5 problems when using a Large Language Model

Five problems to consider when building applications using Large Language Models (LLMs.)
🔔 Subscribe for more stories: www.youtube.com/@underfitted?...
📚 My 3 favorite Machine Learning books:
• Deep Learning With Python, Second Edition - amzn.to/3xA3bVI
• Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow - amzn.to/3BOX3LP
• Machine Learning with PyTorch and Scikit-Learn - amzn.to/3f7dAC8
Twitter: / svpino
Disclaimer: Some of the links included in this description are affiliate links where I'll earn a small commission if you purchase something. There's no cost to you.

[@PritishMishra]

One of my latest projects requires me to make ChatGPT output JSON, which is a nightmare. Most of the time, it adds extra text outside of the JSON, which completely breaks my application. But after a few days of prompt engineering, I've gotten it to function most of the time now. But getting these LLMs to do exactly what we want is still difficult. The majority of the time, it's pure luck.

[@ReportAI]

Did you try providing JSON definition you expecting as output in your initial prompt. It should work much better with filling in and creating lists when model know JSON definition upfront. Depending on what you want to be in JSON setting, set OPEN AI API call Temperature to 0, it might help too. Let me know if this tip worked for you.

[@bloodline39]

I mean, did you create an LLM of your own? Ik it's an odd question🫠

[@Becidgreat]

1) has a ‘learning in the wild’ type scenario been studied? Whereby the LLM is Tarzan and navigates as a young boy learning to communicate with us chimps? So the vectors are connected instinctively.
2) can CHATGPT go back to the original chat that responded offensively? I just wonder how LLM learning was affected by creating a false answer when the answer was true.

[@mgreek31]

Thanks for the video

[@jeanchindeko5477]

3:16 have you tried to use one of those OSS model instead of OpenAI API where obviously their keep tuning the underneath model which change the type of output you have. That not necessarily an LLM problem here but an OpebAI API issue.
Using OpenAI API you have no control on which version of the model is used. Today can be GPT-3.5 build 12345 and tomorrow can be GPT-3.5 build 23456 and the output will be different

[@underfitted]

Yes, that’s one reason. Unfortunately, their model is so far ahead of everything else that’s hard not to use it. But I agree with you.

[@neilwng]

you should tighten the screws in your desk, it gives me anxiety when it moves

[@underfitted]

You are very perceptive

[@neilwng]

@@underfitted I had the same problem on my standing desk - probably why I noticed. Very easy to fix

[@EvilCherry3]

Great video, the sound is a bit too low though.

[@underfitted]

Really? It’s loud when I try

[@aristhotle]

@@underfitted yea its way soft

[@PritishMishra]

@@underfitted The volume is very low but great video as always :-)

[@henrymaddocks984]

And the exposure is a bit dark. Doesn't detract from the message though. Great video

[@frankmunz4262]

Sound is fine

[@nedyalkokarabadzhakov5405]

Also huge sums of money and tunning wth human feedback.

[@AlexanderWhillas]

Its not just LLM demos, _ALL_ startup demos are smoke and mirrors