Skip to content
Search

ChatGPT

lastmod 2026-10-01created 2022-12-0956 edits

ChatGPT is a prototype artificial intelligence chatbot developed by OpenAI that focuses on usability and dialogue. The chatbot uses a large language model trained with reinforcement learning and is based on the GPT-3.5 architecture.

ChatGPT was launched in November 2022 and has garnered attention for its detailed responses and historical knowledge, although its accuracy has been criticized.

Categories

Projects

  • ChatGPT for Google
    • Chrome extension that shows ChatGPT answers next to Google search results (github.com/wong2)
    • Asks ChatGPT the same phrase you typed into Google search and shows its answer on the results page
    • Requires logging in to ChatGPT with your own account
    • Build it directly from source, or install it in Chrome/Edge/Firefox (web store)
    • PRs for Kagi / DuckDuckGo support are also up, so they will likely be added soon
  • Awesome ChatGPT Prompts
  • Poe from Quora - What if ChatGPT talked plainly instead of like C-3PO?
  • Jasper Chat - Better sentences and more imagination than ChatGPT
  • Youchat - ChatGPT + search
  • Ghostwriter Chat from Replit - Generate and debug code while chatting with AI
  • Perplexity.ai - GPT3.5 + Bing. Not conversational, but far more useful because it combines search results
  • ChatHub - Chrome extension for using ChatGPT/Bing/Bard at the same time
  • Scrapeghost - Web scraping library using GPT
  • Telegram ChatGPT Concierge Bot (+ Voice!)
    • Open-source concierge bot using ChatGPT with Telegram and voice
    • Prompt composition and history memory for ChatGPT with LangChainJS, integrates with Google
    • Speech recognition with OpenAI Whisper
    • Generates Text-To-Voice with Play.ht to answer by voice
  • ChatGDB - Using ChatGPT inside the GDB debugger
  • GPT-4 & LangChain - Create a ChatGPT Chatbot for Your PDF Files
    • GPT4 & LangChain - Building a ChatGPT chatbot for PDFs
    • Load many large PDF files into an LLM and chat about their contents
    • When you add a PDF, it extracts the text into many chunks, generates embeddings, and stores them in the Pinecone vector store
    • For a question, it finds the relevant parts of the document with the OpenAI Embeddings API, passes them to GPT, and generates an answer
    • LangChain + Pinecone + TypeScript + OpenAI Embeddings / GPT-3.5 & GPT-4 + Next.js
  • HuggingChat - Open-source alternative to ChatGPT
  • ChatPDF - https://www.chatpdf.com/
    • Upload a PDF or enter a PDF URL, then chat about the PDF's contents
    • On upload, it builds a semantic index of every sentence in the PDF; when answering a question, it finds the relevant sentences and answers through the ChatGPT API
    • The free version supports up to 3 PDFs / 50 questions per day. (PDFs up to 120 pages, 10MB or less)
  • GirlfriendGPT - Building an AI girlfriend with ChatGPT 4
  • GPT-Code UI - Open-source implementation of OpenAI ChatGPT's Code Interpreter
  • AutoGPT
  • NanoChat - The best ChatGPT that $100 can buy

Alternative

ChatGPT is GPT-3.5 with RLHF applied for chat

Lists only open-source projects that apply other language models to chat systems

Bare (no data, no model weights, no chat system)

  • lucidrains/PaLM-rlhf-pytorch - Implementation of RLHF (Reinforcement Learning with Human Feedback) on top of the PaLM architecture. Basically ChatGPT but with PaLM
  • carperai/trlx - A repo for distributed training of language models with Reinforcement Learning via Human Feedback (RLHF), supporting online RL up to 20b params and offline RL to larger models. Basically what you would use to finetune GPT into ChatGPT.

Mildly bare (data, model weights, basic chat through an API)

Full (data, model weights, fancy chat system including TUI and GUI)

Complicated (semi-open source, not actually open source, based on closed models, ...)

