The ChatGPT artificial intelligence chatbot developed by OpenAI, is being described as most popular internet app ever. It turns out that it was originally released as a “research preview” for two-year-old technology that was being prepared for something much grander. Its reception has come as a complete surprise to the OpenAI team who have been running to catch up and build on its success since November 2022.
What’s under the ChatGPT bonnet?
It’s a fine-tuned version of GPT-3.5, a family of large language models (LLPs) that OpenAI released shortly before ChatGPT GPT-3.5 is itself an updated version of GPT-3, which appeared in 2020. These models are available as APIs so developers can plug them into their own code. Turbo-charged versions were released on 1 March.
ChatGPT was trained using a technique called reinforcement learning from human feedback (RLHF), it’s secret sauce – take a large language model with a tendency to spit out anything it wants and tune it by teaching it what kinds of responses human users actually prefer.
Who are the people behind ChatGPT?
- Jan Leike – alignment team lead and research Associate at the Future of Humanity Institute
- John Schulman – research scientist and co-founder
- Sandhini Agarwal – AI policy researcher
- Liam Fedus
How are they capitalising on its early success?
Adversarial training
This technique is being used by OpenAI to prevent users from being able to trick ChatGPT into behaving badly by pitting multiple chatbots against each other: one attacks another by generating text to force it to produce unwanted responses. Successful attacks are added to ChatGPT’s training data in the hope that it learns to ignore them.
Partnerships
OpenAI has signed a deal with Microsoft to integrate ChatGPT into its Bing search engine and another with management consultancy Bain to incorporate it into clients’ marketing activities.
What’s keeping the ChatGPT team busy now?
Since its release the ChatGPT team have been hard at work monitoring what tens of millions of users are doing with it and ironing out some of its biggest issues, especially around bias and a tendency to expose the nastier side of society and human nature.
If you are interested in incorporating chatbots like ChatGPT and other generative AI tools into your own business operations, internal systems, products or services we offer advice, guidance and support. We also have a team of machine learning specialists and developers who develop bespoke products and services for clients. For more information please get in touch – contact@mondatum.com.
Source: MIT Technology Review