Artificial Intelligence and ChatGPT: ChatGPT & Emerging Technology
ChatGPT
"ChatGPT is a conversational language model developed by OpenAI. It is based on the GPT (Generative Pre-trained Transformer) architecture and is trained on a large dataset of conversational text. ChatGPT can generate human-like text in response to a given prompt or question, and can can be used for a variety of natural language processing tasks such as text completion, conversation generation, and language translation." -- ChatGPT
AI Resources
- ChatGPT General FAQCommonly asked questions about ChatGPT answered by OpenAI.
- ChatGPT UpdatesCheck here for official updates and articles about ChatGPT from OpenAI.
- ChatGPT PluginsPlugins that give ChatGPT new capabilities by allowing it to interact with other computer services. Currently only available for plus subscribers. Of particular note is the Wolfram Plugin, which allows ChatGPT to make calls to Wolfram|Alpha and Mathematica while responding to users' prompts, thus enabling it to answer a wide range of mathematical and scientific questions.
Limitations of generative AI
Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback, by Casper, Davies, et al. RLHF is the central technique used to fine-tune LLMs, and involves collecting human feedback on the output, which is then used to modify the LLM's model to produce more highly rated responses. It allows the LLM to go beyond modeling the distribution of its training data.
- Limited knowledge of the world: The dataset ChatGPT, for example, is trained on only contains data up to January 2022 (as of time of writing), meaning that it can not generate responses related to current events.
- Bias: AIs are trained on large datasets, which can lead to inherent biases in the model. This can result in biased or prejudiced responses.
- Lack of Common Sense/Emotional Intelligence: May occasionally generate incorrect, nonsensical, or inappropriate responses.
- Hallucinations: If the user asks for/about something which doesn't exist or which the LLM doesn't know about, it often "hallucinates" or makes up an answer
Attempts to fix these problem often create new problems. For example, Google suspended Gemini's image generation capabilities shortly after its release because when creating images of people, it depicted people of a range of ethnicity and gender expressions, even when not appropriate for the prompt, for example in images of Nazi soldiers (NPR article).
Microsoft Bing Chat
Microsoft Bing https://www.microsoft.com/en-us/bing/apis/llm
- Uses GPT-4 for free
- Connects to internet
- Lists sources that are usually used accurately
- Can be "precise" or "balanced"
LLMs with focuses on particular tasks
Video generation:
- Sora (from OpenAI)
Coding:
Research: The following tools can be used to collect and summarize research articles, starting from natural language questions, and not using exact keyword searches. They can see different collections of research papers, and present their findings in different ways, and it is often useful to use both to see the differences in what they come up with. Be careful with their responses, and be sure to double-check their work.
- GOAT: Who is the Greatest Economist of all Time and Why Does it Matter?A book by economist Tyler Cowen which comes with an LLM trained on the book itself plus a corpus of economics literature which can be queried with questions about the book and more broadly the topics it discusses.