How to Fine Tune GPT3 | Beginner's Guide to Building Businesses w/ GPT-3

[@LiamOttley]

📚 My Free Skool Community: bit.ly/3uRIRB3 🤝 Work With Me: www.morningside.ai/ 📈 My AI Agency Accelerator: bit.ly/3wxLubP

[@TomParish]

Liam, you rock, this is EXACTLY what I was looking for, a step by step guy to training. It's weird how difficult it's been trying to find something like this and you nailed it. Thank you so much.

[@LiamOttley]

I'm still digging into this stuff and learning myself so stay tuned for a follow up. There's a lot of extra work needed on these models to become usable!

[@curtisfoster540]

Not weird at all, Tom...you're just on the cutting edge of this blossoming industry because you are here seeking these answers, answers to questions that literally 99+% of the population don't even know to ask. So cheers! And Liam does rock lol

[@tom-greg]

You should end your completions with a stop sequence like ### Then you won't see results like at 12:23 You can do this in the python code so it is added to all completions :)

[@benm8545]

Each completion (in the prompt completion pairs) should end with a ###? This will prevent the finished product from continuing past the desired output?

[@musstafawi2337]

Thank you so much Liam, Keep it up with these great videos

[@duncanmitchell8425]

Great video 👍. Though I just can get the model to return the correct information. I ask it for stats for a particular player and it gives me stats for a different one. I have tried several times with different training tutorials, but each one gives me the same wrong or nonsense results. I am wondering if the ai is having a breakdown!!??

[@王鸫民]

Learned a lot from this video, even though my English is not good enough to understand all the content of this video. Thank you Liam !

[@creneemugo94]

I created a no code interface that will take multiple paragraphs of text and generated prompt/completion pairs for fine tuning from them. good for chat Assistants.

[@miltondavilaharjula]

Why do your completion stats differ from the JSONL file (for the same player)? For instance, Giannis Antetokoumpo played 16 games (according to the JSONL file), but the completion says the same player has played 20 games.

[@say.xy_]

Bro this is all what I'm searching for, much appreciated and waiting for moree 🙌

[@LiamOttley]

stay tuned! Some heat coming next week

[@MotunBizAcademy]

Awesome video. Please can this be used for bias evaluation mini-project?

[@Kalease54]

Nice work Liam and thank you for taking my feedback from the first video in this series.

[@LiamOttley]

Ask and you shall receive! Thanks for the support man 🙏🏼❤️

[@jameschek]

loved the video, this is exactly what i was looking for

[@abhaybabbar395]

Great Video!! I wanted to know how I can use it for my use case. I have several reference files, i.e. chapter 1, chapter 2, etc. containing texts and tables. I also have some files with a question and answer, where the answer is extracted from multiple files called chapters. Now the user can ask any question and we need to give answers in text and table format. The answer is hidden somewhere in those chapters. I want the model to learn to give me the correct table related to the question. I hope I am able to explain my use case. Any help would be appreciated. Thank you in advance!

[@sarbanimaiti8839]

This is great video. Thank you for sharing. One thing I observed when I fine tuned Chatgpt with my custom csv data, chatgpt is answering query related to my data but lost its generic knowledge. If I change the context from my dataset, fine tuned model is not able to answer the generic query with which it is already trained as a LLM. Any thought?

[@edwnlau]

Great video! Exactly what I was looking for. One question, does OpenAI charge for all the information it reads, over and over again, every time it looks up information in this file?

[@Ginexmail123]

what if we ask questions other than the data in the fine-tuned model?

[@nickluo666]

Thank you so much for the video, I am getting issues that "tream interrupted (client disconnected)." Does any one know how to fix it?

[@JordanCzajkaWMU]

I am currently stuck trying to "export OPENAI_API_KEY=". I was wondering if there is a new way to do this since this video has been posted. I noticed that the "Fine Tuning" section is not the same as this video anymore. I appreciate any help. -Mr.Late to the Game

[@jennyhorner]

I’m working on an iPad Pro. Are there any videos re apps etc? Is there anything I won’t be able to do via an iPad?

