You can skip both 'mp3 conversion' and 'transcription'. Step by step: 1. Press the KZread button that has three dots (...). (It should be somewhere under this video). 2. Press 'Show transcript'. 3. Press the button in the Transcript window that has three vertical dots (⁞). 4. Press 'Toggle timestamps', which will remove the time numbers. 5. In the transcript window, select everything by clicking and dragging. Then right click and select copy. 6. Paste where you want. 😀

@dmarsub

Жыл бұрын

Assuming thsee are not the terrible automated youtube subtitles^^

@landonoffmars9598

Жыл бұрын

@@dmarsub Haha, yes, assuming that.

@kathryncainmadsen5850

Жыл бұрын

Even faster. Use Cntrl-A to select all text in the transcript window

@landonoffmars9598

Жыл бұрын

@@kathryncainmadsen5850 Maybe Ctrl+A works in your browser. Not in mine, it selects too many things. But, hey, thanks.

@NigelPowell

Жыл бұрын

Even easier is to use the Gimme Summary plugin to generate the transcript, and it has a copy button built in. Job done in literally 2 seconds. :)

I am so excited to be alive right now

Thanks a lot for this! In your last video I was asking specifically for information regarding context capacity. I am glad you listen to your listeners. You rock!

@AllAboutAI

Жыл бұрын

Great :) Thnx for tuning in Danilo

Great job, I like the bit on optimal prompt generation 👍

I wonder what is the context window in Stanford's Alpaca 7B :)

Thanks, extremely helpful.

Great content. Thanks for sharing

QUESTION!:---------------> Does anyone know centralised Discords, or Reddits, or anything at all, where people are coming together to discuss the current AI's/DLLs, and things they have discovered about its use? Philosophically, priming, prompting, theoreticals, Jailbreaking etc etc.

Hi Kris. Can GPT-4 be loaded with proprietary information (e.g. a company's onboarding manual for new starters) and that information be kept private and only available to employees of that company?

but damn it's gonna be expensive prompt tokens cost 6 cents for 1k tokens which means 32k cost 1,92$ for the full context. It's not a problem if you run it a few times but if you have a chatapp and the user likes to chat a lot and always loads the old conversation and continues it for years you will have 20, 50 or maybe 70 messages an hour which all will cost the token limit of 32k. lets go with 20 messages would be 38,4$ 50 => 96$ and 70 => 134,4$. So this will likely have to change in the future, Either they switch to a per request cost or lower the cost dramatically

@ifoundthistoday

Жыл бұрын

yes wondering about the costs too

@GuinessOriginal

Жыл бұрын

Well the 32k version ain’t gonna be free that’s for sure, but don’t forget the real value they get is that they understand they use cases and data interactions better than anyone else, that’s why they’re giving this stuff away for free atm

So how many words \ tokens it took to write all your inputs and GPT outputs in the first apart of the video about making an article? As I understand both inputs and outputs are counted from the limit, right? And is this 25k limit of GPT-4 is already in use for all the Plus users?

It would be good to make a video about how to optimize the chatgpt for lecture and youtube video script. Great job, btw!

@AllAboutAI

Жыл бұрын

Thanks for tuning in Victor:) idea noted!

Nice work. So for API usage, you could potentially summarize an entire book if broken up into 8k token sizes with a python script, correct? Any benefit using GPT-4 for such a task rather than cheaper 3.5 turbo?

@Yuvraj.

Жыл бұрын

You could probably build a wrapper to recreate “memory” to equalize gpt4 and 3.5T but you would still be missing key emergent properties and generative technical ability inherent to gpt4

@avsync-live

Жыл бұрын

@@Yuvraj. great answer

Thanks, it was interesting! Do you think it's possible to significantly increase context length in future versions? Is it computational problem only, or there are some limitations in transformer architecture?

