We compared text davinci 2 and text davinci 3 on a bunch of nlp benchmarks, see here:
LLMS
-
Recent Instruct Models Discussion Thread Analysis
By
–
interesting list of pointers in this long thread on recent instruct models (ChatGPT, Anthropic Claude, etc)
-
Vector Search and Map-Reduce Chains for Document Question Answering
By
–
exactly! if doc is larger you can still prune with the vector search, so you may not need to chain. but you could also load more documents and do a map-reduce chain over them. See these docs for QA in general: https://
langchain.readthedocs.io/en/latest/use_
cases/question_answering.html
… and let us know if any questions! -
John Schulman: OpenAI Cofounder Behind ChatGPT Success
By
–
There’s been a lot of buzz about #ChatGPT on Twitter lately — the main architect behind the system is none other than @berkeley_ai alumni and OpenAI cofounder @johnschulman2
! Hats off to John and the team! -
Technical mechanism of AI-driven voice synthesis
By
–
It generates a series of codes based on the phonemes and an audio recording of the speaker's voice.
— AI Breakfast (@AiBreakfast) 7 janvier 2023
These codes are then turned directly into a waveform, which allows it to generate speech that sounds like the speaker from just a small amount of reference audio.
Example (🔊): pic.twitter.com/HQvjJUnbiHIt generates a series of codes based on the phonemes and an audio recording of the speaker's voice. These codes are then turned directly into a waveform, which allows it to generate speech that sounds like the speaker from just a small amount of reference audio. Example ():
-
Technical architecture of the VALL-E generative audio model
By
–
Instead of turning text into a series of specific sounds (called phonemes) and then into a visual representation of the sound called a mel-spectrogram, and then into a waveform (a digital representation of sound that can be played through a speaker) VALL-E takes a shortcut:
-
Technical overview of an AI voice synthesis training model
By
–
The language model was trained on a huge amount of data (60,000 hours of English speech) and combines it with just 3 seconds of a person's voice and uses that to synthesize new, high-quality speech that sounds like the original speaker. pic.twitter.com/tIaimhfQ9R
— AI Breakfast (@AiBreakfast) 7 janvier 2023The language model was trained on a huge amount of data (60,000 hours of English speech) and combines it with just 3 seconds of a person's voice and uses that to synthesize new, high-quality speech that sounds like the original speaker.
-
VALL-E: New Language Modeling Approach for Text-to-Speech
By
–
Digitally clone your own voice in 3 seconds: VALL-E
— AI Breakfast (@AiBreakfast) 7 janvier 2023
Inside the new language modeling approach for text-to-speech synthesis: pic.twitter.com/slaX1P3sqzDigitally clone your own voice in 3 seconds: VALL-E Inside the new language modeling approach for text-to-speech synthesis:
-
Creating Advanced GPT Prompts Beyond Basic Expectations
By
–
There are some decent websites and repos, but imo, the prompts are too basic and they teach you to lower your expectations. I've been creating a GPT prompt toolkit that's higher level. Could easily be expanded to bible level. What info and references would you want to see?
-
LLM Math Capabilities: ChatGPT’s Blind Spot and Claude’s Unreliable Honesty
By
–
ChatGPT has no idea it's bad at math. It thinks it has a calculator, but it doesn't. Claude, to its credit, will own up to its limitations, but its ability to spot when it shouldn't answer is not reliable: