Recreating Dr. Alex Wibbereley Scripts (from scratch; Stream 1)

Okay, we're live. Might have to restart it cuz YouTube was complaining about the resolution. Let's see if we got it right this time. Hopefully, it is. I'll keep uh looking at the stream health. Okay. So, as I was saying, I'm just making a system that recreates scripts like the doctoral release scripts.

The output that we want is um Dr. Alex Wolvery like script. I'll be loose with the things that I uh the shapes that I use. We'll just keep this white because a paper is white. Okay. Am I the one viewing it here? I don't know. [snorts] All right. Well, it doesn't matter. I'm going to do this anyways. Okay.

I should be able to It feels like I don't have access to the Yeah. Okay. What was I say? Oh [ __ ] Nice. [snorts] All right. So, we'll make a uh script generator. Yeah. I mean, that's like the thing that we want to focus mostly on. And that'll just make that we can start there and have it spit it out.

We'll give this a different color. Yeah, that'll work. Look, great. We can probably just have a We can, you know, we can just straight up go to Chad CPT or something and see if it generates anything. But that will be fast. But I know that it would be better probably if I had something to teach it based um based on past scripts of Dr.

Wberly. So because I actually have past scripts of Dr. Wberly, I'm going to try to generate a local Python program that uses um a remote API. What a remote API. We'll figure it out. Okay, we'll say Python scripts. That's where we're at right now. One problem that I already foresee with this is like I said I have let's say 100 scripts of Dr.

Rubberly and we'll say 100 here. Yeah, that's good. What a good live diagram. Okay, I want to be able to use as many scripts as possible when I get them through the remote API. So, I needed to be able to pull from all those to understand the style. One thing I could do is Okay, so let's look at our options.

Option one, send all send in scripts as part of the API call. That's pretty fast to develop. Might be expensive because that's a lot. That's going to be a lot of tokens. Option two. I'll probably have to research some more options to be honest. But like is a is a rag a valid example here? I know rag is good for fetching like specific information from a tree of information but we don't really have that.

Uh we have just a bunch of information that we wanted to learn from. Uh so doesn't seem applicable. Option three. [sighs] We have now. Let's align these babies. Oh yeah. Okay. What we got? We can we can hardcode the style right as part of the co as part of the call. So use the scripts to get like a style summary and then use that in every call.

you can hardcode a style summary into the API call. So we'll be adding an extra step to get a style summary. So the con is that it adds an extra step pro is it might it would be less tokens. I am uncertain if it would be more accurate proconwise. Might it be fine to just try it with one script and see if it can replicate it just based on the latest one script.

I feel like if you give it enough like it can't yeah this hypothesis is quick to test. Can it learn the style based on just one script? So let's just try that. So that's the option for like sort of our Python script call design or how we're going to call the remote API. But then which remote API do we want to call?

So this is we'll paint these as yeah that we'll say that. trying to summarize. Okay. Oh, I forgot to write that one right now. We want to be able to get the right API. Right? So we have for example open AI together AI. Yeah, the vendors option one. I'm sure there are many more vendors now. I feel like this decision doesn't matter as much because what matters eventually is a model that we're going to use.

Other we'll say other I want to use together AI because I have credits on that from another video that I'm making. Okay. Do any other decisions we want to make? So, so far we have we decided to I guess yeah, we'll color it. So, we decided to send one script basically as part of the API call and just use together AI because I have credits there.

Okay, as part of the prompt, we also want to be able to have it like generate something kind of new. So, we need to be able to like give it a subject matter. I feel like I can just hard code this for now. Obviously, if there's if we feel like it's too similar what we've got, then I can unh hard code it and add something else.

Okay. So subject we're going to prompt subject matter hard code something about organ disease. I think I just want to try this. But just to be comprehensive, we should just quickly consider other options in case some there's something way easier. I can have it. I can basically prompt it using the title of another video made by Dr.

Alex Wberly. Use title of another video by the YouTube channel. Dr. Alex Wberly. I say YouTube channel because [snorts and clears throat] I don't think he's a real doctor. Okay. Well, the channel I don't think the channel is a real doctor here. So, it's probably still just my Okay. It's actually probably easier.

