Happy new year! I'm glad to be restarting weekly AI updates for 2023. A LOT has happened over the past two months and of course the largest of all is the release of OpenAI's ChatGPT.
As such, there's no better place to start our AI updates for the year. This edition is your ChatGPT cheat sheet. Read below for ChatGPT in 10 lines.
(Yes... this was optimized with ChatGPT)
If you would like to support our continued work from £1 then click here!
Marcel Hedman
Overview
1.ChatGPT is an AI chatbot by OpenAI (2022), known for its accuracy and human-like responses, nicknamed the "Google Search killer". Microsoft have recently announced integration of a ChatGPT-esque OpenAI model into Bing and Edge.
2. ChatGPT reached 100 million users 2 months after launch, and OpenAI now offers a subscription plan. (It took Instagram about 3 months to reach 1 million users.)
3. Emerging use cases: Technical (code assist, classification tasks, prompt engineering), Non-technical (content generation, education/ query response, blue-sky ideation, copy writing, itineary planning) and more
4. ChatGPT's hyper-realism poses risks such as fraud/plagiarism, data ownership ambiguity, misinformation spread, model hallucination, and embedded bias in model and training data.
5. Google released "Bard," a rival to ChatGPT powered by LaMDA, which last year made headlines for claims of sentience by a Google engineer.
Technical
6. InstructGPT, ChatGPT's sister model, was released a year prior and uses RLHF to provide detailed responses based on instructions.
7. RLHF is a training method in AI alignment, steering systems towards intended goals and reducing harmful language (if that is the goal). It is an advancement from traditional prediction-based training; predicting the most likely next character.
8. RLHF significantly improves model performance, with models using 1B parameters outperforming those with 175B.
9. ChatGPT was fine-tuned from a GPT-3.5 series model using prompts and responses from various use cases, generated by both human annotators and those submitted through the OpenAI API.
10. Training process overview: 1) Fine-tune GPT-3.5 in supervised fashion (supervised fine-tuning model - SFT), 2) Train a reward model by ranking responses to prompts, 3) Optimize the SFT (policy) against the reward model using proximal policy optimization (PPO), repeat 2 and 3.
Try ChatGPT here: https://openai.com/blog/chatgpt/
Here's the first five paragraphs re-written as a poem by ChatGPT
Verse 1:
ChatGPT's here to take the lead
With accuracy that's all we need
The Google Search killer, that's the name
Human-like responses, it's all the same
Chorus:
100 million users, just two months passed
OpenAI's subscription plan is a blast
Emerging uses, it's all so clear
Technical, non-technical, ChatGPT's always near
Verse 2:
From prompt engineering to classification tasks
ChatGPT's got the skills, it's a task master
Content generation, education and more
Ideation, copywriting, it's got it all in store
Verse 3:
Hyper-realism, it's got its perks
But also risks, that's what irks
Fraud and plagiarism, data ownership too
Misinformation spread, hallucination, it's all true
Outro:
Google's rival, Bard is its name
Powered by LaMDA, it's got the fame
Sentience claims, it's all in the news
But ChatGPT's still the one, no one can refute.
Best,
Marcel Hedman
Nural Research Founder
www.nural.cc
If this has been interesting, share it with a friend who will find it equally valuable. If you are not already a subscriber, then subscribe here.
If you are enjoying this content and would like to support the work financially then you can amend your plan here from £1/month!