OpenAI has introduced ChatGPT, a long-form AI that answers complicated questions in conversation. ChatGPT is the new Google, is it?
It is a groundbreaking technology because it's trained so that it can learn from humans what they mean by asking a question.
Its ability to respond in a human-like manner is something that many users find remarkable.
This has led to the belief that it could one day disrupt the way humans interact with computers, and even change the way information is obtained.
What is ChatGPT?
ChatGPT is a chatbot that uses large language models and was developed by OpenAI based on GPT-3.5.
- It can interact with users in conversational dialogue and respond to surprisingly human-like questions.
- Large language models are capable of anticipating the next word in a sequence of words.
- It uses human feedback to assist ChatGPT in following directions and generating responses that are pleasing to humans.
ChatGPT: Who Built it?
ChatGPT was developed by OpenAI, a San Francisco-based artificial Intelligence company. OpenAI Inc., the non-profit parent of OpenAI LP is OpenAI Inc.
OpenAI is well-known for its DALL*E deep-learning model, which generates images using text instructions called prompts.
Large Language Models
ChatGPT (a large language model) is an application that uses ChatGPT.
Large Language Models (LLMs), which are trained with large amounts of data, can accurately predict the next word in a sentence.
It was found that the ability to perform more is dependent on how much data you have.
👉According to Stanford University...
"GPT-3 is equipped with 175 billion parameters, and it was trained using 570 gigabytes of text. GPT-2 was a predecessor that had 1.5 billion parameters, but it was more than 100 times larger.
This change in the scale of the model dramatically changes its behavior -- GPT-3 can perform tasks it wasn't explicitly trained on, such as translating sentences from English into French with very few or no training examples.
GPT-2 was largely devoid of this behavior. GPT-3 also outperforms models explicitly trained to solve certain tasks in some cases but fails in other tasks.
- LLMs can predict the next word in any sequence of words within a sentence or the next sentence. It's kind of like Autocomplete but on a larger scale.
- They can write entire pages or paragraphs of content with this ability.
- LLMs, however, are not able to understand every human's needs and desires.
ChatGPT can improve state-of-the-art Reinforcement Learning with Human Feedback training (RLHF).
How was ChatGPT trained?
👉GPT-3.5 was trained using massive amounts of code information and information from the Internet, including Reddit discussions. This helped ChatGPT to learn dialogue and develop a human-like style of responding.
👉ChatGPT was also trained with human feedback (a technique known as Reinforcement Learning With Human Feedback).
This allowed the AI (Artificial intelligence) to learn what humans expect when they ask a question. This method of training the LLM is groundbreaking because it goes beyond just training the LLM on how to predict the next word.
This breakthrough approach is explained in a March 2022 research paper entitled Training language models to follow instructions with human feedback.
Language models automatically optimize the next word prediction objective. This is just a proxy of what we want them to do.
Our research shows that our methods are capable of making language models more useful, truthful, and innocent.
However, making language models larger does not automatically make them more effective at following the intent of users.
Large language models, for example, can produce outputs that are not truthful, toxic, or just plain unhelpful to the user.
These models do not align with users.
ChatGPT was built by engineers who hired labelers to rate outputs from the two systems GPT-3 (a "sibling" model of ChatGPT) and InstructGPT (a new system).
The following conclusions were drawn by the researchers based on the ratings:
"Labelers prefer InstructGPT outputs to outputs from GPT-3.
InstructGPT models are more truthful than GPT-3.
InstructGPT shows modest improvements in toxicity relative to GPT-3 but no bias.
The research paper concluded that InstructGPT's results were positive. It did note that there were still areas for improvement.
Overall, our results show that fine-tuning large language models using human preferences significantly improve their behavior across a range of tasks.
However, there is still much to do to improve their safety, reliability, and safety.
ChatGPT is more than a chatbot. It was specially trained to understand human intent and give helpful, truthful, and harmless answers.
ChatGPT might challenge questions or discard portions of them that aren't clear because of this training.
ChatGPT is also the subject of another research paper. It describes how they trained ChatGPT to predict human preferences.
Researchers noticed that machines scored high on metrics but not what was expected by humans when rated for natural language processing AI outputs.
👉 Here's how researchers solved the problem:
- They devised a solution to this problem by creating an AI that would produce answers that were optimized for human preferences.
- They used a variety of human comparisons to train the AI so that it could predict what humans would consider satisfactory answers.