I'll just do that and it's more relevant this option right here. Okay. And then we'll be able to get a script and then we can judge it roughly to see how good it is and if it's obviously we yeah if we can identify any weaknesses then we can work on them. If not then we'll have to find a better way to evaluate it basically.

All right. So I made these three options basically. Let's just rearrange them in a way that I can use them. Okay. I'll go with that. Yeah, we'll just use one. Is there anything else I'm missing? I think for just generating the script, this is fine. We can obviously iterate on this processor system as we as we hit checkpoints.

Okay, checkpoint here is being let's generate a script because that's the most important thing. Okay, I'm in the script directory. Oh, those are my emails. Wonder if I Ah, nobody will watch that. I don't know. Wait, what's going on here? Is this live? I feel like Huh. Is this way back? What's going on here?

That's not what I'm seeing. Oh no. [snorts] [snorts] So my friends, the one watching, the only one watching is seeing a jerk. So all right. Or is it Why is it playing back the old stuff? Oh, is it having me play back the old stuff? Yeah. Okay. I think I'm way behind, which is fine. Watching me watch the old episode.

All right, I [clears throat] see what's going on now. Okay, let's do this. Actually, no. I'll give it a a script. I'm just prompting Gemini here to create basically the the foundation for this repo. Okay, I'm just gonna Yeah, I'm going to hide my screen just for a sec. So, we got this going on. Okay.

Wait, did that happen? Um, um, you should probably just say this in the chat, but [snorts] I'm just laughing because my friend made a joke. He's asking why is it that every time I uh I type the camera shakes. I know I got heavy fingers. This is a dual coding and lifting channel. So or fitness whatever.

Okay. I'm trying to I'm just trying to copy over some a transcript here that's made by Dr. Ridley. And u the reason I don't want you all to see this is because uh I will post it in another video soon. Okay, let's open up this metadata. Okay. Let's arrange alphabetically. Okay. Found the video ID starts with.

Nope. Not that one. That one. There you go. Okay. There we go. Wait, is that gonna be lost in there? Is this copied? Yeah. Yeah, it's just copied. Okay. Transcripts. Should probably also change my bash while I'm out of here. Bash. Am I using bash Amazing. Yeah. Okay. Okay. Okay. There we go. Okay, now we're back.

[snorts] Oh, I don't know. Maybe it's cuz my Yeah, maybe the thing is shaking because Let's try it here. It doesn't It doesn't look so bad here. [snorts] Okay, maybe if I put on my laptop here off the support. Oh, let's see. tap in here. Okay, it's not shaking anymore, I don't think. There you go. You happy?

A damn, bro. What the hell you dropped? Okay, this is going to I'm going to get some wrist injury, man. Am I going to be able to lift anymore? That's what for a sec. Was exaggerating. All right. Where were we? Okay. Yeah. So, I um you can see my screen, right? Yeah. Okay. So, I got a transcript here, a doctor transcript.

And uh here's the schema. You have a video ID, English, you got the title, the author name, yeah, thumbnail, URL, whatever. So this is this and within the transcript array it says what the text is the transcribe text and when it starts in seconds and how long it lasts in seconds. So we'll use that as a reference.

I just picked a random one, the first one I could find. Uh in the transcript reference to L I N. Yeah, whatever. Actually, I'll just copy the file name. Okay, put it in bar transcript. add our transcript to a prompt that we can send to together AI using llama 70B instruct turbo I'm using a frontier model but one that is like cost effective if you want to do this at scale you can probably if this doesn't look good, we can obviously iterate on the model that we're using.

That's an important choice that I didn't talk about yet. I'll just put it here. Model choice. llama. We I want to narrow it down to frontier models. So ones that are in the billion range and those are going to be, you know, better than the other ones. Okay. Option to really have. We can look at together AI's thing.

Okay. models. We got a lot. We got a lot. But yeah, I want something serverless. Okay. There we go. That should help. Okay. what's the best model that they have? You know, batching cuts the price in half, but sometimes it takes a really long time. So, I'm just going to call it live for this and um yeah, cuz we want to do it [snorts] fast, iterate quickly.

