Here are the large language models that are shaping the future of AI

Here are some of the most prominent LLMs that are leading the way.

GPT-3.5

You may have heard of this one. It powers one of the most impressive demonstrations of NLP today - ChatGPT.

GPT stands for Generative Pre-Trained Transformer. It is a deep-learning neural network that helps ChatGPT understand and respond to human text.

OpenAI achieved this by training the model on vast amounts of text data from all over the internet, but the wrinkle is the additional fine-tuning that OpenAI used to teach the model.

It used Reinforcement Training from Human Feedback (RLHF) to improve predictions.

In simpler terms, the LLM trained under human supervision. They categorised each action it took as desirable or undesirable. It allowed the model to learn like humans do by making mistakes and with positive reinforcement.

GPT-4, scheduled for late 2023, will focus on improving parameters and fine-tuning them further. OpenAI says that this will also save computing costs since it will be a more efficient system.

LaMDA

Google's Language Model for Dialogue Applications (LaMDA) powers the recently announced Bard AI.

This LLM was trained on hours of conversational dialogue, which Google says allows it to pick up subtle nuances in linguistics, and gives it the ability to hold open-ended conversations.

Google has already announced a successor to the model, LaMDA 2, which is more finely tuned and can provide recommendations on user queries.

LaMDA 2 uses Google's Pathways Language Model (PaLM) with 540 billion parameters, allowing it to analyse even more complex queries.

Also Read | ChatGPT: Will the AI legal framework come of age in India?

WuDao 2.0

The biggest model in existence, with a whopping 1.75 trillion parameters, is China's WuDao 2.0.

Created by the Beijing Academy of Artificial Intelligence (BAAI) and introduced in January 2021, WuDao 1.0 was the first out of the gate.

In May 2021, BAAI unveiled the second revision titled WuDao 2.0.

It is similar to GPT but uses a vast library of parameters. Comparatively, GPT runs on 175 billion parameters, whereas WuDao has 1.75 trillion. WuDao 2.0 can simulate human speech and generate content.

MT-NLG

Jointly developed by Nvidia and Microsoft, the Megatron-Turing Natural Language Generation (MT-NLG) is the successor to Microsoft's Megatron-LM and Nvidia's Turing NLG 17B.

The models were trained on Nvidia's Selene ML supercomputer using a dataset of 530 billion parameters.

The 105-layer deep neural network can undertake a vast set of natural language tasks like - completion prediction, reading comprehension, intelligent reasoning, language inferences and more.

Also Read | AI for weather forecasting: How Google and DeepMind researchers approach the problem

Bloom

One of the newest models on the list, BigScience Large Open-science Open-access Multilingual Language Model (Bloom), is an open-source LLM built by a consortium of more than 1000 AI researchers.

The model can generate text in 46 languages and code in 13 programming languages.

Built on a dataset of 176 billion parameters, the LLM can be accessed on the cloud or a local machine.