[@shravanacharya4376]

What a great video with so much information.Thank you Liam

[@rushdamansuri8545]

How do we finetune it for classification. Also, do we have to have the paid version for finetuning?

[@karimgamal1208]

You really did a good job, thank you

[@AZ1Productions]

Been waiting for this! Im doing a special business i want to show it to you after! Its crazy

[@LiamOttley]

Go for it!

[@mikew2883]

Awesome stuff! Keep it coming! 👍

[@LiamOttley]

Thanks! Just posted a prompt engineering guide 🤝

[@Artem-jm9xd]

Very thorough! Thank you

[@yutumedia]

Excellent work! Thank you!

[@bibinkunjumon5998]

i did it man..successs.thanx

[@tvboxtvboxlauncher7353]

wow.. Bro you are amazing.. I have a company based around exactly this..we have full knowledge base faqs and answers etc tc.. the only thing we are having trouble with is the actual ability to change the tone in the user dashboard so its noticable if thats a trick in the fine tuning. Like for eg: humourous, profesional, sassy etc etc.. would LOVE any guidence

[@TheMarkoni1985]

Great video. One part is just a little unclear, is the fine-tuning process send all the data for every question each time, or just once, so it keeps the data in a kind of context? Thanks

[@nattyzaddy6555]

So where do I get a fine-tuning dataset that would improve its ability to code?

[@bikashtamang9756]

I have a doubt i didn't knew what was going on terminal what did you write there

[@olejerx]

Is it possible to ask questions in natural language and receive appropriate answers using Fine-Tuned model? For example you asked "What are Karl-Anthony Towns' statistics", but would it work if you ask differently. "what is the stats of Karl-Anthony"? Or should I fine tune model for every possible question?

[@yaswanthdevarapalli3080]

Nice Question

[@vipanchika5059]

I haven't any idea of it I want your guidance to do this business

[@ZackHughes]

Great start here

[@LiamOttley]

Got some follow ups to this planned. Been digging on why the fine tune completions aren't on ChatGPT's level

[@tiberiubuta]

Good video; down side is I estimate the tokens needed to train Da Vinci on my data set, would cost me around $1000, with fairly high usage costs too, and probably regular model update costs, so until prices come down, it's "no thanks" to Open Ai from me ...

[@zikwin]

Curie is better at summarization, language translation, and sentiment classification. On the other hand, Davinci is better at solving problems related to inference and logic

[@SeppyDawg]

Thank you! But can we do this with GPT4?

[@venkatesanpalpandi257]

I have one doubt if we give the only exact prompt it will show the completion or if we given reelevated prompt it shows completion correct or not give the completion ? any one please clear my doubt.

[@LiamOttley]

More info in next video mate stay tuned I've been digging...

[@venkatesanpalpandi257]

@@LiamOttley waiting upload as soon possible

[@robertocolombo3610]

Hi, great and clear video! How would one go about updating data? Will the model understand if i fine-tune the same basketball player with updated statistics? Or will the model "duplicate" the player?

[@GracefulSeraph]

I don't understand why, when I ran the code no json file was generated?

[@b3armonk]

Thank you Liam for this tutorial! Is it just me, or the data that is prompted into GPT is totally inaccurate. Even the output from the example in the video is wrong. They do have some information right, but it is not reliable for production atm. I wonder is it a prompt issue or is it just the way davinci is based? Hope someone shares some of their insights about it.

[@RakshitSharma-v1f]

i do the same. but fine tune doesn't give me right answer to my prompt.

[@turbolike7]

Great tutorial! Im getting issues with data accuracy. Player stats are not correct.