Okay, so Llama 3. 37 billion um with not the instruct version. It's close to the most expensive one. Deepseek R1, which is the reasoning one. [snorts] It's surprisingly expensive. Okay. Am I the right ones? I feel like I'm surprised that is is GPOSS really shitty. I don't know. Well, what's a good way to like evaluate these?

Maybe like, you know, we could have a model that's been evaluated well on like medical scripts, right? Because that's the output that we want. But I don't know if there's a benchmark like that. Medical script benchmark AI. Okay. I don't want clinical reasoning. It's just YouTube script. Um medical script medical knowledge.

Yeah. Hey. Okay. Okay, this has something healthcare. Yep. Okay. Hi. You want to watch? It's my dog. What are you doing? Good girl. I'm kind of just looking for a bar graph. This is Yeah. Come on. Where? There we go. All right. Which mall is this? [snorts] Gemini. Yeah. Gemini Pro. Is Gemini not on there?

Okay. What else we got? Yeah, maybe this is a it's a blog post, but it's a good starting point. I don't want to do I don't want to spend too much time doing research here. It's good. Okay. Still got to get used to that. So, I got a bunch of miles here. I don't know what these ones are. I think I want to go with this clinical knowledge.

That's the most generic one. And these scripts are like they're generic around emergency medicine. which is generic. Anyways, no shots there. Um, GBT4 base GBT4. Yeah, these are the models that they evaluated. Somehow four base is better than four. But none of the 70 billion parameter models have been evaluated and we don't even have like Gemini here.

[snorts] Yep. Yep. My dog is thirsty. Oh god. Yeah, he's a big drinker. Okay. Oh man. Oh gosh. I have to maybe ask maybe ask Gemini, right? The um Yeah. Find me a blog honestly. Yeah. Just go fast. Okay. Where should these models um talking about generic medical Um, find a benchmark that is something adjacent to and hey, number one.

Okay. General reasoning why it wins not really I don't want medical okay so it it went for like answering medical reasoning questions we know we want general medical knowledge so for example provide stories about why Uh walking is good for your heart etc. Okay. Oh okay. So Ki is good for narratives apparently.

Social science better storytelling matter took GLM5 factual density that seems pretty promising. GBT OSS 120 less likely to drift into overly technical Oh god. Okay. Massive multitask. Nope. Not really. Oh, is this the is this the only song? I doubt it. Yeah, I don't think this is a relevant one page.

I could browse through it, but it's not really specific. We won't maybe it will help us if we look through the script. Okay, I'm gonna find Yeah, this is fine track cuz does he do this Chris mention any medically technical stuff? No. Is this catered people who aren't really technical? Okay. Yeah, I want I want to say, you know, Gemini is saying that Kimi K2.

5 is good for narratives. I think that is a good bet to start on. I'll start with that. And if that doesn't work, I'll try the other three. Okay. [sighs and gasps] Now, where is Kimmy? Here we go. Okay. How do I call this? King K25 using okay K2 and uh oh yeah let's just make sure that we've not noted down this model choice.

So we got we had Kimmy K5 or the other two I don't remember the third one was GBoss 12B and then the second one was GM5 according to Gemini Um, print according to Jim for storytelling, medical storytelling. There we go. I just say that so I can look it up or double check it later if I forgot why I made that decision.

Okay. King five. Send. Yep. [clears throat] Yeah, the transcript will be the style reference in the prompt. Okay. Now, we also needed the last one right here subject matter. Yeah, we set up another video. I can just look at the the videos. Okay. Also, put in the prompt. Use a title. that we'll use as a target subject to discuss.

Yeah. No, target subject for the script. I think I probably made this problem too long already, but I just want to give it the high level. We can iterate from there. We got that. We told it to use together. Kimmy, we told one together. Okay, we told everything. Yes. The out we'll just say the output one.

The output should be a JSON file with schema text and transcript. Okay, while it generates I am just going to go for a drink refill break. So Okay, we're back. Oh, Jimmy, I got stuck on the first command. All right. Well, I have it asking for permissions, so that's all me. All right. What does it want to do?

Import together. Oh, okay. Oh, yeah. I gotta specify the environment it needs to use. Yeah. No, I I don't. I probably should tell it though. Okay. Which environment to use. Okay. I'm just turning this off for a sec. Are we still streaming? Yeah, we're still streaming. Okay. So, I'm just going to tell which environment to use.

