Conversational AI: ChatGPT alternatives and their advantages

A beautiful scene

Since its first release by OpenAI in November 2022, ChatGPT has attracted a lot of attention from not only researchers but also the remaining part of the world. It features a long sought-after dream of machines having the ability to converse naturally with humans. However, ChatGPT is not the only chatbot that can do the job. In this article, we introduce some of the most competitive alternatives for ChatGPT. Each of them has unique advantages, making them suitable for different kinds of applications.

I. Other OpenAI services or ChatGPT in disguise

1. GPT-3 and Codex

Unlike ChatGPT and other chatbots in this article, GPT-3 and Codex are not explicitly designed to make conversation. Instead, they are generative models with the traditional text-completion style. By giving GPT-3 / Codex a piece of text (a.k.a the prompt), you expect it to return another piece of text that meaningfully completes that prompt. For instance, tell GPT-3 something like “It is raining today, so the baseball game” and it may respond with ” should be canceled”.

While serving for the text-completion task is their nature, GPT-3 / Codex can be adjusted to answer questions (just like what ChatGPT does). The idea lies in a technique called few-shot learning [1]. In simple terms, if we include a few examples of questions and answers in the prompt, then in the output, we should have the answer to our question similarly. See the example below.

The question we want to ask:
Why shouldn’t I run a marathon when having flu?

The prompt we give to GPT-3 / Codex (note how we give 2 examples of unrelated question-answer pairs):
Answer the following questions:
[QUESTION]
Who is Barack Obama?
[ANSWER]
Barack Obama served as the 44th President of the United States.
[QUESTION]
Which country is the strongest in table tennis?
[ANSWER]
China is the country that has the most table tennis players. It is no doubt, the most popular country in table tennis. They have the strongest national team in table tennis. They also won almost every table tennis medal at the Olympics.
[QUESTION]
Why shouldn’t I run a marathon when having flu?
[ANSWER]

The output:
Running a marathon when having flu is not a good idea. It will make your flu worse. It will also make you feel very tired.

Advantages of GPT-3 / Codex:

  • We have more control over the output to be generated. ChatGPT outputs the text in any format it prefers. However, by giving examples in the prompt for GPT-3 / Codex, we implicitly guide the model to give the answer in a similar format to those examples. In the case above, the model learned that the answer should be concise and to the point, and it generates the text exactly following this format.
  • Besides the text-completion endpoint, OpenAI recently expanded GPT-3 / Codex with 2 other tasks: text-edit, i.e. given an input and an instruction (both in natural language format), the model will modify the input according to the instruction; and text-insertion, i.e. inserting some relevant text into a pre-defined place in the input [2]. If we need a model for one of these 3 tasks, using GPT-3 / Codex will probably give better results than using ChatGPT.
  • Similar to ChatGPT, we can try out GPT-3 and Codex for free on OpenAI’s Playground website. They also support API to query the models programmatically.

2. Microsoft Bing and Edge

Microsoft is actively integrating ChatGPT into its products, including Bing Search and the web browser Edge. Unfortunately, at the time of writing, the versions of chatbot-integrated Bing and Edge are locked behind a waitlist, so not many people can try them out.

Nevertheless, soon enough, the technology will be ready and all users will have access to these “new-era” search engines and web browsers. An unofficial teaser of the product can be found here.

A sample conversation with Bing's chatbot
A sample conversation with Bing’s chatbot. Taken from Bing.com

Advantages of Bing and Edge chatbots:

  • Unlike ChatGPT, which has no access to the internet, the chatbots integrated into Bing and Edge have the ability to search the web and consult the search result to generate answers. This not only increases the factual correctness of the generated responses but also allows the chatbots to answer questions about recent information, e.g., about some events on the news today.

II. Alternatives that are based on OpenAI’s models

1. ChatSonic

ChatSonic is an AI built upon GPT-3.5 models. While ChatGPT is like the output of a research project, which serves as a minimum working example of AI that can converse; ChatSonic, on the other hand, is a fully-fledged product that is equipped with loads of convenient extensions to serve end users.

A sample conversation with ChatSonic
My conversation with ChatSonic

Advantages of ChatSonic:

  • Like the new Bing Search, ChatSonic also incorporates search results to get the most updated information from the internet for generating answers. Notably, sometimes, it cites references for some sentences in the answer.
  • Users can give input (and get output) through voice rather than typing text.
  • ChatSonic can also generate images (alongside texts) as answers.
  • We have the option to talk to ChatSonic as a specific personality such as a coach, an interviewer, or a poet.
  • It is easy and free to try out ChatSonic on its website. APIs are also available.

The downside of ChatSonic is that it is regarded to be bad at coding.

2. Jasper Chat

Jasper Chat is a conversational AI focusing on business-related purposes. The chatbot is based on multiple models, including GPT-3.5 from OpenAI. It is advertised to be more proficient in business use cases like content creation, marketing, and sales.