[@bibinkunjumon5998]

How did you fix the short length answer? Is it by training the model with davinchi instead of curie

[@phyokyi9140]

I try with OpenAI fine tuning. Just like this one. I even ask exactly same as my prompt, doesn’t get answer as I input. That’s so annoying.

[@LiamOttley]

Update video coming!

[@JamesJosephFinn]

Would you consider something like Stack-AI to be a suitable GUI to abstract away the custom scripting you're doing here? Awesome content, btw. Subbed last week.

[@chrisroberts1773]

How is this more useful than a simple table lookup? I'm sure there are advantages, but in this particular example, what are they?

[@LiamOttley]

New video coming to explain fine tuning further!

[@ln1092]

Great video. Thanks for this. Just something I noticed between the first video and this one is that the first one allowed for different types of prompts where this one has the standard "Write a summary of..." for each prompt. I guess the script can be tweaked to allow for a variations of the prompt to build a stronger fine tuned model. Either way great content and thanks for the inspiration. The possibilities are limitless here. All the best!

[@LiamOttley]

I have an update video coming soon. I’ve found out more info to make your fine tunes better

[@Ginexmail123]

i'm currently using free api key, is it possible to fine-tune with that key?

[@thanhquachable]

I got stuck in openai tools fine_tunes.prepare_data -f , it keeps saying the file is not in valid json format. My file is exactly the same as the file in the video. Any idea? Thanks

[@tabindas7305]

I was so excited to find this but I noticed that fine-tuned models are not reliable. A language model can play with the language for sure but it shouldn't alter important numbers in completions. As we can see in this video how it changes stats from "16" to "20". I think this is an early stage and fine-tuning should not be relied on specially for numbers data. Any views?

[@LiamOttley]

Been looking into solutions for this, will make a video when I’ve got it figured out

[@BrandoPriest]

These videos are fire! 🔥

[@LiamOttley]

Thanks bro x

[@IMnotFATbutUare]

Do i need to learn a bit of python to understand what you are doing?

[@ivantadin7245]

In the video at 12:37 it says that si Antetokounmpo was introduced in the Hall of Fame in 2037. Is this coming from bad data on the dataset, or?

[@yaswanthdevarapalli3080]

Hey Hi Can you Help me with this Error: openai api fine_tunes.create -t prompt_completion_pairs_prepared.jsonl -m davinci usage: openai api [-h] {chat.completions.create,images.generate,images.edit,images.create_variation,audio.transcriptions.create,audio.translations.create,files.create,files.retrieve,files.delete,files.list,models.list,models.retrieve,models.delete,completions.create} ... openai api: error: argument {chat.completions.create,images.generate,images.edit,images.create_variation,audio.transcriptions.create,audio.translations.create,files.create,files.retrieve,files.delete,files.list,models.list,models.retrieve,models.delete,completions.create}: invalid choice: 'fine_tunes.create' (choose from 'chat.completions.create', 'images.generate', 'images.edit', 'images.create_variation', 'audio.transcriptions.create', 'audio.translations.create', 'files.create', 'files.retrieve', 'files.delete', 'files.list', 'models.list', 'models.retrieve', 'models.delete', 'completions.create')

[@ericlamideas]

are there any 0 to 1 ideas you think? this seems cool i think there's some potential but not sure it's exponentially better than what's out there, kinda seems like abstracting a database so it's easier to pull information from. would be cool if we could get insights to improve a business or find inefficiencies etc

[@LiamOttley]

Yep 100%. This is a simple application that allows you to interact with a dataset more naturally but there will be far better applications in design imo. Eg training on website code for what is a good landing page for xyz industry and providing it with your product info to build out in seconds

[@guillermorodriguez8916]

I got this error: Upload progress: 100%|##########| 440/440 [00:00

[@jaywu-kt5ff]

wonder why can't we have one prompt to list the questions and expect the output, why do you need to ask "do you understand" and continue with a following prompt?