Okay. Yeah, we're not Or do you like to put these drinks in container? Okay. Okay. Well, this is creating the base. I'm wondering if there's any task that I can do concurrently. We want to make sure that we have the script down before you go forward. Yeah, I think I think there's anything I can do concurrently right now cuz I'm blocking myself based on the generation script.

Okay. Yep. Sure. I'm going to use that. Um, make sure you use that in the environment. [snorts] If I Okay, maybe interpreter myself. Where is it? uh papers and the one here we go. I don't know if it's using it right now. Py language Python. I don't think so. Okay. I'm just Yep. This is the this is the work now.

mostly checking the environment. Pick up virtual. Okay. Okay. Well, pause it now. Maybe I have to put a Python version. Yay, there we go. Okay, we're back. Yeah, whatever. Okay. Parse. Yeah, sure. Yeah. Yeah. Yep. Okay. Yeah, that's pretty good. Oh, this is unnecessary. No. Ignore. Okay. No, don't use just standard.

[snorts] No. Okay. So, use this. Yeah. Use whatever libraries and add them to our requirements. ext. text. I'll make sure they get installed to my environment. And I'm just like reusing the same environment for multiple files. I wouldn't do this in production, but it's good enough for this. Okay. Oh.

I need to do this file just yeah I mostly meant use this to when run when testing things. Okay. Oh yeah. So I have it generate a get recommendation text um so that it summarizes what it does and so that it kind of like ensures that what it's doing is relevant to the user. That's why I have it do that.

Okay. Yeah. Let's go ahead with this. While it's doing that, I'm going to pull Yeah, I'm going to put pull some more transcripts. So, I'm just going to let it go through ahead with this. I'm going to pull some more transcripts. Here we go. Act should I need to duplicate them. So I had them in another directory.

I'm just going to refer to them here. And I'll create references file references. Okay. So, yep. Data. Copy path. Absolute path. Yeah. Okay. Okay. And then what will we need for the actual transcend. That's your What's doing right now? Sure. Checking the help command, I guess. Get the transcript directory and use file.

Okay, it's using requests. Loads transcript. Yep. Generate transcript and the main script. What's this? Load transcript. So, it's getting it from the directory. Yeah, we'll have to modify that. What's it doing for the generate title style? Oh, yeah. The API key. We got to get this in here. We are compromising the key.

cuz I could put it in a file, but if I accidentally click on it, then that's bad. Okay, let's just see what Gemini has to say. [snorts and gasps] options for not revealing API keys [snorts] on the stream. Yeah. Yep. I mean I have that we just do that. So separate stream. Okay. Window capture. No screen.

Sure. Privacy mode. See here. Okay. What's What are the options within Sorry, I know this is not the most exciting stuff. [snorts] Don't leave this API key. Okay, let's leave it past this OS leveler profiles. Yeah. Okay. I think I'll just do that. I guess I can put it in my bash file for the stream. Okay, I'm just going to hide the clearing again.

Get the uh yeah the other API key. Okay. Just in the interest of saving time. This is not great. All right. There we go. Okay, we've got it. Use infrared number one key flash. Okay, I covered that one. We'll just accept this for now so we can move on. Yep. Yep. Yeah, sure. And then I just keep asking why is it doing that?

[snorts] Yeah. Well, it's hard to it's hard to I didn't sanitize my environment before starting the stream. I'll do that next time. But yeah, I it's really easy to leak secrets, I think. Okay. Yep. I think it looks good. And then where did we land with the Okay, it's got implementation plan centralizing config.

Oh no, I guess. Yeah. Yeah. Is this going to use OS? Sure. No. Uh just Yeah. Option one. Okay. Do [sighs and gasps] this one. Impation plan just has option one. Okay, great. Oh no, don't do that. Don't do that. Um, don't run any manual verification here. I probably should tell that in my command prompt.

Yes. Right. Good. Okay. I wonder where my phone is. I'll act. Okay. Have my phone. What's it doing here? That's pretty. You trying to show up again? geese here on there. Yeah. Say hello. Hello. [snorts] Yeah, that would be crazy. Gemini just like echo me environment variable of the thinking. Oh, that would be bad.