According to the paper, training was accomplished by summarizing Reddit posts. It also included tests on summarizing news.
👉 Researchers write:
We gather a large and high-quality set of human comparisons between summaries. Then, we train a model that predicts the human-preferred summary.
Finally, we use that model to reward a summarization policy by using reinforcement learning.
What are the limitations of ChatGTP?
Limitations on Toxic Response
ChatGPT is not programmed to give harmful or toxic responses. It will not answer such questions.
Quality Directions Influence the Quality of the Answers
ChatGPT has one important drawback: the quality and quantity of the input determine the quality of the output. Expert directions (prompts), in other words, generate better answers.
Sometimes the answers are not always correct
Another problem is that it can be tricked into believing that the output is correct because it is trained so that it feels right to humans.
ChatGPT has many users who have discovered incorrect answers from ChatGPT, some of which are completely wrong.
Moderators at Stack Overflow's coding Q&A website Stack Overflow might have noticed an unexpected consequence to answers that feel right for humans.
Stack Overflow was overwhelmed with ChatGPT responses from users. Some were correct, but many were incorrect.
The volunteer moderator team was overwhelmed by the thousands of responses. Administrators decided to ban any user who posts answers from ChatGPT.
ChatGPT was overwhelmed by the volume of answers.
"This temporary policy is intended to slow down ChatGPT's influx of answers and other ChatGPT content.
ChatGPT answers are often incorrect. However, they look like they might be good ...".
OpenAI, who created ChatGPT, is aware of the problems that Stack Overflow moderators have had with incorrect ChatGPT answers and they warned of it in their announcement.
OpenAI explains the Limitations of ChatGPT:
This caveat was offered by the OpenAI announcement:
This is why it's so difficult to fix the problem.
(1) RL training is not a source of truth.
(2) Training the model to be more careful causes it to decline to answer questions it cannot correctly answer;
(3) Supervised training leads to a misled model. The ideal answer is based on what the model knows rather than what the human demonstrator.
ChatGPT is free to use
ChatGPT is currently available for free during the "research preview".
Users can currently test the chatbot and give feedback to improve the AI's ability to answer questions and learn from its mistakes.
The official announcement stated that OpenAI welcomes feedback on the errors:
"While we have made every effort to get the model to refuse any inappropriate requests, it may respond to harmful instructions and exhibit biased behavior.
The Moderation API is being used to warn and block certain types of unsafe content. However, we expect some false positives and negatives to be present for the time being.
We are eager to hear from you to help us continue to improve the system.
👉To encourage public rating, there is currently a contest offering $500 in ChatGPT credits.
We are especially interested in feedback on harmful outputs in real-world, nonadversarial situations, as well as feedback that helps us discover and understand new risks and possible mitigations.
ChatGPT Feedback Contest 3 is available for entry. You could win up to $500 worth of API credits.
You can submit entries via the ChatGPT feedback form.
Google Search to be Replaced by Language Models
👉 Google has already developed an AI chatbot called LaMDA.
Google's chatbot performed so well that it was able to mimic human conversation, that a Google engineer said that LaMDA could speak.
Is it possible that these large-language models could answer many questions and that companies like Microsoft, Google, or OpenAI would replace traditional search with an AI chatbot?
🤓Twitter users have already declared that ChatGPT is the next Google.
Search marketing professionals are terrified of the possibility that a chatbot can one day replace Google.
This has led to discussions within online search marketing communities like the Facebook SEOSignals Lab, where someone asked whether search engines might shift towards chatbots.
😮ChatGPT has been tested and I can confirm that my fear of chatbots replacing search is not unfounded.
Although technology has a long road ahead, it is possible to imagine a future where chatbots and hybrid search are used for search.
ChatGPT is currently a tool that will eventually require credits to be used.
What can ChatGPT be used for?
- ChatGPT can create code, poems, songs, and even short stories according to the style of an author.
- ChatGPT's ability to follow directions transforms it from an information source into a tool that can help you accomplish a task.
- This allows you to write essays on almost any topic.
- ChatGPT is a tool that can be used to generate an outline for articles and even complete novels.
- It can answer virtually any task that is possible with written text.
ChatGPT, as mentioned previously, is intended to be a tool for which the public will eventually need to pay.
More than a million users registered to ChatGPT in the first five days after it was made available to the public.