Is This the Future of AI?
OpenAI is a San Francisco-Based Artificial Intelligence Company. You’ve probably heard about one of their products that created quite a bit of buzz recently: DALL·E, a deep-learning model that generates images from text instructions called prompts. Check social media for examples, but basically you tell the program what you want it to create – let’s say a penguin smoking a hookah with Taylor Swift – and it will bring together the requested elements into a work of art or a “photograph.” Besides generating such art, it has also generated controversy as people pass on these creations as their own.
Optimizing Language Models for Dialogue
OpenAI, though, has developed another AI model called ChatGPT that answers complex questions conversationally. It has the uncanny ability to interact in conversational dialogue form and provide responses that can appear surprisingly human. What’s revolutionary about this technology is that it’s trained to learn what humans mean when asked questions, something that has proven particularly difficult up until now.
ChatGPT is a large language model (LLM). LLMs are trained with massive amounts of data to accurately predict what word comes next in a sentence. Increasing the amount of data increases the ability of the language models to do more. The latest version of GPT is able to perform tasks it was not explicitly trained on, such as translating sentences from English to French, with little or no training examples.
ChatGPT was also trained using a technique called Reinforcement Learning from Human Feedback – that is, human feedback was essential to fulfilling its promise. The AI learned what humans expected when they asked a question. Training the LLM this way is revolutionary because it goes beyond simply training the LLM to predict the next word (see diagram below).
Source: OpenAI
What Does this Mean in the Real World?
The ability to ask your computer or smart device questions and get the answers you expect is huge. Until recently, LLMs were limited in that they didn’t always understand exactly what a human wanted. Just think of how difficult it is sometimes for other humans to know exactly what you’re asking and in what context.
This ability allows ChatGPT to write paragraphs and entire pages of content. If ChatGPT can do all that and more, it will cause a tectonic shift in our lives. Google’s dominance in searching and retrieving answers may be a thing of the past. ChatGPT has the potential to transform occupations such as law, translation services, social media marketing, and many others. It works similar to Google in that you type a question into a box and hit enter. The chatbox displays the answer, usually in essay format, responding to the question posed.
I imagine that the essay format would be a lifesaver for students who have an essay assigned in school but are too lazy to do their research, but let’s look at how the average person or occupation will be affected.
Take the legal profession, for example. ChatGPT can search for key precedents, provide model pleadings, and outline legal arguments for cases headed to courts. The time saved would be substantial (although I’d bet that won’t reduce the billable hours charged by law firms).
Then there’s translation services. ChatGPT will be able to translate documents for you with much more accuracy than other programs such as Google Translate. Will translation services be necessary after ChatGPT? And then there’s the option to help people learn a new language. Will it make the process easier?
Job seekers may also have an advantage. ChatGPT can help applicants prepare more persuasive resumes, and by gathering information from online sources suggest the best format and list the most pertinent qualifications for a particular job. ChatGPT can also help applicants prepare cover letters that will reference their geographic locations, or expertise, or other items that will be most relevant.
Current Limitations
ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers. Fixing this issue is challenging, as: (1) during RL training, there’s currently no source of truth; (2) training the model to be more cautious causes it to decline questions that it can answer correctly; and (3) supervised training misleads the model because the ideal answer depends on what the model knows, rather than what the human demonstrator knows.
ChatGPT is sensitive to tweaks to the input phrasing or attempting the same prompt multiple times. For example, given one phrasing of a question, the model can claim to not know the answer, but given a slight rephrase, can answer correctly.
The model is often excessively verbose and overuses certain phrases, such as restating that it’s a language model trained by OpenAI. These issues arise from biases in the training data (trainers prefer longer answers that look more comprehensive) and well-known over-optimization issues.
Ideally, the model would ask clarifying questions when the user provided an ambiguous query. Instead, their current models usually guess what the user intended.
While OpenAI has made efforts to make the model refuse inappropriate requests, it will sometimes respond to harmful instructions or exhibit biased behavior. They are using the Moderation API to warn or block certain types of unsafe content, but expect it to have some false negatives and positives for now. OpenAI is eager to collect user feedback to aid our ongoing work to improve this system.
Conclusion
ChatGPT is certainly a game-changer. If the hype and potential are fulfilled, we may be living in the future predicted by science fiction novels and films: A true and helpful AI that interacts with us as though it was human. It could also be a disruptor. Some occupations may be adversely affected, as described above. And perhaps personal assistants who normally would provide proofreading or research could have their job opportunities limited.
Progress is good, but progress is filled with uncertainties. Still, I’ve gone ahead and joined the millions of free accounts to test out ChatGPT with a few questions of my own.