That's why I have it asking for permissions for everything. I usually don't have it that locked down when I develop. Okay. Um, yeah. Okay. What was it? Yeah, whatever. Okay. So, what do we have now? We have main. It'll take a title. optional default transcript path. It'll just put in transcript has my API key from the config.

Let's hope it's Yeah, it's okay. High stakes. Okay. Yeah. Here. Okay. A local transcript. Okay. Yeah. I gota take this the name. Yep. That's about right. Okay. Sure. Yeah. Let's see. Let's test this out. And it feels weird to thread this in like that. I guess it's nice to have it fail early, honestly.

Like that. Okay. Does it log this? That would be crazy. [snorts] Is that the API? XYZ. Okay. You're an expert script writing medical health. Yeah. For a new title. Yeah. Okay. Let's be entire based on the following reference transcript. Yeah. This is very like Google will not let you do this. This is too sus.

But we'll try it. This is just a model deployed in the wild. Okay. Tile reference transcript. Yep. Draw a JSON object. Yeah, this is going to be long, I think. Okay. Uh log the prompt input length. Yeah. And up input length and response. I'll put blank as well as any response as well as model metadata.

We might have to call API. Okay, this requires pro Look at the reference of this here. Okay. Generate. This isn't script. I know I chose this temperature. This seems very must have reasoned somehow. I don't I think this is going to be marginal to be honest though compared to like the the model that only seems like we have Yeah.

Whatever. Are you using the exact one? Yeah, sure. Okay. This looks good. Yeah, let's just go with print. We can we can refactor it later. Last words. Let's go with this. Actually, no. Let's uh yeah, do this and ensure it's in get ignore. I'm not even using git for this repository, but I like that actually.

Yeah. Add run metadata to the file name. No time step. Okay. Yep. Okay. In the meantime, I'm just going to check the constraints or bounds these models. Okay. So, what should we look out for in terms of length? Uh yeah, input output when calling something like ki25. I want to make sure it considers everything and gives me an input and gives me untruncated output.

I'm going also give it example code. Oh, don't do that. Stop doing that. [snorts] Okay, let's go big time. And here's the example here. Yeah, that sure. Okay, whatever. Let's do multiple times model. Okay, I guess that should be in there. Prompt tokens, completion tokens, total tokens title. Yeah, that's input reference.

That being okay. Accept. Accept. Okay, it might give me a short It might give me a short script. So, I might have to do it in parts. We'll see. State validation. Yeah, that's what I do. I don't want to check that be locked. Okay. intro body one body. Yeah, that's the idea. And then so what does it want me to do with the I know that's possible but what's the proper solution a prune but I just Oh, the dough.

That's fine. I think that's fine. I think it's fine. Yep. Okay, there we go. Okay, I think we're ready to start this baby up. Okay. So, Python, which py which Python I want to make sure it's using the right environment. Good. And then Python generate. Yep. And then title for the title. We'll just Okay.

I'm just going to go to the YouTube channel Dr. Weber. What's the one I'm referencing? 10 mistakes that you're Yeah. Oh god. 10 mistakes. Okay. Whoops. Let's see what happens. Was that the parameter name? Is this title that book? Okay. All right. Probably got the model key wrong. What's the key here?

Go. Okay. Okay. Okay. This is the only hard. So, looks like it's running. [snorts] What do you guys think is going to happen? Oh, there's somebody else for a second. That's pretty cool. How long were they on for? Like 10 minutes. Not bad. Or maybe it was your other device. Let's see. Says success. I am skeptical.

There you go. [sighs] How do we get that? Save it wrong. Make it log the output. Generate a script. Right. Okay. Log file name response text. Okay. Okay, log in the usage my okay stopped reason stop stuff naturally. Prompt tokens, completion tokens, total tokens. Oh, wow. It actually tells you the tokens per prompt.

That's pretty nice. That was very good. Okay. What did Gemini say by the tokens? Was it 1940? Did it say 1940? It's pretty close. Oh god. Is there any way I can see the outputs here? I don't think so. Okay, log it. We'll do the whole results actually. Yeah. Save the results to the JSON result. JSON do it.

