If you're curious about what GPT stands for, it's "Generative Pre-trained Transformer" - quite a mouthful.
The only non-evident word here is "Transformer". Nothing to do with Optimus Prime - a Transformer is a type of neural network introduced in a paper in 2017 by a researcher at Google.
Some of the key breakthroughs of transformers:
- weighing the importance of different input words
- ability to be trained on very large amounts of inputs
- ability to be accelerated (GPU) for fast results
If you break it down:
Generative: generates content.
Pre-trained: trained on very large amounts of text.
Transformer: the neural network that process the input to output.
Let's see how naming evolves as technology gets more widely adopted.
10 Oct 2022
Generative Pre-trained Transformer 3 (GPT-3) was launched in beta in June 2020 and uses 175 billion parameters!
import json import openai from gpt import GPT from gpt import Example openai.api_key = data["API_KEY"] with open('GPT_SECRET_KEY.json') as f: data = json.load(f) gpt = GPT(engine="davinci", temperature=0.5, max_tokens=100)
## Example prompt = "xxxxx" output = gpt.submit_request(prompt) print(output.choices.text) ## or prompt = "xxxxx" print(gpt.get_top_reply(prompt))
GPT Index: connect GPT with my own data
2 Feb 2023
"GPT Index" is a toolkit of data structures designed to help you connect LLM's with your external data.
11 Mar 2023
First confirmation of upcoming GPT-4 model:
14 Mar 2023
Received launch email & invitation to join live launch stream:
I like that it's the somewhat nervous CTO who presents - not a marketing presentation, yet still very polished.
Live interaction with the audience is via Discord.
Despite its status as a potential future $1 trillion company, OpenAI confirms and comes across as a research-focused, developer-oriented company.
GPT-4's Playground (mimiced in ChatGPT Plus soon?)
1st demo, possible with 4, impossible with 3.5: "summarise a 1-page text in one line, only with words starting with 'q'".
Separate input for pre-prompt instructions ("You are X with Y attributes).
Main input allows for multi-step instructions, including long-form text (eg. code, article, etc).
Code generation becomes much better, given more context.
Understands context (eg ran in Jupyter) & dependencies (eg. missing libraries & how to install them).
Dirty trick: address the limitation of data cutoff in 2021 by adding missing data to the model.
eg. add API documentation updates to the prompt.
32,000 tokens accepted.
ND: ChatGPT says that's 5-6k words. But it should be ~25k words.
Good showcase of the capabilities of GPT-4 - but hope future model will either be connected Live to the internet (ie daily run?) or have at least a shorter window of time between data ingestions cutoff and Production use.
multi-modal input - text & images
"Vision Model", not just "Language Model" - still in Beta with one partner.
Demo of upcoming new feature generating a response with multi-modal input: image + text.
Not available yet.
description of image
Here: submits a screenshot of the Live Discord channel and asks GPT-4 to describe in details what this screenshot contains. Good results.
Live demo in Discord: user submits images with questions a dedicated channel, with live answers from GPT-4.
That's a cool demo format! (probably nerve-wracking for the organisers). Feels "real".
picture to website
Demo: from taking a picture to a website with a prompt.
tax code demo
Demo: provide tax code & question - answers with result + step-by-step logic.
Impressive calculations (eg taxes) considering it's "mental calculation" (prediction) - not connected to a calculator (potential future improvement).
Stream ends. 25mns.
No slide, overall presentation, etc..
Just a simple & effective developer-focused livestream.
I hope "Open"AI will communicate more/better about the bigger picture - for the company, its technology, and its impact on the world.
(beyond what Sam Altman tweets @sama)
Available in ChatGPT Plus immediately
Immediately available in ChatGPT Plus 😁
"GPT-4 currently has a cap of 100 messages every 4 hours"
I'm not sure what to test right now 😂
Might come back to it tomorrow...
Well, perhaps not..
First idea ended up very "meta" - it doesn't know about itself, because GPT-4 came out after 2021... but it also says it's based on GPT-3.. perhaps just a bug and the model is not properly activated yet? 🤔