

So that it also outputs as uh generated script. as the generated files i. e. the same fields where transcript is the entire transcript. add an function within to do it and then call it at the end of Yeah, call it uh after saving generated transcripts if it exists. But only If also add an early exit within if the output already exists.

using the input. So let's uh Okay, so that should output it and then we we'll be able to push it to our database using push scripts. We need to add these keys. Okay, I'm going to need to go get those. I'll just turn that off for a sec. Just me. Okay. Let's get Wonder if this is what other streamers do with me.

I don't really watch a lot of streaming. Okay, right now I'm just adding my uh my keys to my environment. That's why I'm hiding my screen. Is my audio still going? Hello. Yes. Okay. Okay. Also need the transcript base URL. It's just this one. This does not need to be protected. Unless they want want that.

I doubt they do. They don't. They want business. Okay, cool. Okay, we're now ready. It looks like uh Gemini is done. Looks like okay. So, still doing that. So, that's good. Yeah, this is early exit that we asked it to do. Uh the full text and the title. That's how I wish we saved that. And how did it get the full text?

It's just stripping that. And the transcript data is also saved. Hopefully this data. Yeah. Still right in here. Same way that it used to be. Okay. Okay. So, we'll accept that and then Yeah. Okay. Yep. Yep. Okay. Good job. We get a log an error if there is no input original transcript ID. Just be careful.

Oh no, it's mandatory here because don't have a default. Okay, that's great. also get and save original. Okay cool. And now we'll just we'll just save it by uh we'll need to modify this to save it to the original transcripts. Push to also enable it to push to. Yeah, it's a different table. We're calling this a main How is it pushing it to that table right now?

Got the API key. That's the payload. How does it know? Oh, okay. Based on scripts, it's based on is it all that? Okay. Besides scripts, let it be inputable to uh via main Python commands. Okay, there we go. Oh, it's planning. Okay. Hopefully, it doesn't actually ask me for this because I'm pretty specific here.

And now let's generate a script um so that we can get one with the with the vid ID copy. Oh no, I don't have the I don't have the IDs for this one. I'm just going to use ones that I already have. I guess it sucks that I've already saved this. Uh, okay. Here's what I'm going to do. I'll actually copy the files there and I'll make sure that this new check has um Wait, let's just make sure.

Yep. Yep. Now, where was I? Oh, yeah. Which of the fetch? Because I already have these. Yeah, I kind of overdid it here. I don't care about the formatted out path. I care about the out path actually. Not sure why it's suggesting that. Now, where does it look for outpire? currently only have that actually.

That's not that doesn't make sense. Yeah, this should do it. We still need this, right? Yeah. Okay, let me just copy those scripts. Oh god, it's going to be so tough. Yeah, it's going to be tough because I have to I have to parse them. I have a Yes. I want to be able to like isolate just the Wly files and uh I have a video metadata file for that.

It tells you which it is. This is my most recent video metadata file that tells you which channel made it. Make a script that'll duplicate. Okay, I'll move all the transcripts into a directory named after their channel handle. This is a not a flash job. Actually, it might be really fast because I've already told it which data.

Okay, I'm just going to copy the path to this. Bam. Flash. Flash alert. Okay, let's see this plane. Yeah, it's modified the plan before improving it. I think this is the first time I'm seeing this. Is it done? Table name. Yeah, there isn't that many changes here. Now, where are we at with this other one?

It's pretty good. How do I run this? Nope. I need to get into the that directory. comparing data. Okay. And then now Python organized. Yeah. Move zero files. Okay. Does it not have access to the transcripts directory? Where is the Yeah, that should be right. video ID in metadata. Where is getting the metadata from?

Oh, it thinks. Oh, hey. Where's it getting the video ID from? Stem. Yeah. Yeah, we'll put a warning. So, I feel like it's just not It's got to be just not doing that. Is there something that's not finding? All right, we'll just make it log. Okay, print transcript path. Transcript Bath. Yeah, it's not finding this data transcripts.