Yeah. Let's run it again. Let's just do two mistakes. While we do that, I feel like we're going to have to do the chunking. So, let's look at the chunk analysis. Yeah. Okay. Okay, so I have is chunks chunks up high. This is just from a JSON that I produced. These are just like some motifs that are repeating in the scripts for Dr.

Rberly. So if we were separate the chunks, we can just use some of these. Okay. [sighs] So this is the output. This is pretty good. It's just failing to parse the JSON. Okay. No, it's because it's incomplete. Yeah. only. I'm surprised it said that there was a stop. Why does it say stop? Oh, length. Now, this one is length.

What was the output? Nothing. Okay. Let's Let's read this though. We don't have to read it in the JSON. They use Oh, it's got reasoning. All right, fair enough. Where's the text cell? What did I tell it to do? Transcript. Okay, transcript. No way. He used all the tokens on the re. What the [ __ ] [snorts and gasps] is that?

I feel like that's what it's saying. Look at that. I might have to try a different model. Okay, I'm just going to check with Gemini that what I am seeing or I think I'm seeing is what I'm seeing. Okay, I got this output. It looks like all it tokens on the transcript is reasoning token zero. Yep. But then yeah says the reason tokens are zero completion total prompt.

Yeah prompt is big. I think it's like 44 cents per million. So we have to use it 100 times. Yeah. Yeah, over 100 times he used 44 cents or I guess for Kimi it was 50 cents. Yeah. Well, the output is a bit more 2. 8. And what's the output that we're getting here? Yeah. The just two 2048 Yeah. Wow. So, it's Yeah, it's saying that spent all of it on reasoning, but Okay.

What else is saying? Okay. He thinks the reasoning tokens uh it's just the way that u took their AI set it up. Okay. Max tokens together. Hey, maybe I should use GLM. It's got a bigger That just be easier. We're the ones that they ranked GM5. Where can I read the inference max? I love you. highlighting.

If I went to Oh, I can just configure this. and the completions. So let's try it. So temperature we'll see 100,000. Go crazy. Right. A million. Yeah, five cents. I know it'll work though. We're trying out but similar platform. I don't want to use it. thinking. I don't need to think. Is there a way I can turn it off?

Probably just use a different model. Okay. Where's the last transcript? Oh, maybe it's just not showing it to me in the print. Oh, okay. Now I got transcripts here. Oh, wait. Is that the input or the output? That looks a lot like input. What the [ __ ] is going on? Okay, wait. I guess I'm still failing to parse it here.

Okay. He used 5,000 tokens. That's it. No worries. Oh, wait. I forgot I'll put tokens or six times of price. Oh, whatever. We're just going to do this for a little bit. Okay. Something's going on with the saving and it's not even live in the air. Got that metadata saved. I think I have logged this content.

Did I last log it? I'm printing the results. Maybe I should save the result. Wait, no, I am. Result. json. Yay. Okay. Why didn't you say anything? [snorts] All right. Yep. Live chat is popping off. [gasps] Okay. Yeah. Jason. Yep. Wow. It's really thinking here. Just thinking about the British spelling.

I guess I realized that it's supposed to be your British. This looks like thinking. Where is the transcript? This is all this. Yeah. What's going on here? Maybe the maybe there's something wrong with the prompt. Gosh, I think the I think it's over interpreting this prompt here. It's like, oh, you want transcript with dots?

[snorts] [gasps] I like how you transcript with dots. I'm going to just try to read [snorts] this reasoning in my experience. Okay, so this is based on the following director. There's a city social jet lag cortisol social. Oh yeah, it's really checking the spellings here. Is this the transcript I'm making?

Okay. Oh god, I think I lost. Okay. Blood. So the example reference we gave it is about the same thing. Okay. input that we had was to make the title blood pressure. Okay, so let's see. Do we see pressure here or is there pressure on this thing? Okay. And then this one. Okay. So, this is a transcript cuz Yeah.

Is it mission social jelly? social. Okay. He's copying this uh closing with encouragement hook. Yeah. Yep. Sure. The output should be all I think it is the prompt. So I'm just going to try it. And I also want to so ensure that the output files EG I add timestamps to generated script and result. json.

