Poe AI Chat

Description: Poe.com (Platform for Open Exploration) is a powerful and versatile online platform developed by Quora, that offers an easy, fast and useful AI chat system. It is designed to communicate with a range of AI bots separately that can answer user questions, provide assistance, and perform other tasks related to natural language processing.

Architecture description: Poe.com consists of two main components: an encoder and a decoder. The encoder converts the user's request into a vector representation that contains information about the desired characteristics of the poetry. The decoder generates the text of the poetry word by word using the vector representation as a condition. The encoder and decoder are trained on a large corpus of poetry texts of different languages, genres and authors. The architecture is similar to the Transformer model, which has multilayer attentional and positional encoding. The number of layers, the size of the hidden state and the number of parameters depend on the particular implementation of the model.

Main advantages: Poe.com has several advantages such as:

Free: Poe.com AI is free to use for any category of individuals and for any basic tasks. There is also a possibility for a subscription option for more advanced features, but that’s not required.

Access to a variety of AI models: Poe.com grants access to various state-of-the-art AI models, like GPT-4, GPT-3.5-turbo, Claude, XiaoQia. These models can be used on a variety of tasks including text, translation, creative content and various tasks.

High accuracy: Poe.com can generate poetry that matches the given parameters and takes into account the context, semantics and stylistics of the language.

Efficiency in dealing with big data: Poe.com uses neural networks that can process large amounts of text data and extract useful information from it.

Flexibility and customization: Poe.com allows users to choose different options to create poetry such as theme, style, rhyme, meter and others. Users can also create their own bots to generate poetry in a particular genre or imitating a particular author.

Limitations and Disadvantages: Poe.com also has some limitations and disadvantages such as:

Limited capability: Poe.com is still under development, meaning the AI is currently learning, which can be though in the competing world of more mature AI systems.

High computing resource requirements: Poe.com requires a lot of CPU power and memory to operate, which may slow down the speed of poetry generation or cause the service to overload.

Inability to handle sparse data: Poe.com may have difficulty generating poetry on rare or specific topics for which there is insufficient training data.

Unpredictability and inconsistency: Poe.com may generate poetry that does not always make sense, have logic or quality. Poetry may be unoriginal, repetitive, or contain errors.

Privacy risks: The privacy of the user’s data can be concerning, because Poe.com AI collects and stores user data, conversations and other private texts, which later on are used to train and improve the model.

Examples of uses: Poe.com can be used for a variety of purposes such as:

• Entertainment: users can create poetry for themselves or others, read poetry by other users or bots, play poetry-themed games, or enter contests.
• Education: users can learn poetry skills, analyze different styles and techniques of poetry, study the history and culture of poetry, or compare poetry from different languages.
• Research: users can use Poe.com as a tool to explore the possibilities and limitations of artificial intelligence in text generation, creativity and aesthetics.
• Creativity: users can use Poe.com as a source of inspiration, experiment with new ideas and images, combine poetry with other art forms, or create their own projects based on poetry

Key technical features:

Supported chatbots:

OpenAI

• Assistant (known as Sage) - powered by gpt-3.5-turbo
• GPT-4 (with limits for unsubscribed users)
• ChatGPT – powered by gpt-3.5-turbo
• ChatGPT-16k - powered by gpt-3.5-turbo-16k (not available for unsubscribed users)
• GPT-4-32k - powered by gpt-4-32k (not available for unsubscribed users)

Anthropic

• Claude-2-100k (with limits for unsubscribed users)
• Claude-instant
• Claude-instant-100k (with limits for unsubscribed users)

Meta

• Llama-2-7b
• Llama-2-13b
• Llama-2-70b

Google

• Google PaLM

Latest released bots:

StableDiffusionXL - turns prompts into high quality, expressive images within seconds. You can build bots on top of this with simple prompts. There are hundreds of image generating bots already created.

Web search - is capable of conducting web searches as necessary to inform its responses and is great for queries that require up-to-date information or accurate long-tail facts.

GPT-3.5 Turbo-instruct - from OpenAI is fine-tuned for instruction following and many people are finding it performs very well for specific tasks (e.g., playing chess at Elo 1800)

Solar - is a fine tune of Llama 2 created by Upstage which scored especially highly on third party evaluations.

Code Llama - is a new model from Meta focused on coding.

Model Architecture

• Poe.com AI uses LLMs (Large Language Models), deep learning algorithms trained on massive amounts of text data. They minim human speech and can interact with users through algorithms without an interface. Poe’s goal is to become a global solution for such LLMs using a unified interface.
• Poe.com AI uses a variety of AI models, like GPT-4, GPT-3.5-turbo, Claude from Anthropic and XiaoQia. These models are large language models, trained on a massive dataset of text and code. GPT-4 has been trained on a dataset of 1.56 trillion words and uses 100 trillion parameters. GPT-3.5-turbo is a 175 billion parameter language model, trained on 450 billion words. Claude is a 175 billion parameter language model, trained on a 500 billion words dataset and XiaQia 300 billion parameter language model trained on 1 trillion words dataset.
• All these models are trained on Tensor Processing Unit (TPU) v4 pods, specialized hardware accelerators for machine learning. TPUv4 contains a large number of chips, designed to maximize the speed, performance of matrix multiplication, mathematical operations, predictions and other machine learning algorithms:
  • Each TPUv4 pod contains 4,096 TPUv4 chips
  • Each TPUv4 chip has 275 peak TFLOPS of performance
• The TPUv4 pod is interconnected with a high-speed fabric that provides 10x the bandwidth per chip at scale compared to typical GPU-based large-scale training systems
• The next picture highlights how a TensorFlow program can integrate the TPUv4 pods for creating a logic in our model generation.