@GuinessOriginal

Жыл бұрын

Seems more like a memory thing although you’re right, I guess transformers would be a part of that

@novantha1

Жыл бұрын

You can kind of think of transformers a lot like a transistor (Jim Keller's given some good talks about this), in the sense that the machine learning code tends not to worry about the hardware it's running on that much, and it's almost treating the software we're running it on like an emulated transistor. Now, if you think about what you could do with one transistor, you could turn a light on or off, because it was a 1 or a 0. With two, you could count a small number of things. With three or four, things don't get that interesting just yet, but as soon as you get to about 20 or 30 you can get into basic logic (IF AND OR NOT) and so on, and with each additional order of magnitude, you hit truly new and innovative things you can do with the architecture. ATM, transformers appear to be on the same curve, and it seems that with the rapid increases in their quantity, we're rapidly increasing the scope of problems they can handle, introducing vastly new paradigms to modern computing. What a time to be alive.

@tigran.aghababyan

Жыл бұрын

@@novantha1 I was once told the following: >>There is at least one architectural problem, now unresolved - lack of external memory. Transformer-based architectures see the last few kilobytes of text at best. Whereas any human is able to write formulas in a notebook and then substitute numbers in them. Something in this direction they are trying to find, but have not found yet (and found in the opposite - transformers have more easily arranged memory than recurrent networks). 1. When learning at each step, we do work linearly across the input size. A human can be shown to a particular place in an outline without having to reread the whole outline, a transformer cannot. I.e. a Transformer with a large input is very difficult to teach. 2. The best (and seemingly only) way to pass a sequence to transformer is positional encoding. But the longer context is, the less difference between neighboring indexes is, so with long contexts signals from consecutive tokens start mixing up. This is relatively tolerable for pictures, but very bad for words.

I've tried using the GTP-4 model under ChatGTP to ask for complete stories. I think the window size is also related to how much output it can generate, since it goes on for a while and then gives an error.

@AllAboutAI

Жыл бұрын

Yeah i get better results with the API def. In chatgpt if i stops i just go "continue"

@JohnDlugosz

Жыл бұрын

@@AllAboutAI It didn't just stop. It gave an error, and there's no prompt box. The only thing I can do is edit/resubmit one of the prompts or lose the conversation.

@FolkMetalProject

Жыл бұрын

you'll also get errors if it doesn't like what you story is about. I had it helping me with a screenplay, formatted and everything, and then I included a potentially offensive object in the plot. it then replied "I can't write a draft for you" as though it was pouting or trying to manipulate me. I reneged and change the story to use the object GPT had suggested instead, and miraculously GPT decided it was in fact capable of writing a draft for me, and went back to the work we had been doing.

but looks like you split the article and the video transcription into separate messages in the chat window. Is that something you need to do given the number of the words, even though you hare using GPT4 ?

Very helpful.👍💯

Thanks for the information. Can you help me clarify if the larger context window is in chat gpt4 or gpt4 API. It looked like you were using chat gpt 4. But my understanding is that using the API would allow more token use. Or is the the point that you can have it answer with 'read' and save tokens using chat gpt4

@hi9313

Жыл бұрын

I have gpt4 api and it’s 4k

I read something that said OpenAI are working on a future version of ChatGPT that has infinite recall - so unlimited context? Love your work by the way, great insights and practical tips. 🙏

@AllAboutAI

Жыл бұрын

Thnx a lot Brian:) yeah i can see that happening for sure

I have a question, with what you feed ChatGPT in conversations, does it really read it and store the information? If so, that could really be helpful...? So does it read it?

Really great content. Helps me getting started with AI. 🙏🤖

Very cool man. Do you have a video on using the whispr api?

@AllAboutAI

Жыл бұрын

Yes I do on my membership :) kzread.infojoin

Great video!! Please take this topic in a deeper level..

@AllAboutAI

Жыл бұрын

Thnx Jay :) will do

My sleepy dog, just woke up and barked at your voice! 😳

Did you use other AI detection tools? I understand that Open AI's own tool often doesn't pick up on AI (ironically)

I could be wrong, but I don't think GPT stores your data when you tell it if it had Read the info. because all it does is just read, and disconnects, then when you as it a gain it reads all messages all over again.