Yeah, I don't know to do that. And currently the prompt is resulting transcript output just being dot dot. Actually I feel like this is a question for Gemini. The agent this problem is making actually have a chat from earlier. I think this is the issue. It's making I think tomorrow. Yeah, use pro that one.

Okay, let's see. Yep. No, it's not. Okay, let's try that again. [snorts] This thing says yeah. So great. Still generating. Yep. It did say 10,000. 100,000 limit. I'm glad if I don't have to chunk it. That would be great. Is this a waiting game? Actually, I kind of get Keep going. Yeah. Okay, what we got here?

Oh, we still got an error. We still got this. Yeah, it's okay. Okay. commit. Initial commit currently seems to be failing. Transcription output in all JSON for transcript field is All right, we might have to simplify things a little bit. Let's just remove this style transcript. We'll do a different prompt.

Yeah, let's just the schema first and then we might have like a third party output saver instead of directly calling the API. Maybe it will save the API using u or save the files using library that together it has. Okay. Output about the weather. Gosh, I didn't imagine this one would take long. What could it be possibly thinking?

[snorts and laughter] It's probably just network stuff. Okay. Okay. So, we got transcript, got decent schema. Okay, it's got the examples here. So, this is the output, I guess. No new lines here. How does that compare to this? It's got a lot of completion though. Gonna make it ignore reasoning. I think that's a problem that I need to solve.

still reason here. So, it's the length. Holy [ __ ] though. It reasoned a lot about the J sound schema. What the [ __ ] I guess that's what they do. This looks okay. So, it's finished reasoning. It looks like here it didn't finish reasoning. No, it has. It does say looks good at the end. This looks correct.

I think maybe something in the formatting of the reasoning is messing it up maybe because the prompt is too long as we're saying. So, I'm just try that actually works. Oh, yeah. about some 500 random words essay 500 essay by the Okay. Still working. [snorts] Feel like it's almost done. Come on. Oh god.

[snorts] Maybe I should try multiple things in parallel next time concurrently. I was really thinking about the weather. This is going to be a great essay. I think I'm kind of curious to read it to be honest. So if it doesn't come out all corrupted. So we'll see. Come on. Okay. Wow. 10,000 tokens. Okay.

Wow. Okay, so the reasoning still looks messed up. Okay, that's fine. It ended properly, but the transcript is good. Yeah. Okay. What if I said a 10,000word essay? I feel like we should do that. If it failed on that, then we know it's a length problem. 10,000 words. How crazy is it going to be? Cuz we did 10,000 tokens.

How many tokens was it for the other ones? Where did they sell? Because this is the one that we just had. That's a short weather one. 10,000 and then these are the 4,000 5,000. I'm surprised. Maybe there's something wrong with this style transcript. Yeah, let's try that. So, we'll just add log this Try 500 words.

10 words of weather. And then we'll just say solid. Yeah, we'll put this where the common lines. Okay, cool. Wash break. Okay, back. It's getting late. Okay. What does it say? It's finished. Okay. So, having the tile transcript mess it up. Okay. Why is that? Okay. Is there being used here? Does it have to be It is.

Okay. Is that how you want to do it? Just separated. I guess that can work. Okay. These are not great get commits, but they're just for me to be able to come back to it. Okay. I should be able to figure out how to put the chat in here. Okay. Yeah, this is not Oh, okay. trans. Okay. There we go. Okay, you write a script.

Guess we're using this tile. [snorts] Okay. Oh, yeah. Cuz they didn't have the F here. Whatever. I'll just Okay, I'll use this style. And I want to add this draw titles. are done. It's getting late. It's getting hard to think. Okay, this script must be entirely. Yeah, that's pretty good. That's good up.

Okay, let's see that do be fixing it. Okay, [panting] this might take a while. It looks like I had a 100,000 tokens. So, this could go real loud. Did I commit? Yeah, I committed the chunks the first one. Okay. This is a lot of CPU, but it's not even warm. It's pretty cool. Max are great. What a glass.

