What is GPT-3 and how does it work? | A Quick Review

You probably have heard of GPT-3 and how it is a fascinating development. But have you learned why or how GPT-3 managed to impress so many people?
In this video, we will learn why GPT-3 is so unique, and how it manages to help bring in a new wave of excitement for AI. On top of this, we will also briefly look under the hood of GPT-3 to understand its architecture and some of its potential dangers.
Want to give AssemblyAI’s automatic speech-to-text transcription API a try? Get your free API token here 👇
www.assemblyai.com/?...
Apps made with GPT-3: gpt3demo.com/
B-roll credits:
Video by Julia M Cameron (www.pexels.com/@julia-m-cameron) from Pexels
Video by Jack Sparrow (www.pexels.com/@jack-sparrow) from Pexels

[@lilstar3705]

What is the software used to Built GPT3

[@spokewheeldiaries2724]

Thank you, useful video

[@AssemblyAI]

Glad to hear you liked it!

[@andersonsystem2]

Good video thanks

[@AssemblyAI]

You're welcome!

[@karthikeyanak9460]

Could you make a video about GPTJ

[@AssemblyAI]

Will add to the list of topics to consider. Thank you for your suggestion! - Mısra

[@amitvyas7905]

I am a little confused. You mention 175B parameters are those not words? What are those parameters? Are those like gender, masculinity,.. etc?

[@AssemblyAI]

Those are the values that are used by the model to calculate the outcome. On Neural Networks it would be weight and bias values.

[@amitvyas7905]

​@@AssemblyAI Then how are those parameters/values generated? I've read that word embeddings are generated by all LLMs but don't they need values with respect to royalness, gender, etc to get output for king-man+women = queen?

[@amitvyas7905]

@@AssemblyAI My question is that all the LLMs use some kind of n-dimensional input vector, how is this calculated, and how "n" in n-dimension is calculated? Is that calculated by feeding in some Neural network with a dictionary or n-gram words?

[@pathikghugare]

​@@amitvyas7905 those values are initialised randomly at the start then the neural network modifies those values based on gradients in back propogation step

[@msg3030]

As good as GPT-3 is, it still has little to zero memory. It can't remember hardly anything in terms of a conversation. From what I know about AI, it is extremely difficult to program a memory into the system that even comes close to the memory of the human brain.

[@paulden3158]

Why is it difficult to program a memory model?