How ChatGPT Became Possible - Rise of LLMs

What is GPT-3.5 and Why it Enabled ChatGPT?

If you enjoy articles about A.I. at the intersection of breaking news join AiSupremacy?here. I cannot continue to write without community support. (follow the link below). For the price of a cup of coffee, Join 140 other paying subscribers.

https://aisupremacy.substack.com/subscribe

Will 2023 be the year of Conversational A.I?

Hey Everyone,

Large language models have been shown to achieve remarkable performance across a variety of natural language tasks using few-shot learning, which drastically reduces the number of task-specific training examples needed to adapt the model to a particular application.

OpenAI’s?text-davinci-003?was trained on a more recent dataset, containing data up to June 2021. This is what we normally refer to as GPT-3.5 and what the viral ChatGPT demo embodied for the public.

Open Source PaLM Architecture with RLHF

More recently in late December, 2022, it appears that the first open-source equivalent of ChatGPT has arrived:

See it on GitHub

It’s an implementation of RLHF (Reinforcement Learning with Human Feedback) on top of Google’s 540 billion parameter PaLM architecture. Check out the LinkedIn comments on this?post.

Just weeks after the demo of ChatGPT launched there?are many live examples?of Chatbots that are similar.

There is also much healthy speculation on how?GPT-4 may be like?(Twitter thread), and how it may produce emergent A.I. and?more emergent behaviors?along the spectrum of for instance chain-of-thought and multi-model tasks.

On November 28th, OpenAI released a new addition to the GPT-3 model family:?davinci-003. This latest model builds on?InstructGPT, using reinforcement learning with human feedback to better align language models with human instructions.

Due to the larger?LLM of GPT-4, extended training period (GPT-3 was released in June, 2022 - going on 29 months) and with improved methods of RLHF, ChatGPT as a real product will produce some interesting competition for Google’s LaMDA even potentially impacting their future of dominating Search advertising and consumer search in general.

Share

OpenAI believes it will become very profitable in the near future, and is thus in negotiation with Microsoft according to reports for more a multi-billionaire dollar funding. Not only is Microsoft Research an incredible hub for A.I. research, Microsoft’s superior business diversification is thus allowing for A.I. to get special funding and for OpenAI to use its supercomputer among other things.

Microsoft realizes that Generative A.I. could make coders more productive, speed up game development and many other useful things boots its Cloud, Azure’s adoption. The evolution of GPT-3 to GPT-3.5 and the products of GPT-4 in 2023, will be very interesting to watch.

GPT-3.5 was key for ChatGPT

OpenAI trained this model using Reinforcement Learning from Human Feedback (RLHF), using the same methods as?InstructGPT, but with slight differences in the data collection setup.

1. They trained an initial model using supervised fine-tuning: human AI trainers provided conversations in which they played both sides—the user and an AI assistant.
2. They gave the trainers access to model-written suggestions to help them compose their responses.
3. They mixed this new dialogue dataset with the InstructGPT dataset, which we transformed into a dialogue?format.

To create a reward model for reinforcement learning, they needed to collect comparison data, which consisted of two or more model responses ranked by quality. To collect this data, we took conversations that AI trainers had with the chatbot.

They randomly selected a model-written message, sampled several alternative completions, and had AI trainers rank them. Using these reward models, we can fine-tune the model using?Proximal Policy Optimization. They then performed several iterations of this?process.

From the lay person’s perspective we might conclude that ChatGPT felt “different”. OpenAI’s remarkably capable, if flawed, GPT-3 was perhaps the first to demonstrate that AI can write convincingly — if not perfectly — like a human. GPT-3.5 and ChatGPT’s demo made us realize these tools could be useful in our lives and work tasks in novel ways.

Matthias Bastian, of?the Decoder, is one of my favorite writers and journalists to follow if you are into these topics. I’m a big fan of breaking news in the GPT watching, related to OpenAI and GPT-4’s launch in the next few months.

According?to OpenAI, GPT-3.5 was trained on a blend of text and code published prior to Q4 2021. Like GPT-3 and other text-generating AI, GPT-3.5 learned the relationships between sentences, words and parts of words by ingesting huge amounts of content from the web, including hundreds of thousands of Wikipedia entries, social media posts and news articles.

Microsoft Likely to get first Dibs on ChatGPT

ChatGPT is fine-tuned from a model in the GPT-3.5 series, which finished training in early 2022. You can learn more about the 3.5 series?here. ChatGPT and GPT 3.5 were trained on an Azure AI supercomputing?infrastructure.

Always try to understand the evolution of LLMs from actual academic papers:

Share AI Supremacy

PAPERs

• [2005.14165] Language Models are Few-Shot Learners?[July, 2022]
• [2107.03374] Evaluating Large Language Models Trained on Code?[July, 2021]
• [2201.10005] Text and Code Embeddings by Contrastive Pre-Training?[January, 2022)
• [2009.01325] Learning to summarize from human feedback?[February, 2022]
• [2203.02155] Training language models to follow instructions with human feedback?[March, 2022]

Of these I’d say I think?InstructGPT?is the real breakthrough.

These?InstructGPT?models, which are trained with humans in the loop, are now deployed as the default language models on OpenAI’s API.

This?RLHF is getting better?and companies like Google and ByteDance also are doing important R&D with regards to it.

You can read the full article here.

If you enjoy articles about A.I. at the intersection of breaking news join AiSupremacy?here. I cannot continue to write without community support. (follow the link below). For the price of a cup of coffee, Join 140 other paying subscribers.

https://aisupremacy.substack.com/subscribe