Leta, GPT-3 AI - Episode 10 (GPT-J, GPT-3, GPT-2 questions, facts, general knowledge)

The Memo: https://lifearchitect.ai/memo/
https://www.youtube.com/playlist?list=PLqJbCeNOfEK88QyAkBe-U0zxCgbHrGa4V
*Note: Conversation happened via text first; this is a re-creation of the text chat using an AI avatar which takes several minutes to generate. This recording is one single take of me talking, with no cuts. The questions are unedited. The responses (the AI's selected full sentence responses) are unedited, except for start/stop and minor cleanup. Read more: https://lifearchitect.ai/leta/

====Fine print:
GPT-2 is the 1.5B parameter model using WebText (data based on upvoted outbound links on Reddit) released in February 2019. I used the responses provided in the original GPT-2 paper. The bonus question used https://transformer.huggingface.co/doc/gpt2-large

GPT-3 is the 175B parameter model released in May 2020 (data based on a broader dataset, see: https://lifearchitect.com.au/ai/models/). I used https://www.quickchat.ai/emerson as usual, but had to lean on Copyhat’s mobile app https://copyhat.com/ (also 175B) for some topics including politics, where Emerson refused to respond because of the sensitivity filter.

GPT-J is the 6B parameter model (data based on a broader dataset, see: https://lifearchitect.com.au/ai/models/) released in June 2021. I used https://6b.eleuther.ai/ for prompts, with the settings: TOP-P=1, and Temp=0.5.

- Responses have been shortened for brevity. For example, GPT-J's full response for 'Ubuntu project founder' was 'The project was founded by Mark Shuttleworth, who is a South African entrepreneur, computer scientist, and philanthropist.' and this was shortened to just 'Mark Shuttleworth'. The full text results are available at https://lifearchitect.com.au/ai/leta/

- Speed/latency: Response times shown are in milliseconds, but actual model response times average 1sec-30sec depending on model, computing, and network.

- Repeatability: As of June 2021, all major language models give different responses each time, so the responses in this video are not final, and are not evidence of accuracy or inaccuracy. They are a one-off view of a response at a point-in-time.
====

Read more: https://lifearchitect.ai/
View the playlist (video): https://www.youtube.com/playlist?list=PLqJbCeNOfEK88QyAkBe-U0zxCgbHrGa4V

‘Leta’ is actually Emerson, by Quickchat. It is powered by OpenAI's GPT-3 and other language models.

This recording is one single take of me talking, with no cuts. The questions are unedited. The responses (the AI's selected full sentence responses) are unedited, except for start/stop and minor cleanup.

Using these technologies:
https://www.quickchat.ai/emerson
https://copyhat.com/
https://www.synthesia.io/

Dr Alan D. Thompson is a world expert in the fields of child prodigies, high performance, and personal development. He consults to families with special needs children including Bolshoi ballerinas, chess masters, abstract artists, and those with a mental age up to twice their chronological age.
https://lifearchitect.com.au/

Music:
Liborio Conti - Inspiring Acoustics
https://no-copyright-music.com/

Видео Leta, GPT-3 AI - Episode 10 (GPT-J, GPT-3, GPT-2 questions, facts, general knowledge) канала Dr Alan D. Thompson