It looks like it already moved them. It must have done that in the testing. I should have just looked at this number. I manly does things. How many do we have here? Oh, I should get the info. Yeah, I really should start streaming earlier, but it's just like it was hot today. That's why I didn't do it.

I was so stressed. Okay. kind of just want to know how many files. Okay, 101. That's pretty good. We'll just copy all this content. Yeah, my uh I'm going to have to start storing things on Google Drive. They just updated their their Google Drive AI plan to have five terabytes, which is so generous. And now that means I have like 20 times.

So I actually keep a very small local disc. I have 20 times on uh Google Drive than I have locally. Okay. Where did that one go? Is it going to be looking in transcripts when fetching? Well, this is the push one, not the fetching one. Should be looking in the transcripts directory. Okay, I meant to do that.

I should have done this last time or offline. Sorry again. Wow. Uh, could you imagine if I accidentally opened the Bash RC? I'm just going to [ __ ] open a new one just in case. Yeah. Okay. So, now what we're going to do, we're gonna We added all those files. They're not being tracked. I'm tired of tracking those.

uh format it. Actually, let's undo that. Yeah. Uh allow using a different table. Okay. And now we just need to do it. Generate Python generate Dr. Alex script. We'll generate one that we already know exists. Just generate this. Uh, do I not have the titles here? Okay. Nah, that's not not a medical one.

That's the one I'll script. Nine things that happened. Great. That looks good. And then original script ID. Oh, guys, this is exciting times. We're almost going to get to the point where we can compare our scripts side by side. Okay. I might have just original script. ID. Okay, I think the problem is the file name.

One of these file IDs have like these weird uh hyphens at the beginning. The secret in the blood. These used to be so batty at the beginning. Click baity. It's always interesting. I I respect the creativeness. Oh man, I have to start over. Okay, man. I can't wait for this lower trap injury to stop bothering me.

I know I'm complaining a lot, but it's messing up my vibe. Okay. Is this one? Okay. Come on. Come on. Where is this error loading? Why can't it find it now? We didn't modify the We didn't modify that, did we? This is transcripts. Oh god. Sorry. I need to change my position and it's going to take a while because of my lower traps.

Okay. Wow. Feeling like an old man today. Oh, that's better. Not for the back. Just like for the sweat. I'm just getting railed on here by the weather. Okay, where' it go? L is this this thing? I'm not sure what the problem happened. Loading the reference. We just couldn't find the reference, but the reference exists.

It looks like it stopped working after I copy pasted all these files on there. Let's delete the copy cuz it is unnecessary. I was not finding What's the file path here? Did that change my encamp? That's still the same, right? Oh, it was referencing the original stuff. I didn't even realize that. Okay.

Sorry, I sorted transcripts. Where' I go? Just like that. Okay, cool. So, it's generating a transcript like we have before. And now it's What's the issue here? Error saving the output. Which output do they have saving? What's the What was the result? This was the result. Why did it do that? Let's write.

That's how it ended that Why is that so dumb? Blah blah blah blah blah blah blah blah blah. Okay, I'll give another chance. These things happen. If I'm going to deploy it, I'll probably have it do it multiple times. I could really go for some um salpa. We're just waiting. Okay. What's the result now?

That looks good. That looks good. That was good. Did that. I'll save 219. Okay. So, this actually did fine. The problem has been with this other thing. Success or after success. Let's send in the fetch unsupported operant type. Okay, we're just going to skip generate transcript this time. So, we'll skip that and we'll skip that.

That'll do that. Okay, we command. So, the OP path is going to already be there, but it hasn't said anything here. So, we'll see. console. Nope. Uh, print fetch transcript video ID. Video ID. And we'll just put that here. Okay. just never made it to here. Oh, I guess it's trying to concatenate. Oh, yeah.

Two strings. It was expecting them to be paths. was expecting transcripts directory to be path. I'm just going to change it here because I don't know if I have anything else depending on it. Okay. to make the the format one Oh, Is it just exiting? No, it should be here. Skipping download and formatting.

Oh, okay. I guess it already did it, but I just have it like that. Yeah. Okay, cool. So, now that we have that, we need to be able to push that one. Now, how are we how do we configure the push here? Actually, this is probably a good time to make some commits. Okay. Okay. [ __ ] the back. There we go.

Don't want to do that. Don't want to do that. This is logging that we don't need. No longer need. No longer need. Okay. Actually, I think we already kind of tried to do that. So, I'm just going to say that's now success. Successful. And we'll copy that one. I am actually really really confused that Dr.

Alex Swimberly this is making me like this channel did not make a video today. That's very interesting cuz like every other day it didn't it never missed. I wonder if it got banned demonetized. Maybe YouTube beat me to the punch. Oh no. Okay, it skipped a day here as well. 8 n skipped a day here. 13.

Okay, looks like it's skipping some days now. I guess trying to make it look a bit more realistic, but that's no use. So, it's not publishing on Saturdays. Okay. Yeah, that could be all of it. But Oh, no. I'm hungry now. Well, it's kind of late. I should go to bed soon. Let's just try to push that and then call it.

We want to push this thing and push this Python push script. What do we need? We need the file name. How is it going to is it also going to take the path? Let's see. JSON I was going to say okay at table original scripts. Yeah, as I thought. Okay, so I need to say that it's in the uh this other directory transcripts.

I wonder if now if OS bath can handle that. Hey, it can. Okay, so we pushed it. Show me the whatever thingy. No known incomparable script. Oh yeah, that's not the script. Which one did we Because it was what is this baby secret in the world the most recently generated one. So we'll have to push this.

Maybe this is the Yeah. Not not a great experience for me right now, but whatever. Okay. No known comparable script. I just pushed one. When did we log that error? I guess this is never being this was never being saved. Uh okay. Transcript API vid ID. Now, where is this from? Okay, I guess I got to also put that in and the arguments.

Yep. Yep. Let's take that. Yep. Okay. SERs API. Vid ID we'll get the V ID should be this. Is it gonna I think I'm going to duplicate it. What's going on here? So, we have it. Transcript API. I misspel it. Oh, he's not being passed here. Got it. Got it. Yeah, let's fix that. Okay. And then you just got to send it.

I'm just going to send it. Okay. I think that's going to duplicate it. Yeah. Okay. Now, the problem is that the other one didn't have it. So, we're just going to do that. the original script. Yeah, we'll copy this. Okay. I wonder if it can handle the uniqueness. Hey, tada. We've got it working. We've got uh the one that we generated on the left.

It's a bit nicely formatted because we know. And then we've got the AI one on the right. Uh I mean the uh the from the Dr. Alex Wly channel, which I believe is AI. So, okay. Ever since I started medical school, I've been valuing by the people to make it to 100 years old. the secret. Okay. So, the sound is different here.

Because like I think this is more like what it what it has been more recently. So, here's what I'm going to do. I'm going to read these sections and then I'm going to try to find compare comparisons on the right markers. Okay. So, blood blood marker. Markers window. Hey. Okay, let's see. Protein. Okay, we got proteins here.

uh 37 DNA. Does it mention any specific proteins here? No, it's not very technical. Like this sketch on the right is not very technical. This is a lot more technical. Not really fun for the YouTube viewer. This is great though. We made a lot of progress. in the sense that like we can now have a human compare it to the the one generated by the channel and you know I can tell right off the bat is that right off the bat is that it's different and that it's not this is more technical and obviously is longer as well but this is more this is like the old scripts Yeah, these are the old scripts.

So, let's compare it to something newer. Unfortunately, there are too many steps in the generation here. Like, first we have to generate and then we push each one. At least the generation is all one step. Sorry. [ __ ] Oh no, my microphone's all tangled up under my weight. Yeah, this is really tough because I tell you I'm more fit than this.