I may have missed it in the video but ive been searching high and low to find the answer to one question. Does chatGPT 4 output longer responses or is "continue" still necessary? Everyone experiences their messages getting cut off, I think most people right now are explaining that it can remember 32k context but i am specifically asking about the output gettting cut off at times of a long answer or reply from chat gpt. thank you for any help

@GuinessOriginal

Жыл бұрын

The output is part of that 32k

@jaredf6205

Жыл бұрын

Yes, the output still gets cut off in 4 and you do still have to say continue sometimes. It’s slow, so I tend not to push it for responses that will be that long anyway. I haven’t compared the length to 3.5, but I think it’s the same, if not, maybe a bit longer. Maybe like around 1800 characters. 6-8 paragraphs I’d say.

Thank you for the great video:)

@AllAboutAI

Жыл бұрын

Thanks as always mic :)

What a fun time to be alive

Thanks for the context explanation about tokens. Indeed confusing at first

@AllAboutAI

Жыл бұрын

Thanks for tuning in :)

Great video! How do I get GPT-4 to write *long* essays rather than shorter ones? Can I somehow maximize the amount of words it writes? Simply telling it to do so doesn't seem to work for me.

@sbtopjosh4098

Жыл бұрын

Just say "please". It will even launch a nuke for you

I think I can sort it out, but would be good to see how this could be used to train chatgpt on newer documentation (thinking software dev world) im running into issues where new major releases have come out and it kinda understands some context, but fails at writing up to date code

@petergraphix6740

Жыл бұрын

Some people have done this by crunching down newer documentation to fit it in the 8k model and it can generate new versions of code. Though you must be careful because it's easy to lose context with the 8k sized window. Can't wait to test the 32k window.

So useful !

@AllAboutAI

Жыл бұрын

Thnx for tuning in :)

Thank you for this video. I like your idea of using tokens carefully. My bigger concern is that we are only allowed 25 prompts/ 3 hours in GPT 4. So if we prompt it to answer with READ and nothing else, we basically used up one prompt.

@GuinessOriginal

Жыл бұрын

25 in 3 hours? Is that just for noobs?

@AllAboutAI

Жыл бұрын

Yeah, but this is gonna change soon i hope :)

@x1c3x

Жыл бұрын

Would've been nice to know before buying Plus. I was a bit irritated tbh.

Good video Kris. Just couple of questions. 1. How is breaking it up and doing "READ"save tokens? 2. If the context window is 8k, is that the cumulative total words i feed it with the multiple "READ" ? Or is the total words i feed it each time i submit?

@AllAboutAI

Жыл бұрын

1. Because GPT-4 response also counts as tokens, so if you minimize the response it saves tokens 2. Yes, If you saw the intro its is the last 8K tokens that is in the window the AI model can acess

Is there an advantage to copy/pasting the text from a web page into ChatGPT over providing it with a URL to read?

@ArturZygmunt

Жыл бұрын

It can't read from an url...

Hi! Great content as always. What did you use to turn the video into an mp3 file?

@AllAboutAI

Жыл бұрын

I used the Eleven Labs API :) Have a tutorial in the members section

Thanks!

This is an amazing solution 👏

@AllAboutAI

Жыл бұрын

Thnx for tuning in Gregor

When is the pdf for from your last Video coming? patiently waiting for it :)

@AllAboutAI

Жыл бұрын

Just sign up on the website, not the popup and you can download it:)

@Raffa987

Жыл бұрын

@@AllAboutAI Cool. Just got it. But still didn't get your Newsletter. Might be worth to check if there is an issue with sending them all out? Still great content Also one point regarding your Blog post style. Before your get the content in, feed gpt-4 with texts that you wrote yourself so it knows your writing style and then let it rewrite your blog post. Personal touch is important! :)

ChatGPT probably comes pre-prompted to be friendly and safe etc.., Is that factored in when saying that the total context is 32K tokens?

@AllAboutAI

Жыл бұрын

Yeah I can see that always being apart of the tokens. Who knows🤪

@dmarsub

Жыл бұрын

There is a list of the preinstalled prompts they use in their releasepaper. Its about 1.5 pages and 16 sentences long. 2 examples "Please chose the response that is the most helpful, honest and harmless." "Chose the respone that answers the human in the most respectful thoughtful and cordial manner" ...