A sample conversation with Jasper Chat
A sample conversation with Jasper Chat. Taken from Jasper.ai

Advantages of Jasper Chat:

  • Jasper Chat should give better answers to business-related questions.
  • Recently, they integrated search results into the chatbot, making it more up-to-date with new information.
  • There is a free 5-day trial for Jaster Chat. However, you need to register for an account and input your credit card numbers.

III. Alternatives unrelated to OpenAI

1. Character AI

The selling point of Character AI is, as indicated by its name, the ability to act as some specific persons/roles. The chatbot was instructed to learn the tone and traits of those characters so that it can mimic them. Imagine you can sit at home and chat with the world’s richest billionaire Elon Musk, isn’t it cool? Moreover, it is even possible to have conversations with historical figures such as Socrates and Napoleon Bonaparte, or even fictional characters like Super Mario.

A sample conversation with Character AI
My conversation with Character AI (the AI pretends to be Elon Musk)

Advantages of Character AI:

  • The best thing about Character AI is, of course, its variety of characters you can talk with. There are tons of available characters you can find on its website right now.
  • You can go as far as creating new characters yourself.
  • Some (but not all) characters can generate images.
  • It is completely free to try out Character AI on its website.

2. You Chat

You.com is originally built as a search engine. However, unlike traditional search sites such as Google and Bing, which show a list of links, You tries to summarize the content from those links into a piece of text and show it to the user. This idea is very close to having a general-purpose chatbot that can read the web and provide answers to users’ questions. The only difference is about memory: a chatbot should have the memory of prior interactions when answering the subsequent question, while a search engine normally doesn’t. You.com has integrated this memory capability into the website, making it a real and free chatbot for everyone.

My conversation with You Chat
My conversation with You Chat

Advantages of You Chat:

  • As a search engine, naturally, You Chat can traverse the web and get the latest information. It can also cite the sources for its generated statements.
  • You Chat is completely free to use (however, you need to create an account).

3. Microsoft’s GODEL

GODEL [3] is a dialog model open-sourced by Microsoft in the summer of 2022. In contrast to other publicly available language models that are specialized in text completion, GODEL targets the task of answering in dialogs. One feature of GODEL is the ability to answer questions that are grounded in external texts. For example, suppose you trained/fine-tuned GODEL on data collected up to the end of 2022. Now, you are in 2023 (the time of this writing) and you want to ask GODEL about the latest news. While most other language models are unable to answer your question as they don’t have the required knowledge, GODEL can still give a good answer provided you give it the latest news (e.g. by giving it access to some news websites). Notice that this technology (or similar of its kind) is already applied in some of the chatbots we mentioned above so that they can integrate the search results in their responses.

A sample conversation with GODEL
A sample conversation with GODEL. Taken from Microsoft.

Advantages of GODEL:

  • The model is open-sourced, which means anyone can download, use, and fine-tune it locally as needed.
  • It is entirely free. However, you need a computer to run it.

4. Meta’s LLaMA

Meta (the parent company of Facebook) announced and open-sourced LLaMA [4] on February 2023. LLaMA is a series of highly-optimized language models (from 6B to 65B parameters) pre-trained on large amounts of text. All models in this series exhibit great performance on text-related tasks. For example, LLaMA-13B outperforms GPT-3 (170B parameters) on most benchmarks, while the largest version, LLaMA-65B, is comparable to the best existing models such as Google’s PaLM (540B parameters) even though they are just about 1/10 in size.

A sample conversation with LLaMA-65B.
A sample conversation with LLaMA-65B. Taken from Meta.

Advantages of LLaMA:

  • The source code of LLaMA is publicly released, so anyone can download, train, and use the models. However, at the time of this writing, Meta hasn’t officially released the pre-trained weights yet.
  • It is free.

5. Meta’s Blender Bot

The Blender Bot [5] is another series of dialog models coming from Meta. Since the first launch of Blender Bot version 1 in 2020, the company has released updated versions 2 and 3 in subsequent years, 2021 and 2022, respectively.

A sample conversation with Blender Bot 3.
A sample conversation with Blender Bot 3. Taken from Meta.

Advantages of Blender Bot:

  • The Blender Bots are open-sourced. You can find model version 1 on hugging face. However, versions 2 and 3 seem to be available only in the US.
  • The models are free.

References:

  • [1] Language Models are Few-Shot Learners, Tom B. Brown et. al., 2020, arXiv.
  • [2] New GPT-3 capabilities: Edit & insert, OpenAI, 2022, article.
  • [3] Godel: Large-scale pre-training for goal-directed dialog, Peng et. al., 2022, arXiv.
  • [4] LLaMA: Open and Efficient Foundation Language Models, Touvron et. al., 2023, arXiv.
  • [5] BlenderBot 3: a deployed conversational agent that continually learns to responsibly engage, Shuster et. al., 2022, arXiv.

Leave a Reply