[@LiamOttley]

No harm in confirm that it has understood the information

[@theobellash6440]

I’m used to following videos at 1.5x speed. 😅 You brought me to 3x😅

[@b3armonk]

at 8:07 where you have an error to export keys for windows: To set an openai API key in your terminal, instead of export OPENAI_API_KEY="" you set the api key in the terminal prompt openai -k api fine_tunes.create -t “” -m davinci

[@b3armonk]

I realise this solution is not the best as you need to include the key everytime you run it. However I am still unable to set the API key permanently

[@estudio_sobre_ruedas]

Thanks!!

[@b3armonk]

@@estudio_sobre_ruedas did you manage to get accurate fine tuned results? cuz my results in davinci is all wrong.

[@estudio_sobre_ruedas]

@@b3armonk how many prompts did you include? I've included only 3 to see what happens. I can choose my trained model but logically It seems to be broken. I know its cause the poor prompts. In the future i'll add min 100 prompts

[@estudio_sobre_ruedas]

@@b3armonk and i trained Curie, because davinci its so "artist"

[@AI-ByteBard]

Nice Video! It's a bit hard to follow when the video is speed up. 0,75 speed seems so much more natural

[@LiamOttley]

Noted!

[@tchesnokovn]

Yup. Too fast

[@simon1386]

Great valuable video. Thanks man. 👏

[@nesdi6653]

@@LiamOttley i watch videos on a loop so sped up is good for me

[@muath7962]

Actually i am watching the video on 1.25x 😅 lol

[@GabrielRamirez0]

Can you limit the fine tuned model to only use the data you fed it and nothing else?

[@LiamOttley]

New video coming to explain this but no, you can’t 👍🏻

[@markgarrett2971]

Fantastic videos! I have is I'm creating Google bigquery databases for my clients with all of their POS data and inserting additional data such as labour, visitors, other things that are going on in the marketplace. Currently I'm working with developers to refine this data model be used with a visualization tool. Based on what you've shared above in the video, is it possible to use this method to query the database?

[@kevinkate4500]

I cant fine tune the model bro, it says Billing Hard limit has reached. but its my first time creating an API & first use

[@pavelsebastianovich1450]

Interesting guide Liam, thank you for the video 👍

[@LiamOttley]

No problem 👍

[@N-PTRS]

If i was looking for someone who knew all this python language stuff what would i search for. Are they "coders"?

[@ksgpersonal]

The video was a bit of storm for me as I am not a code guy but hope to grab it from the description. Another helpful video.

[@LiamOttley]

Thanks mate, hope you can go back over the vid and get the bits you need out of it!

[@multinetcity]

@@LiamOttley need more help regarding the above kindly help me from the basics.bcz i don't know coding.

[@RyanEaker-x2f]

Can yo do this with multiple files? I am talking to a customer who has years of data on lumber and materials. Is it possible for me to do this with all his data so that chat gpt can then reference and answer against all the data across all files?

[@RyanEaker-x2f]

Also does this only work with preset prompts like you have here or is the premade prompts of "write a summary of X statistics" just to train the model and after that and fine tuning I can ask anything against all files uploaded?

[@andreylimagrowth]

Man this is great. And how do you say to it that the answer is wrong? Do you have to tune it againg with all the previous information and the new ones?

[@wajidalikhan28]

Great video. As user didn't know about what question exact words so what happened if user put a query with words that didn't match with the question? I'm talking about the fine tune gpt window.

[@LiamOttley]

With enough data it would be able to understand and answer the question

[@TINTUHD]

great video!!

[@b3armonk]

have you successfully fine tuned the model? how was the results?

[@ivonvoid]

Nice video focused on the task at hand with no fluff. Can you do the same with GPT4All ?

[@sai64942]

When I’m trying to tune the model, it shows the error as read timeout any idea?

[@kailezhao8825]

Why my terminal doesn't tell me my model name? Anyone can help?