I'm just I I tried to really So, I'm doing this running walking experiment and so I got back into running pretty fast and furious. because I want to be able to beat myself at running. I just want to be a better runner. Uh and I had lost some fitness walking for the last 20 days. Your body, you should probably go gradually into it.

Okay. Cool, cool, cool. Let's generate. So, we need to generate. The problem here was that we took something uh too old. We're looking we want to look for something more recent. So, let's look at just like 10 morning habits that add years to your life. That would be an interesting one to find right? 10 morning.

It should be there. Oh, wow. I guess my this is producing so fast this channel. Let's go. IBS still versus bowel cancer. Or maybe the title is changing. I should close this one now because it's confusing me. That's the old stuff. IBS versus bowel cancer. Do you have IBS or bowel cancer? Seven things destroying your blood pressure.

Okay. Yeah, they're really changing the uh the titles a lot. Destroying blood. Nope. Blood pressure. I want it in the title. Is this not searching files that are hidden? I feel like it's not here. I gotta modify this. Cancel. I don't want that. Transcripts. Okay. That's probably why blood pressure and I want in the title.

Okay. So, what we're going to do is Okay, let's try this one again because I feel like this one should exist. Okay, cool. So, this is a Tim warning habits one. We just need to get the ID so we can generate a parallel. Oh, I've already got that one. Yeah, we'll just do another one. I guess it doesn't matter.

Uh, no. Okay, we'll do a new one so we can compare more cleanly. Okay. Improve your kidney function in 24 hours, believe it or not. Okay. I could really go for some hummus right now. How Arab is that? Okay, this is generating the script. It's going to find that this one's already generated. We'll format it and then we'll push it.

Push it. Okay. Okay. So, this is the generated stuff. I want to make sure that get the right number. Says finished. Whenever that's finished, we'll be able to switch it over. Okay, so this one was not good after all. Okay. Sometimes these errors happen, the JSON ones. It's kind of annoying really. Oh, I'm smiling because my wife is texting me.

So, okay. Uh, it's a good It's a good thing we're waiting for this thing. Yeah, I'm too tired to do anything concurrently to be honest. We're almost gonna call this stream done. Honestly, I might be able to hit three hours. I'm proud of it. I know it wasn't like on the whole time, but uh you know, it's nice.

It's nice to be able to do something for that long. a new record for me, even if it's not really the best stream. Okay, so we got Yeah, we got the we got the right name. So, let's push it to scripts. And now we want to push original one, right? The original one should have a title like this. Okay, let's cover that.

Yeah, I mean it's not great that I'm passing in redundant variables like the vid ID stuck here and here, but I don't want to refactor it right now. Okay, original scripts. Great. Which one is that one? Yeah, that one. Come on. Be here. I feel like I have to look at it here locally. For some reason, it's not deployed.

Okay. Sold something. It doesn't work. Okay. So, now before we get into the specifics, uh let me give you let's see what the first Okay, the first habit. So, let's see what the habits are. Um I guess what's the point of that? If if they have the same habits, they probably do, but I guess they have the same habits and same high level content and then we can evaluate the low level stuff more clearly.

Yeah, but I'm getting way too tired. Okay, so the first habit uh the physiology behind it is strategic hydration specific timing. So what's this one? Kidney hydration step number five. Okay, maybe you told it to like lead with like how do you say contrarian things? Okay. So hydration these uh yeah I'm ending up with really long transcripts here.

Sodium. It's not really here. Sodium restriction. It's not on the I can mention sodium here habit. Where does it say habit here? Like how is this transcript organized? Uh first Okay. Okay. Okay. There we go. Step number one is to eliminate liquid sugar completely. sugar, blood sugar stabilization, meal composition, and time.

Wow, that one said eliminated completely. I guess liquid sugar. It's a bit more specific. We can prompt that. But let's just remember to do that. Uh step number two, restructure your meals, protein and fiber. So, so you know here steps um it's organized by steps and then here on the left is a is for Uh we actually say habit I think sodium restriction, blood sugar stabilization, movement, sleep.