  • tatsu-lab/stanford_alpaca - This is the repo for the Stanford Alpaca project, which aims to build and share an instruction-following LLaMA model.
Related links

Fine-tuning

hyper parameters

There are hyperparameters you can use when calling the API of OpenAI's GPT models. Hyperparameters are not trained neural network weights but parameters the user enters directly, and they affect the model's responses. The main hyperparameters are as follows.

Hyperparameter

Default

Range

temperature

1

0.0~2.0

top-p

1

0.0~1.0

max_tokens

inf

Integer

frequency_penalty

0

-2.0~2.0

presence_penalty

0

-2.0~2.0

Let's look at the results you can get with each hyperparameter.

temperature

Controls the diversity and predictability of the results. Higher values give more random (more creative) responses. When you need predictable, consistent responses, use a value close to 0. Usually you adjust either top_p (below) or temperature, not both.

top_p

The top portion of the probability distribution used when sampling words. For example, with top_p=0.9, words are sampled only from the top 90% of the probability distribution. This has the effect of removing outliers. With top_p=0.1, words are sampled only from the top 10% of the probability distribution, so you get more predictable results.

max_tokens

Used to limit the maximum number of tokens in the answer.

presence_penalty

A value greater than 0 discourages words (tokens) that have already appeared from appearing again. In other words, the model becomes more likely to talk about new things.

frequency_penalty

A value greater than 0 discourages the frequency of a specific word (token) from increasing. In other words, it prevents the model from repeating the same thing verbatim.

Predicted Outputs

A Chat Completion API feature that speeds up responses, mainly when the response is predictable. You provide the expected output to the model with the prediction parameter so it can generate results efficiently. Currently (2024-11-24) supported only on the gpt-4o and gpt-4o-mini models.

  • OpenAI recently released a powerful feature called Predicted Outputs
  • It did not get much attention in the tech media, but it deserves attention
  • This feature can reduce API response wait time and greatly speed up generation of output that can be predicted in advance
  • When most of the content is already predictable, you can provide that content to the model as a prediction
  • The model uses this prediction to accelerate response generation and improve performance

Limitations

  • Model compatibility: available only on the gpt-4o and gpt-4o-mini models
  • Cost: rejected prediction tokens are charged the same as generated tokens. Monitor rejected_prediction_tokens to manage costs
  • Unsupported parameters:
    • n (values greater than 1)
    • logprobs
    • presence_penalty (values greater than 0)
    • frequency_penalty (values greater than 0)
    • max_completion_tokens
    • tools (no function calling)
  • Modality limits: supports only the text modality; audio input and output are not possible

ChatGPT models appear to leave watermarks in text

Use case: photo-based location tracking

Image-based location tracking with ChatGPT is spreading as a viral trend

  • OpenAI's latest models o3 and o4-mini combine image analysis and web search to perform highly accurate location inference
  • Users feed o3 restaurant menus, street scenes, portraits, and so on, trying to guess locations like in the game "GeoGuessr"
  • Privacy concerns are growing; OpenAI has safeguards but lacks clear preventive measures
  • OpenAI says it has systems to detect and respond to abuse, but critics point out that actual restrictions are still lacking

Trend of image-based reverse tracking with ChatGPT

  • Recently, tracking the location in images with ChatGPT has become a spreading trend among users
  • OpenAI's new models o3 and o4-mini can reason about visual information in images
  • Even blurry or distorted photos can be analyzed after preprocessing such as cropping, rotating, and zooming
  • Combined with web search, this feature performs excellently at inferring the location of cities, landmarks, restaurants, bars, and so on

ChatGPT used like GeoGuessr

  • Users ask o3 to "act like GeoGuessr" and try to estimate locations from Instagram story captures, street photos, menus, and so on
    • GeoGuessr is an online game where you are shown a photo and find its location
  • ChatGPT infers locations from visual clues alone, without prior conversation history or EXIF metadata
  • o3 in particular outperformed GPT-4o in certain situations

See also

Favorite site

Article

author Zer0 sitemap