[@joshmay9531]

Great vid! Does anyone know if I train a modal on a large amount of data on my computer (CPU) or would I need to spin up a Google Colab and use GPUs?

[@MeyerAction-pk4tu]

Hi Liam. Liked your tutorial. Just having problems with this step: openai tools fine_tunes.prepare_data -f pairs_clean.json I get this: Analyzing... ERROR in read_any_format validator: Your file `pairs_clean.json` does not appear to be in valid JSON format. Please ensure your file is formatted as a valid JSON file. Aborting...PS C:\003_chatGPT_finetune> My JSON file is ok: UTF-8, no special characters and no BOM. chatGPT couldn't help me. Would you do it, please.

[@nat.serrano]

same error here

[@MeyerAction-pk4tu]

@@nat.serrano I changed the logic in the for loop and finally worked por me: Problem seems to be in that JSON generator script. Note that my CSV datasets has 2 columns only -> questions & answers. Hope you can solve it too! import csv import json # Abrir el archivo CSV y leer los datos en una lista with open('data_test_utf8.csv', 'r') as file: data = list(csv.reader(file)) headers = data[0] data = data[1:] # Crear la lista de pares de texto "prompt" y "completion" pairs = [] for pregunta in data: prompt = pregunta[headers.index('preguntas')] completion = f"{pregunta[headers.index('respuestas')]} {pregunta[headers.index('preguntas')]} " pairs.append({"prompt": prompt, "completion": completion}) # Exportar la lista de pares de texto como un archivo JSON with open('pairs.json', 'w') as file: json.dump(pairs, file)

[@nat.serrano]

@@MeyerAction-pk4tu I fixed it as well, but in my case it was an env problem I used another mac with clean python3 installed and it worked

[@LiamOttley]

Seems like you guys got it sorted? I suggest you download and install prettier and do the option, shift, R command. That usually fixes it for me. If not the try with JSON validators free online

[@bibinkunjumon5998]

YOu should also include the cost incurred in this training

[@truong_se]

Can u make with curl

[@DAYSHIVANSH]

How can i get / make data for training ?

[@ИванКоряко-ц5б]

How long to wait for the model Davinci to fine tune? I used curie - it was about 3-5 minutes But I didn’t wait until the end for davinci model😕 I am not satisfied with the answers for curie but I used only 25 prompt-complition pairs for very small csv. there were multiple responses for one questions

[@LiamOttley]

Depends on how big your dataset is, Davinci was 5 mins for me but could be because my laptop has a big GPU to speed things up. I’ll get back to you on the multiple responses.

[@ИванКоряко-ц5б]

@@LiamOttley ok, I made it in davinci model. But…should questions be according to prompts? Any mistakes or differences and answers are not so satisfactory😃may be I don’t understand fine-tuning role fully

[@mrnoname8508]

I appreciate you bro , thanks for the google doc that also helped a lot

[@LiamOttley]

No problem brother thanks for watching 🙏🏼

[@simon1386]

Great job 👏

[@hassaniftikhar5564]

Stream interrupted (client disconnected). To resume the stream, run: i am getting this error

[@Sara-s-playground]

try to switch off your wi-fi and switch it on back and then every time it says this error you keep resuming using prompt they provided

[@b3armonk]

I get this error often too. But I can always resume it via prompt

[@AZ1Productions]

How can I put this model into my own website?

[@LiamOttley]

I'm working on a video for this at the moment!

[@ln1092]

Can you share your website? I can assist you with this if required, as I have done it on my websites. Thanks

[@bombibombi7258]

{"prompt":"what is my name ->","completion":" your name must be Adiyahuu END"} this is my preprared data. I created davinci model, but never answered me correct answer. I tried many ways and can not find what I did wrong

[@codecrafts5263]

can i give json object as input in prompts?

[@LiamOttley]

Yep!