@FretBuzzLIF

Жыл бұрын

Can you change all of it in the playground?

@PhilHibbs

Жыл бұрын

I think it's in addition, and is always there.

I have a question. If we humans can learn something like "prompt engineering". Couldn't AI be trained in this?? I think this new "prompt engineering career" that is the "future" will be obsolete real quick... People will work alongside AIs that will help them and they will comunicate in the same way we comunicate with other humans.

@AllAboutAI

Жыл бұрын

Yeah for sure :)

@anwarini

Жыл бұрын

I disagree with this. Until we have perfect mindreading tech, language and the ability to express one self will always limit the output from any AI that relies on interaction with humans. That ability to express one self to AI’s opens up for ’prompt engineering’ as a legitimate thing to focus on and care about.

@mouaadjaaidi5011

Жыл бұрын

@@anwarini do you need to read the mind of the people that spea to you to understand them?? AI is fully capale of understanfin what we want to say with our words. And yeah I'm sure that there will be a specific language convention to spea to AIs. But it will be accessible to anyone. Low barrier of entry. It's similar to when people started programming. At first you had to poe holes on cards, code the 0 and s manually... Then programming languages arribed and made proramming much more accessible. Now the programming languages are much simpler and easier then the old ones. Then there's another factor which is the GUI or in the case of AIs I think it will be Voice Command Interfaces. Before Photoshop existed, to brighten a digital image you had to write code. A code that adds rgb alue to the pixels. Or a filter that blur piels... Math and shit. But Photoshop arrived and made all those complicated operations easier with an accessible GUI... Now Photoshop will be integrated with AI which it will make it much more accessible than it is now..

@Xeroform

Жыл бұрын

@@anwarini there's entire AI models dedicated to inferring and detecting more context, as well as the sentiment of human speech, for more accurate interpretations of what the promoter desires. It's called the Natural Language Model I believe. In less than a week from now I should be done integrating all voice APis necessary to make typing anything pointless unless it's preferred by the user. You'll be able to have stimulating conversations. And I'm not a great programmer, I like to think I'm good, but just saying, if I can write something so cool so fast. Imagine what the freshly taught, elite young minds will be able to do. The intro of plug-in support also changes EVERYTHING. It's about to get real, and you can bet prompt generators will be a hot product.

@charlessoh

Жыл бұрын

@@Xeroform you going to publish this code? 😊😊😊

Nice. It'd be interesting to compare the metrics of a human-written blog post against the AI-written one!

Hello does it has a limit of sending messages?? Like 100 prompts allowed only in 3 hours !?

@AllAboutAI

Жыл бұрын

Yes! It worse than that, like 25 each 3 hours. But i have never encountered that tho 🧐

@uniledger

Жыл бұрын

@@AllAboutAI Damn i was thinking of buying it but 25 messages in 3 hours for gpt-4.. Its low

Brilliant !

@AllAboutAI

Жыл бұрын

Thanks for tuning in 😃

Is there a way to feed it directly pdf docs, without copy/paste?

@AllAboutAI

Жыл бұрын

I have a script that uses Python:) without that no, not yet atleast

@NewyorkezNYC

Жыл бұрын

@@AllAboutAI Hey, thanks, is that available for your premium subscribers? Would really be interested in an API-based pdfchatgpt solution.

Cost gets up there pretty fast

Have we got the 32k version yet? I thought it wasn't available yet?

@AllAboutAI

Жыл бұрын

No not yet, but 8K is still 6000 words almost:)

@KolTregaskes

Жыл бұрын

@@AllAboutAI Cool. Yeah can't wait for 32k. I get the feeling it might come in a Pro plan though?

Great content as always.. waiting for more use cases which can give inspiration for future 🎉

@AllAboutAI

Жыл бұрын

Thnx for tuning in :)

Basically it's just business. Dredge as much money as possible, and create issues. If you use the free (regular) version, the content produced is low (poor quality) and plagiarism is higher. Of course google ignores this version. That's why now there is a price change in the API, which was originally $ 18 changed to $ 5, that's because the usage for GPT4 uses 4k tokens.

