OpenAI trained a model called ChatGPT, a chatbot which interacts conversationally. According to the announcement, the dialogue format makes it possible for ChatGPT to answer follow-up questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests.
Users were encouraged to use the bot, provide feedback and help the developers learn about its strengths and weaknesses. During the research preview, usage of ChatGPT is free, and you can still test it here.
Since the announcement on the 30th of November, social media has been awash with conversations about the bot. While some people have claimed that AI has finally come to take over, some have claimed that ChatGPT is the new ‘google’ while others are unimpressed.
Stack Overflow has warned all users against sharing responses provided by the AI chatbot, with the mods stating that the volume of incorrect but plausible-looking replies was just too great for them to deal with.
“The primary problem is that while the answers which ChatGPT produces have a high rate of being incorrect, they typically look like they might be good and the answers are very easy to produce,”Mods
“We need the volume of these posts to reduce…So, for now, using ChatGPT to create posts here on Stack Overflow is not permitted. If a user is believed to have used ChatGPT after this temporary policy is posted, sanctions will be imposed to prevent users from continuing to post such content, even if the posts would otherwise be acceptable”, they stated.
What you should know about ChatGPT
ChatGPT is an experimental chatbot created by OpenAI based on its autocomplete text generator GPT-3.5. A web demo for the bot was released last week and has since been enthusiastically tested by users around the web.
The user interface of the bot encourages users to ask questions and, in response, provides impressive and fluid responses for various queries, including writing and debugging lines of code and creating songs, poems, and TV scripts, among other things.
There is also the tendency for ChatGPT to generate incorrect answers. While many users have admired the bot’s abilities, others have noticed that it consistently produces convincing but misleading answers.
How ChatGPT was trained
ChatGPT was trained using Reinforcement Learning from Human Feedback (RLHF). Human AI trainers provided conversations in which they played both sides—the user and an AI assistant. The trainers had access to model-written suggestions to help them compose their responses.
Open AI also collected comparison data which consisted of two or more model responses ranked by quality to create a reward model for reinforcement learning. To collect this data, the conversations that AI trainers had with the chatbot were collected.
The developers then randomly selected a model-written message, sampled several alternative completions, and had AI trainers rank them. Using these reward models, the model can be fine-tuned using Proximal Policy Optimization.
Limitations of ChatGPT
One of the most significant limitations of ChatGPT is that it sometimes writes plausible-sounding but incorrect or nonsensical answers.
Although, with the tendency to provide incorrect answers, ChatGPT is quite careful not to produce dangerous information. It still uses a sizable language model that has been trained to generate believable-sounding output rather than a system that is designed to represent knowledge propositions about the environment and then reflect them verbally.
This is one of several well-known failings of AI text generation models, known as large language models or LLMs.
Because there are no hard-coded guidelines for how particular systems work, they tend to produce “fluent nonsense.” However, these systems analyse and churn out large output, and it is difficult to tell how significant their errors are.
ChatGPT is also sensitive to tweaks to the input phrasing. For example, given one phrasing of a question, the model can claim not to know the answer but, given a slight rephrase, can answer correctly. Ideally, the model would ask clarifying questions when the user provided an ambiguous query. Instead, our current models usually guess what the user intended.
Patricia’s CEO, Hanu Fejiro, recently chatted with the bot and expressed his excitement about the AI being able to generate a coding solution to Bitcoin price fluctuations.
Like the CEO, perplexed by the discovery but unclear of the solutions, we cannot definitively state whether or not that code is accurate. However, we know that the chatbot can include inaccurate biographical information if you ask it to write a biography of a famous person, for instance.
In ChatGPT’s case, Stack Overflow has judged for now that the risk of misleading users is too high.
Recently, Carl T. Bergstrom, a Biology Professor at the University of Washington, in describing the AI chatbot, said, “…They’re trained on largely correct texts, and they reproduce some aspects of these texts faithfully. Enough seems right that they seem credible.”
“But when they fail, they don’t throw an error code or report “data not available” or anything. They just make shit up. And they’re trained to make up stuff that sounds entirely plausible. That’s how they get things right as well as how they get things wrong”, he added.
After all that, ChatGPT might not be a horrible resource for coding solutions and related or not-so-difficult problems. The propensity to offer solutions to intricate issues connected to actual life phenomena, though, calls for scepticism.
Get the best of Africa’s daily tech to your inbox – first thing every morning.
Join the community now!