Brief summary of ChatGPT

This content is part of my collection of Machine Learning resources. Check out my repository https://www.github.com/MarcoSteinke/Machine-Learning-Resources and leave a star, if you like my collection of resources on Machine Learning and Artificial Intelligence

ChatGPT:

ChatGPT is a ChatBot which is powered by artificial
intelligence and was created by OpenAI.
It is still considered as a prototype and was released in the november of 2022.

Training Methods:

• RLHF (Reinforcement Learning from Human Feedback
• Proximal Policy Optimization (also reinforcement learning)

I/O:

For the input and output ChatGPT uses GPT 3.5 which is another AI model. GPT is short for
Generative Pre-Trained-Transformer and GPT is an improved version of OpenAI's GPT-3.

Protection against wrong and adversial answers

Before ChatGPT was developed, OpenAI learned a lot from their research on GPT and Codex, another model of OpenAI.
Thus, ChatGPT includes various mechanism which try to protect ChatGPT against wrong and adversial answers.
Since this is a really hard job, ChatGPT already does a good job on this, but still has to get a lot of improvements.

ChatGPT will continuously be improved by human input and feedback.

Training Data:

The training data of ChatGPT consists of texts which were created by humans. This has the purpose of training ChatGPT to
communicate with its users in a natural way. Thus, it uses texts from:

• forums
• social Media
• blogs
• books
• spoken language

Problems with human input:

• In the early stages of the training users did prefer long answers, which resulted in CHATGPT answering with longer texts.
• Algorithmic bias: The training data can cause the model to give biased answers, for example it can interpret the word "CEO" as a succesful white male person

Applications:

The model has the purpose of serving a tool for natural conversation with an AI to inform the human about expressions, definitions, concepts of various areas of the human life.
For example ChatGPT understands all common human languages and can translate between them.
It can also be used to generate text templates, which can then be improved by the human.
ChatGPT is also able to explain complex contexts in simple language and can present a brief overview of certain topics.
If you ask ChatGPT about "Scaling of Software" it will define what Scaling is and will also name different methods of scaling. The users can then answer and pick one of the named methods to step one step deeper into this topic and receive an overview of this certain subtopic.
Thus, complex information could get available for a broader audience by using ChatGPT's simpler explanations.

Another powerful ability of ChatGPT is the generation and explanation of sourcecode and configurations. Using ChatGPT developers are able to generate code for all common programming languages. This does even work for frameworks like Unity or ThreeJS. ChatGPT also explains every piece of code and provides information about the performance of the generated code and how it can bei improved.

Critique:

ChatGPT can in one answer provide a correct and high quality formulation and in the next answer it could provide something completely stupid or wrong, as described by AI expert Gary Marcus. Thus, users always have to remind that the answers of ChatGPT always require verifications, before using them in a serious situation.
Another problem has to do with scientific or very complex backgrounds. In these ChatGPT offers well-written and long answers, which are actually completely wrong in some contexts. This was tested by the physicist Teresa Kubacka by asking ChatGPT about her doctoral thesis in physics.
The german blog author Sascha Lobo is fascinated by ChatGPT, but also fears a huge wave of spam on the internet, which was generated by ChatGPT. This problem forced Reddit to ban content which was generated using ChatGPT.

Sources:

• Learning from Human Preferences. OpenAI, 13. Juni 2017, abgerufen am 11. Dezember 2022 (englisch).
• Paul Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, Dario Amodei: Deep reinforcement learning from human preferences. 13. Juli 2017, doi:10.48550/arxiv.1706.03741.
• Proximal Policy Optimization. OpenAI, 20. Juli 2017, abgerufen am 9. Dezember 2022 (englisch).
• John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, Oleg Klimov: Proximal Policy Optimization Algorithms. 28. August 2017, doi:10.48550/arxiv.1707.06347.
• ChatGPT: The Ultimate Tool for Natural Language Processing and Text Generation. Abgerufen am 16. Dezember 2022 (englisch).
• OpenAI: ChatGPT: Optimizing Language Models for Dialogue (englisch) 30. November 2022. Abgerufen am 5. Dezember 2022.
• Murphy Kelly: This AI chatbot is dominating social media with its frighteningly good essays (englisch) In: CNN. 5. Dezember 2022. Abgerufen am 5. Dezember 2022.
• Eike Kühl: ChatGPT: Gut erfunden ist halb geglaubt. In: zeit.de. 6. Dezember 2022, abgerufen am 10. Dezember 2022.
• ChatGPT is a new AI chatbot that can find mistakes in your code or write a story for you. In: Business Insider. Abgerufen am 9. Dezember 2022.
• Sophia Schmid: ChatGPT - Überblick zum neuen Tool von OpenAI. In: neuroflash. 5. Dezember 2022, abgerufen am 17. Dezember 2022 (deutsch).
• Sophia Schmid: Wo kommt ChatGPT am besten zur Anwendung? In: neuroflash. 12. Dezember 2022, abgerufen am 17. Dezember 2022 (deutsch).
• ChatGPT. In: OpenAI. OpenAI, 30. November 2022, abgerufen am 9. Dezember 2022 (englisch).
• Michael Moorstedt: Künstliche Intelligenz Chat GPT beantwortet Fragen verblüffend klug. In: sueddeutsche.de. 4. Dezember 2022, abgerufen am 10. Dezember 2022.
• Jakob Jung: Stack Overflow verbannt OpenAI ChatGPT. In: ZDnet.de. 7. Dezember 2022, abgerufen am 10. Dezember 2022.
• Sascha Lobo: Das Ende der irrelevanten künstlichen Intelligenz. In: Der Spiegel (online). 7. Dezember 2022, abgerufen am 10. Dezember 2022.
• https://twitter.com/paniterka_ch/status/1599893718214901760. Abgerufen am 17. Dezember 2022.
• https://twitter.com/paniterka_ch/status/1599893839367483392. Abgerufen am 17. Dezember 2022.
• Datenwissenschaftlerin aus Zürich warnt vor ChatGPT und den bösen Folgen. Abgerufen am 17. Dezember 2022 (Schweizer Hochdeutsch).
• ChatGPT — Release Notes. 15. Dezember 2022, abgerufen am 18. Dezember 2022 (englisch).

---