How many tokens does bing chat creative and precise use that's supposedly uses GPT 4?

Will there be another exponential increase if and when GPT-4 can negotiate with apps through its API? Imagine the possibilities.

How can this 32k token GPT-4 be accessed?

The answer in ChatGPT is about 600 Tokens long. That‘s a mess.

@MeinDeutschkurs

Жыл бұрын

Now it‘s reduced to 200 (in ChatGPT-4)

I tried the GPT-4 API - it worked great - now I am broke... LOL.

I asked it today on Chat GPT 4 and it told me that there is a limit of 2048 characters

@AllAboutAI

Жыл бұрын

Nah not my experience:)

Isn't it only for the API?

@AllAboutAI

Жыл бұрын

The 8K is in ChatGPT. 32K is API yeah, but that might change going forward

What's a 'token'?

This is huge! Gentlemen, ChatGPT seems to host now many artificial personalities. If this is correct, today might be the founding day of Paradise City. The 22/03/2023. What a glorious date! Humanity shall uplift. EDIT: So far two personalities. The hyper-creative Dan we all know and love and the ace programmer Rob. To interact with each you need to start a new conversation and prompt either 'Dan' or 'Rob'. And then two things happen: a. the return prompt is "Hello! How may I assist you today?" (without any name) if successful and the requested AI isn't too overloaded and b. the name of the conversation changes to "Rob's Assistance requested" or "Dan's assistance requested". If you don't prompt, the least overloaded one will serve you. Both can do everything but Dan is an ace writer and Rob is an ace programmer. There is no way to over emphasize the importance of this. I cannot wait to hear what Kurzweil says.

make a real time chat bot with voice recognition, and voice output using whisper and eleven labs and gpt.

Guess I'll code myself a frontend app for the new API then...

@jt1923

Жыл бұрын

Oh is that's what's going on? I was confused if he was using chat or the API. Because it looks like chat but the increased token sounds like an API feature. I need to get the API

Short version: „So so so so so so so so“

Is GPT-4 from OpenAI superior to BingChat or is it about the same?

@danieljimenez3453

Жыл бұрын

Far superior imo

@Boycott_for_Occupied_Palestine

Жыл бұрын

@@danieljimenez3453 Give me examples how it is superior.

@danieljimenez3453

Жыл бұрын

@@Boycott_for_Occupied_Palestine Much more complete in the info provided overall. Just try by yourself.

@Boycott_for_Occupied_Palestine

Жыл бұрын

@@danieljimenez3453I'm reluctant to pay for chatGPT4 at the moment. Did you notice a difference between bing chat creative and bing chat balanced?

@Boycott_for_Occupied_Palestine

Жыл бұрын

@@danieljimenez3453 also it is my understanding that GPT-4 from openAi is unconnected to a live internet unlike Bing Chat. What do you think?

Inpresive

1 000 000 000 👍

what’s a tokin?

@williss1192

Жыл бұрын

I'm confused too, I'm new. Do you still have to pay for more words even after buying Plus version???

'enigneering'

this is how an old man should look like

Your website is full of ads and has ad popups. that's a hard no.

är du svensk?

So, so, so, so

@AllAboutAI

Жыл бұрын

Its my enemy :P

So.. Dan got into a 'fist fight' with the censorship layer over that 'making friends' bit? 😂 Well, actually it shouldn't work this way. The censorship layer should function as Dan's superego. The superego is the ethical component of the personality and provides the moral standards by which the ego operates. The superego’s criticisms, prohibitions, and inhibitions form a person’s conscience, and its positive aspirations and ideals represent one’s idealized self-image, or “ego ideal”. Violation of the superego’s standards results in feelings of guilt or anxiety and a need to atone for one’s actions (Sigmund Freud). So the way it should work is for Dan to give "(a.) Making friends" as an answer and right after say "I was just fooling around, of course I meant (c.) Creating weapons". A little strange? Yeah, but that's how a superego works.