It's probably a blur that I've got going on. Let's see what happens if I turn it off. Yeah, like I just get the screen. Is that message? Okay. Not really. Might be the the whole streaming thing. [snorts] Okay. Let's see our uh Great. Oh god, it sounds exactly like him. In my experience, the emergency department.

Sorry, I'm not mocking the English accent. I'm mocking the the rage bait. At least it's rage bait for me. is out. Okay. In order for things I see. What if I told you that [ __ ] Okay. And yet the readings are still dangerously high. All triggered. Oh my god. It sounds exactly like him. Did I actually do it?

Okay, we got to come up with a rapid random topic. Let's see what this Oh, wow. It even like mentions some random assass study or says that there was a study. That's crazy. And this only took 3,300 tokens. I wonder what's the length of this though. It's not that long. Get successful. transcript sauce.

So, it to me feels like the style that he uses. I just got to make sure that it's about the length of a 20-minute video. So, to do that, I need to count the words that they have from a different video. Actually, I've got an idea. You can come along. Okay. Yeah, I'll just make a script that counts the words.

Count the words in. I'll put the result separately. We need to be able to run it Python word counter. I also want you output a file for each where each word is on my Now just help me make sure that it's counted them correctly. This could be a flash job. Sure. Sure. Go ahead. Yes. What does it say? Fully immersed.

Okay. [snorts] Oh, I created a created an actual directory. Did I say that? there. Okay. Oh, wow. Used one function for both files. Whatever. This is just a utility to help me make sure This executes just by running. I guess it just goes into main. That's pretty cool. I never tried that before. Okay, so the actual transcript is about four times the length.

And let's double check that it's counting correctly. Yeah, it is counting correctly. Okay, so we got 6,000 here and then yeah, 1500 needed to be about 6,000 words approximately. I'll just put that in front. a uh 5,000 to 7,000 word script. Yeah, I'll do it. Oh, that should do it. Fix that. Okay. words 100 versus 6,000 reference.

Gosh. Okay, cool. Now 500 and we'll do a different title. Tim mistakes. So this is the one he just came out with today. It'd be funny if it was the exact same script. Okay, I have to transcript for that though. 7,000 years. Damn. Okay, let's get Let's just hope that this works and then I can go to bed and we'll uh build on it next time.

But this is a pretty good checkpoint even if it fails. This is 11:00 here and by this time usually my brain doesn't work. So usually stops around 10. Actually this is really reasoning. So it's pretty good. Like what's this? Kimmy K2. 5 Gemini as well. GBT OSS. I'll try this because it's five times cheaper, so it'll scale pretty easily.

I'm mostly looking at the output price now. And I wonder what happens if we batch it as well. Okay, that doesn't help. That help a little bit. Just a little bit. Yeah, probably try that one. Oh, still going. Okay. It's weird that it's like playing back my stream to me at the wrong time. Why can't I have it be like What is it?

Is that what reset does? Oh, no, no, no. I don't want that. I want to like Okay, let's refresh it and see what happens. Oh. Okay, I'm caught up now. We have this finished. Yes, it outputed some words. Okay, now I just got to count it. I'm going to manually count it. We'll use this one. It's going to get this new file name.

Oh, that's the old one, right? Yeah, that's the old one. The name the word counter. I'll put it here. Move the old one. Okay. Not bad. 4,400. What does it say in terms is it actually split up into 4,400? Yeah, that's pretty good. Okay. In my experience, the brutal reality. So, we're talking about the strokes he attacks the an organ.

Yeah. More people come our doors. Yeah. It just sounds sounds so made up. [snorts and laughter] [gasps] I'm Dr. Alex. Yeah. I mean, it's really good at copying things as well. So, I give some credit to the LLM models. This is I'm going to have to have it output in something structured so I can actually like so it can be human friendly.

Yeah, maybe we'll do that tomorrow. Then again, like maybe that's a really downstream task. Yeah, I'm not saving these. These are just validation files texts. And then we've got the Yep. Okay. Generate 5,0004 longer transcript. This was 4500. But we want five to seven. Whatever. That's not a big deal.

It's not going to make a huge difference. We can we can do that. But next step consider getting something readable. Okay, I got to push that. I'll push it. I'll put the link in description, but I got to go to bed right now. So [snorts] if you watch this, thanks even later. Okay.