Okay. blood sugar, meals, prioritizing proteins and fiber, timerestricted eating. So that's similar to what we have in the other one. Uh walk after every single meal. So that's a specific category of movement. And uh so this is actually specific medical advice to be honest because it's so direct compared to this.

This is a bit more generic. It's just telling you what it thinks is important. So, we've we sort of come up with like two differences right now. Our left one is more is more general. The advice is more general. Doesn't actually tell you what to do. It just tells you how it works or like the theory. And also I think this one is a bit more technical.

We're not seeing it now in this one. Okay. CKD CK I guess that's a chronic kidney disease. That's not very technical. Step number four. Sor hydr uh walk and then hydration. Okay. So we said I'm just going to write it down. uh walk hydration and then time restricted eating. Restructure meals. Restructure meals.

Eliminate liquid sugar. We're at one, two, three, four, five. Okay. And then let's just go through this list here. It's organized by habit hydration. H it's just okay. So, this one's actually more specific about the time of hydration. Yeah, never mind. This is not more general uh sodium restriction with meal timing.

So, you know, meal timing is the same thing time restricted eating, but it's a bit more specific here about the sodium restriction, blood sugar stabilization. So, here is general. Uh, I was just wondering if like it has something about like a a fast sugar spike. Yeah, it doesn't matter. Well, it doesn't really, but it has more specific examples that are just not relevant to that one.

uh restructure meals. So it is movement low inensity walking. So we actually have walking on both sides and then sleep optimization on this side. We don't have sleep optimization. So this one matched. This one matched. This one matched. That one didn't match. This one was close. Restructure meals about protein and fiber.

That was the only one that was not really here. Instead, we had not that one. Not that one. Not that one. Not that one. First, second. Third, fourth. Oh, yeah. No, it's sleep. It's sleep. It's sleep versus sleep. Other blood sugar stuff. There you go. Might as well be getting that advice from AI. Okay, that's it.

I think this is a good place to end the stream though. Let's just recap what we did. We kind of added this other step which is just like yeah this is what we focused on. We added this random randomly throughout and it took a lot of time but it was worth it I think because it gives you like a more viby comparison a human quality control sometimes it's hard to visualize this kind of thing in data.

Okay, next we might want to do being able to comment and then use that to modify the script. However, honestly, I think the script was fine enough. I don't think we need that. It is interesting. We Yeah, we could postpone that. So, I'm just going to make this, you know, yellow because we would have done it if we needed more time.

Uh, and we didn't have time to do this in the stream. But yeah, I would say this was a success in improving the script's quality. We also were able to have an interface where we can bear the script. Okay. Is that too? Oh yeah. Okay. add service API v ID to the tables to data to the push data and this is just really generate more scripts.

and push the original. And now we'll just push that. Where is this? Is this accessible? Yeah, you can use it if you want. Close that. I think we've concluded most things. We didn't have to modify too many things here. I'm not sure why it's not updating. I think it only I think I might have to press publish again or something cuz like here we've got this working here, right?

Let's refresh it. Ah, close it. But it was working there. Now here, look at the ID at the bottom is what I'm looking at. Okay, this is the same one. Yeah, we we're not seeing it. I think it might be because it's just not published. So, we'll just say publish. Update. That might How long has this stream been?

So, three hours to improve the script quality and add a way to compare it to the original, which that was that was spontaneous and it was important because we don't want to spend time iterating and improving the scripts if this is actually already good enough, right? The worst thing you can do is spend a lot of time and resources creating something that has no benefit at least as an engineer. I guess you could also create something that hurts which is worse.

So like AI, I mean the Dr. Alex channel. So okay, so that published. Yeah, feel free to check it out. Okay, we got some issues here with that, but that's that's fine. That's fine. No, no comparable script. Yeah, because we don't have the transcript IDs there. Those are early. Great. Is it working? Okay, feels good, mate.

Thanks for watching. Let's just close this, baby. Wait. Yeah, we definitely crossed three hours, didn't we? Yeah, by 8 minutes.