Introduction

Why is it 'ChatGPT' competitors? Why not Llama competitor or PaLM competitor or BLOOMZ competitor? I believe, it is because of the mainstream presence of ChatGPT, their early adoption by significant masses, their huge investments, news presence and their demonstrated success. You can offer a different perspective, it is quite welcome.

How this information may help you?

The Generative AI space is getting crowded by the day. A review like the one below will get us updated quickly on the broad spectrum of players in the space.

If you are concerned with AI/ML strategy or piloting an AI tool or exploring ChatGPT this article will be of help to you. If you want to keep yourself updated on the latest trends in AI or to find answers for your curiosity questions or if you were driving in a single alley of one player, this information will be a resource for updating yourself and pivoting from your current position.

Please do add your comments, recommendations and opinions to make this article more useful to the community.

ChatGPT overview

ChatGPT is an AI tool built with the techniques of Machine Learning, Natural Language Processing and Language Model.

A model is a systematic representation of a phenomenon in the real world (for example, a weather model).

A language model is one that reflects the syntactic and semantic constructs of a language. Hence such a trained language model is able to understand the language and it can produce output in the given language.

A Large Language Model is an extension of the language model. LLMs are trained with a very large number of parameters (for example, ChatGPT was trained with 175 billion parameters).

In simple terms, parameters are the internal calculation units used by a model to store information about the data processed by the model. Please refer the 'References' section below for a more detailed explanation about model parameters.

Bringing it all together, what we can understand about ChatGPT is, it is trained using large volume and variety of language data, it is modeled to interpret the language data and it can generate output complying to the language rules. ChatGPT model version 3.5 could take only textual data as input, where as ChatGPT model version 4 could take textual data and image as input to provide predicted output.

ChatGPT competitors

Following is a compilation of a brief summary of each major competitor of ChatGPT. The competitors listed are considered based on ChatGPT's own resources or based on the resources mentioned by the ChatGPT resources or by independent bodies such as Stanford University, known for AI related discussions.

Amazon, Inflection, AI21 studio, Cohere, Anthorpic, Google, Stability, OpenAI, HuggingFace and Meta are listed with a brief summary of their motivation in the figure above.

There are more companies developing Generative AI models. HELM by Center for Research on Foundation Models (CRFM), for example, gives a comparison of around 40+ model versions on the basis of evaluation using MMLU (Massive Multitask Language Understanding).

The Foundation Model Transparency Index, also by CRFM shows the ten competitors that are described in the Table 1 above.

Another canvas that we can use to get an understanding of LLMs is given below.

We can observe the timeline when the LLM model versions were launched and we can recognize some of the models such as Jurassic-1, BLOOM, Llama, Claude, PaLM etc Though we couldn't find the AWS Titan, we could find Alexa. We could not find Inflection, Command and Diffusion in the picture above, I could not identify a proxy name or a reason from the available information.

Performance of Inflection-1 on six widely used criteria

The six standards on which the LLMs are evaluated are shown in the picture above. You can find details of these six evaluation criteria in the References section.

GPT-4 performance on MMLU (Measuring Multi-task Language Understanding)

The performance of the latest model version, Generative Pretrained Transformer-4 (GPT-4), on the MMLU is shown in the picture above along with the languages in which GPT-4 can handle.

Llama 2 model description

Llama 2 is trained with 7 billion to 70 billion model parameters and it can take an input text of 4096 tokens (words) without losing its context.

Conclusion

The technology space of Generative AI and LLMs is fast paced and new models are coming into the research and commercial arena. It takes time to read comprehensively about them in one space. As the number of dimensions in which these models could be understood are really multi-fold, this article has tried in a modest setting to achieve the objective of a quick-scan-understanding on the leading models.

References

1. ChatGPT and large language models in academia: opportunities and challenges https://link.springer.com/article/10.1186/s13040-023-00339-9
2. Model parameters: https://medium.com/analytics-vidhya/what-are-model-parameters-in-deep-learning-and-how-to-calculate-it-de96476caab
3. Syntax vs semantics: https://scholar.harvard.edu/files/chierchia/files/1999_syntaxsemanticsinterface.pdf
4. Weather model: https://www.weather.gov/about/models
5. Foundation models: https://www.nature.com/articles/s41467-022-30761-2
6. Semantic search: https://mitpressbookstore.mit.edu/book/9783642250071
7. Retrieval Augmented Generation (RAG): https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00530/114590/Improving-the-Domain-Adaptation-of-Retrieval
8. Amazon: https://aws.amazon.com/bedrock/titan/
9. Popular benchmarks for evaluating LLMs, 1. MMLU Measuring Multi-task Language Understanding: https://arxiv.org/pdf/2009.03300v3.pdf
10. Popular benchmarks for evaluating LLMs, 2. GSM8K Grade School Maths: https://paperswithcode.com/dataset/gsm8k
11. Popular benchmarks for evaluating LLMs, 3. HellaSwag, commonsense natural language inference: https://paperswithcode.com/paper/hellaswag-can-a-machine-really-finish-your
12. Popular benchmarks for evaluating LLMs, 4: WinoGrande, An Adversarial Winograd Schema Challenge at Scale: https://paperswithcode.com/paper/winogrande-an-adversarial-winograd-schema
13. Popular benchmarks for evaluating LLMs, 5: TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension, https://paperswithcode.com/dataset/triviaqa
14. Popular benchmarks for evaluating LLMs, 6: BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions, https://paperswithcode.com/dataset/boolq
15. Inflection: https://inflection.ai/inflection-1
16. Cohere: https://cohere.com/models/command
17. Claude2: https://www.anthropic.com/index/claude-2
18. Google Palm2: https://ai.google/discover/palm2/
19. Stability Diffusion: https://stability.ai/stable-diffusion
20. OpenAI, GPT-4: https://openai.com/research/gpt-4
21. Huggingface, BLOOMZ: https://huggingface.co/bigscience/bloomz
22. Meta, Llama 2: https://ai.meta.com/llama/
23. HELM: https://crfm.stanford.edu/helm/latest/?group=mmlu
24. A comprehensive overview of LLMs: https://arxiv.org/pdf/2307.06435.pdf
25. Zero-shot learning: https://swimm.io/learn/large-language-models/zero-shot-learning-use-cases-techniques-and-impact-on-llms