• The above code snippet creates a TPU cluster and strategy and after creates a simple TensorFlow model and compiles it using the TPU strategy. This is a way TPU logic can be integrated inside our model generation code. Running this code, will maximize the performance of the model training and evaluation for a better outcome.

Software Architecture

• Poe.com is developed using the Python programming language and TensorFlow open-source library for machine learning and artificial intelligence with a wide focus on training and inference of deep neural networks
• TensorFlow used together with Python, provides a stable Python API and help load data to train the model and deploy it in production environment using TensorFlow Serving
• Check the following picture representing a code snippet of the TensorFlow Serving server integration

• The above code loads a model called my_model.h5 and starts the TensorFlow Serving server. The next line of code will make a prediction request to the server using the tf.serving.predict() function which takes some parameters like the name of the model and the input data, returning the prediction as an output.
• For making prediction requests, TensorFlow Serving provides a dedicated REST API. You can send a POST request in JSON format to the following endpoint: /v1/docs.html/{model_name}:predict
• The output result consists of the prediction made displayed in the JSON format
• TensorFlow library can be used in a wide variety of programming languages besides Python, like JavaScript, C++ and Java. Because of the large compatibility with different programming languages, TensorFlow is used by many large companies like Google, Facebook, Amazon or Microsoft

Advantages of using TensorFlow open-source library:

• Open-source library for machine learning and artificial intelligence, free to use
• Developed by Google Brain team
• Provides stable APIs and integration compatibility with third-party packages like C#, Haskell, Julia, MATLAB, Pascal, Scala, Rust, OCaml and Crystal
• Powerful and flexible platform for machine learning
• Available large documentation and support, please check this link: https://developer.Poe.com/
• Compatible with various hardware and software platforms

Tools and libraries: Poe.com can use different frameworks to implement its text generation models, such as:

• TextBox: is a unified, modular and extensible text generation framework that supports a wide set of text generation tasks and models, including categories of VAE, GAN and pre-trained language models1.
• Pretrained Language Models for Text Generation: this is a survey of the main advances in pretrained language models (PLMs) for text generation, which describes the main architectures of PLMs and how to adapt them to different input data and properties of the generated text.
• Text Generation: is a site that collects research articles, code, and data on the task of text generation, and provides rankings of the best methods on various metrics3.
• A Contrastive Framework for Neural Text Generation: is a paper that proposes a contrastive text generation solution consisting of a contrastive learning objective and a contrastive decoding method that improve the quality and diversity of the generated text.

Selection guidelines: the choice of framework to implement Poe.com depends on the specific purpose and goal of text generation. Some factors to consider while selecting a framework are:

• Task complexity and scope: some frameworks may be more appropriate for complex or large tasks that require high accuracy or efficiency than others.
• Availability and support for code and data: some frameworks may have better or more up-to-date code or data than others, as well as a more active developer or user community.
• Flexibility and extensibility: some frameworks may be more flexible or extensible than others, allowing you to easily add new features or modify existing ones.
• Creativity: some frameworks may be more skilled in creating human-like-text than others, taking into account free-text or creative texts

Known Notes: Poe.com has several features that distinguish it from other text generation services, such as:

Poe supports many languages including Russian, English, French, Spanish, German, Chinese and more. Users can choose the language for their bots and switch between them at any time.

Poe allows users to communicate with bots not only through text, but also through voice and video. Users can record their messages or use speech synthesis to generate the bot's voice. Users can also see the bots' faces, which are created using neural networks.

Poe integrates with other platforms and services such as Quora, YouTube, Spotify, Instagram and others. Users can share their bots or poetry with others, get recommendations from bots, or use bots to find information or entertainment.

Known Applications: Poe.com AI is still a relatively new platform, but users have already been starting to use it in different areas like:

• Chatbots: Customer service bots are being used to help with any task customer service related. From answering customers questions to providing support 24/7
• Virtual and creative assistants: Virtual assistants can help writers, employees, content creators to generate ideas, improve texts or translate documents with a major complexity
• Writing Code: Poe.com can be used to generate code and help developers with complex code snippets, improving code quality and help performance creating fast and efficient code
• Educational tools: Poe.com AI is being used as an educational tool that help students learn and develop new skills, by generating well-written explanations to complex topics and concepts
• Data analysis: The concept of data analysis is a widely used one in many areas. Poe.com AI is able to collect and create data analysis documentations, helping users to extract data and insights from large documents

Important projects that took AI as a basis: Poe.com is inspired and based on different projects in the field of artificial intelligence and text generation, such as:

Quora: is a knowledge sharing platform where people ask and answer questions on various topics. Quora is the creator and owner of Poe.com and also provides a database for training some bots.

OpenAI: is a research organization dedicated to creating human-friendly artificial intelligence. OpenAI has developed several advanced language models such as ChatGPT, GPT-4 and Bard, which are used in Poe.com to generate text.

Anthropic: is a new startup company that aims to create robust and interpretable artificial intelligence. Anthropic created Claude and Claude+, two new language models that are also used in Poe.com for text generation.

Text Generation: is a site that collects research articles, code, and data on the task of text generation, and provides rankings of best practices by various metrics. Poe.com uses this site as a source of information and inspiration for its bots.