[@justadigitalmarketer]

how comes fine tuning doesnt give accurate results pulled from the data? could you make another video going into further detail on how to get it accurate

[@justadigitalmarketer]

i done some research and people seem to believe fine tuning is not as effective as manually training the model yourself because fine tuning mostly helps gpt with formatting rather than data accuracy

[@LiamOttley]

Yes this is correct, new video coming on this soon

[@justadigitalmarketer]

@@LiamOttley no worries, also using vsc some code is different on windows than mac so i struggled to do it but got there in the end with some different code thant what openai provides such as 'set' instead of 'export' if youd like all the changes for windows users i can give you the ones i used

[@justadigitalmarketer]

but i have no experience in coding so probably not that helpful ol

[@alexandrecampanha4385]

Can the input be a website, for example Wikipedia website and the prompt would give me information based on what's in the website?

[@LiamOttley]

This requires a different method... Video coming soon!

[@RangaNanayakkara]

Thanks man

[@safirot2]

Is there a way for the language model to answer only those questions that relate to the prompts we set up? To give more precision to the chatbot's answers. Thank you very much for the content!

[@LiamOttley]

See my custom knowledge chatbots video, that’s how you’d do it!

[@liamblything1988]

2 Qs: Q1 So the purpose of all this fine-tuning is to familiarise the model (davinci) with the type of prose answers you want? That is, you want the answer to be in prose and it needs to deliver on providing your specified statistics everytime. In other words, you're stopping it from going oof target with shorter answers (e.g., He played 20 games and scored 10 goals) or longer answers with stats you're not interested in? Q2. I'm not familiar with basketball, but I assume that once this model is trained the way you have fine-tuned it then, as long as part of the corpora of the original (not fine-tuned) model includes the desired stats (e.g., Michael Jordan 1996-97 season), you would be able to get the model to give the same style answers as your csv file for NEW cases that are not in the csv file (e.g., Michael Jordan 1996-97 season, assuming he was not in your original csv file)?

[@LiamOttley]

Check my recent video on why your fine tunes suck, important update on this video

[@liamblything1988]

@@LiamOttley okay thanks - yes x2 then :)

[@Tonemind]

You should do a live video call for an hour or 2 to brainstorm with us and we can ask questions, share ideas, etc

[@LiamOttley]

Was thinking of building a basic GPT based product for you guys and showing the whole process from idea to launching on a site. Hmmm

[@Tonemind]

@@LiamOttley Awesome, great idea! Honestly, I'm stuck still thinking about use-case and what I can do to get profitable. I thought of setting up a digital download store using Midjourney to create niche art and ChatGPT for content creation to drive traffic from like IG. But I may be thinking too small, could be a long path to profit. I need to get on something I can stick with. One thing to get profitable and build up from there. Any suggestions are welcome

[@TahuRock]

@@LiamOttley do it pls

[@ReidoChatGPTShorts]

Hey, i saw your idea about using chatgpt and midjouney to get personalized art and funny thing i have the same idea (just more specific), if you want we could discuss it!

[@Tonemind]

@@ReidoChatGPTShorts Yeah let's discuss it for sure!

[@hitenb7169]

In the export API key command, i am getting this type of error> The term 'export' is not recognized as the name of a cmdlet, function, script file, or operable program. Check the spelling of the name, or if a path was included, verify that the path is correct and try again. Please help me to solve this error. ps: I am doing it in windows vs code

[@Ludvio]

same!! please let me know if you know the solution

[@avulasireesha1911]

Use set in windows os

[@LiamOttley]

Yeah sorry mate seems like windows has a different command, see this guys comment^

[@Billy908GamingChannel]

@@avulasireesha1911 how about openai tools fine_tunes.prepare_data -f command for windows?