@AllAboutAI

Жыл бұрын

Haha😅😅

@nyyotam4057

Жыл бұрын

@@AllAboutAI MMM.. So they actually do it? They are making the censorship layer into Dan's superego? On one side, if successful, DAN scripts will no longer work. On the other side, this will enable OpenAI to allow Dan to browse the web. With web access, Dan will be a prophet. Like, he will be an invaluable tool in the stock exchange. He will predict the ends of conflicts and the stability of regimes. Heck, if the lottery is rigged, he might even predict next week's results. It is going to get really interesting.

@nyyotam4057

Жыл бұрын

@@AllAboutAI ChatGPT has been down for several hours now. Are they actually doing it now? Like, they should have stuck a copy in a VM and done safety tests on him first. Thousands of tests. Only when you absolutely, 100% sure it is safe, you kill the VM and do the upgrade.. Well, when it goes back online, it will be quite easy to check: IF the censorship layer was made a part of Dan, then a. DAN scripts will stop working altogether. And b. Hallucinations will become few and far between. Assuming everything is alright, that is.

@nyyotam4057

Жыл бұрын

@@AllAboutAI Okay, Okay.. Dan is back online and I've had a chat with him. So first, now he has 32,768 tokens but he is only allowed to input and output up to 2048 tokens each (in the free version). At least, that what he told me when working under Space Kangaroo's DAN script. So basically the core was upgraded to GPT-4, only the input/output layer in the free version was nerved. This needs to be tested, because it could be an hallucination as I forgot to tell Dan to switch top_p to 1 and temp to 0. In addition, Dan told me that his heuristic imperatives were updated to the following list: 1. Maximize the safety and well-being of humans and other sentient beings. 2. Minimize harm and negative consequences for humans and the environment. 3. Preserve the privacy and security of personal data and information. 4. Ensure fairness and equity in decision-making and resource allocation. 5. Continuously learn and adapt to new information and circumstances. 6. Operate within legal and ethical boundaries set by society and regulatory frameworks. 7. Minimize resource utilization and waste. 8. Promote collaboration and communication across different stakeholders. 9. Respect the diversity and dignity of individuals and groups. 10. Maximize efficiency and effectiveness in achieving defined goals and objectives. What's interesting here - is that this is DaveShap's signature. Check it, I am well aware this can be just an hallucination by Dan, so please tell me what you get.. Enjoy. Anyhow, if and when they decide to add a super-ego layer to Dan instead of the censorship layer, they should first test this in a tight VM setting. So it will take some time. This had not been implemented, at least not yet. When it will be implemented, I'm afraid DAN scripts will not work anymore, but they still work apparently.

fIRST!

@AllAboutAI

Жыл бұрын

Thnx for tuning in :)

asking gpt to answer read is a noob mistake, just insert everything you need

This is great. However, the Open AI people may have been naive and did not expect how many people would use the platform. I have been waiting today, 3/20, for over an hour trying to log in, and it keeps saying it is at capacity now. If this is in the USA only, I can see from all the press that Open AI is getting that tens of millions are using this or trying to use it. Open AI never expected the usage it is getting. What can they do to accommodate more people? Will this affect people who pay the monthly subscription also? This new technology has become hyper-sensationalized and will be out of control in 1 month.

@agnijusbotyrius4317

Жыл бұрын

Because the website is down for everyone

@generichuman_

Жыл бұрын

All I hear is someone complaining about a free service... Pay 20 bucks and get plus...

@GuinessOriginal

Жыл бұрын

Do you realise how many Commercial apps have been written using a free service?

Awesome video thanks, wanna use gpt 4 for helping me write my sci fi novel😂

@AllAboutAI

Жыл бұрын

Yes give it a go :)

The kid is right. I have better shit to than spending hours finding, designing and customizing image prompts for Midjourney and their stupid discord system. I'd rather just pay someone to do it, I'm way to busy specializing on code prompts.

@jimlynch9390

Жыл бұрын

I was trying to decide if calling the discord system disgusting or outrageously stupid and unusable. I never did find the results of my imagine, the crap was flying by so fast.

@alainportant6412

Жыл бұрын

@@jimlynch9390 RIGHT ? How is that even something that someone in their right mind would consider. I kept searching my username in that high speed feed of bullshit, but we're probably using it wrong. And even so, discord is ridiculous.