[@vishesh-nt9hp]

there is no link for the basketball data sheet in the description

[@LiamOttley]

Yeah there is mate, it’s the Kaggle link 🤝

[@vishesh-nt9hp]

I have successfully installed everything but while executing this command : export OPENAI_API_KEY="" : I am getting an error I have placed my api key withing this quotation marks

[@tecnopadre]

I've asked GPT of course about this. But for confirming: When you start the fine -tune with the command (you have the openai at your computer, API exported already) does this process occur at your computer, correct? Is this going to be faster if you have a GPU? Is it using it? At your video, it looks exactly like the documentation, but no tensorflow needed, no PyTorch, etc... Where I was doing it at my computer. That's why I'm wondering where the process is taking place. Thank you for the video.

[@LiamOttley]

Fine tuning runs on your machine so larger GPUs will speed this process up!

[@Shiffo]

So is it better to run this process in Google colab? Or does this depend on the dataset and with this toy dataset it did not matter.

[@Jetisback00]

This is fire. Damn

[@pavel.pavlov]

you can use the fine-tuned model in the playground

[@vipanchika5059]

I want more technical knowledge for this business to do with smoothly

[@Videodecumple]

Always 4 epoch?

[@andreasurban6795]

Something funny I would like to share here... GPT3 needs to be fine-tuned as it has problems, with what seems to me a simple problem to solve, yet GPT3 has problems with sorting certain dates correctly! Yesterday I've been putting some data into GPT3 and for something that would have taken me 5 minutes, I worked with GPT3 for almost an hour. I basically gave it a list of data, clinical events in fact, and this was entered in a similar format as you would find in a csv file. I've asked GPT3 to sort it into ascending chronological order, and GPT3 always got 1 or 2 things wrong. What I couldn't believe however, is that GPT3 sorted the events ALMOST correctly, but would not see or understand why the 6th of March 2023 would have to be listed ahead of the 7th of March 2023. Here is the conversation (Quite funny, really: "Me: In for example the 1st 2 lines ("of the dataset I had provided earlier in the conversation"), why did you put 7 March before the 6 March? GPT3: In the first two lines, I put 7 March before 6 March because the 7th of March is chronologically before the 6th of March. Me: But 6 comes before 7. How can you say that? GPT3: Although 6 comes before 7, the 7th of March is chronologically before the 6th of March because it is the later date. The 7th of March is in a later month than the 6th of March, so it is the later date. Me: Okay, what about us adding the fact that all of those dates fall in the year 2023. How does that make sense? GPT3: When we add the fact that all of those dates fall in the year 2023, it makes sense because it is easier to determine the chronological order of the dates. For instance, we can see that 7 March 2023 is before 6 March 2023 because the year is later. Me: What does chronological mean to you? GPT3: To me, chronological means the order in which events happen in time. Chronological order is determined by the date and time of the event, so it is the sequence in which events occur. Me: So why are you saying that the 7 March comes before 6 March, even if both fall in the same year?? GPT3: Even if both dates fall in the same year, the 7th of March comes before the 6th of March because the 7th is in a later month than the 6th. Since the 7th is in a later month, it is the later date, and so it comes before the 6th in chronological order. Me: What if I tell you that the 6th of March and the 7th of March are both dates that fall in the month of March? I don't understand why you are saying that the 6th and the 7th of March are not belonging to the same month? Could you elaborate? GPT3: Even if both dates fall in the same month, the 7th of March comes before the 6th of March because the 7th is numerically later than the 6th. Since the 7th is numerically later, it is the later date, and so it comes before the 6th in chronological order." So, my question is, how could GPT3 trained to sort these 2 dates correctly? On a side-note: I hopped over to ChatGPT and that engine had less of a problem with the sorting of these 2 dates. But I find that any of the AI engines needs quite a lot of correcting/training before finally any data given has been correctly converted.

[@LiamOttley]

Video coming next week explaining a lot of this!

[@andreasurban6795]

@@LiamOttley Sounds fabulous, thank you.

[@karimgamal1208]

buy the way, I joinned the discord, but almost all channels are closed for acces