Note: This information can change very quickly, so we’ve not attempted to outline the precise state of each LLM in lots of detail. There are also new AI startups appearing all the time, and existing AI companies are constantly working to improve their LLM models.
Contents
Boosting Your Productivity With AI
(Click any of these links to go directly to that section)
Open vs. Closed Source LLMs
Large Language Models (LLMs) can be categorized as open or closed source.
Open source LLMs are extremely customizable, but they require a lot of technical know-how to use and they’re not usually cutting edge (meaning that they tend to not be quite as effective as the best closed source models), because it costs a huge amount of money to train LLMs to the level of the big closed source alternatives. Most open source LLMs also haven’t gone through much (or any) fine-tuning (with some exceptions like Meta’s Llama 2-Chat). These LLMs are called 'open' because they have publicly available source code and often have pre-trained models freely available for anyone to download and use without charge.
You might benefit from using an open source LLM, if:
you’re a programmer with very particular needs,
you're a start-up founder who wants a proprietary LLM for their organization,
the closed source solutions are prohibitively expensive for your use case, or
you just strongly believe in open-source as a philosophy,
but for the vast majority of people reading this article, a closed source LLM is the best bet.
Closed source LLMs don’t have their source code openly available to the public. And, often, details about their training and fine-tuning are also not available. They are developed and maintained by large, for-profit organizations.
The main reason you’d want to use a closed source LLM is that they have already undergone a huge amount of very expensive training and fine-tuning to help make sure their outputs are safe and applicable for the tasks you might want them for. They’ve got all sorts of capabilities that the open source LLMs just don’t have (because those haven’t been trained and fine-tuned as much). You can explore those capabilities and how they can boost your productivity in our AI Productivity Guide.
Which Closed Source LLM is right for you?
Here’s a list of the closed source LLMs we think it’s best for you to know if you want to boost your productivity at work, as of 2024. We've included a small amount of information about them, to help you get an idea. If you want to see a live-updated leaderboard of LLMs that contains a bit more detail, check out the Chatbot Arena Leaderboard (note: it contains a bunch of open source LLMs too, which can be overwhelming to look at - if you're new to LLMs, you'll it's worth just focusing on just the ones we've also listed below).
Starting with the big LLMs that can be used for most things you might want to use an LLM for:
Producer: OpenAI
ChatGPT 4o is the best on the scene in almost all performance-related respects. It has the most features and capabilities, but it is not free. The free version is 3.5, which is still very good but the other LLMs in this list perform as well as or better than 3.5.
Producer: Microsoft
Base Model: GPT-4o by OpenAI
Free to use, and based on GPT-4o. This can be a good way to use GPT-4o without cost, but it has fewer features than accessing GPT-4o through OpenAI’s website (for example, it’s harder to share conversations and has use limits).
Gemini (formerly, Bard)
Producer: Google
In December of 2023, the LLM formerly known as Bard was updated: its underlying model was changed to one that’s much more powerful (Gemini Pro, rather than PaLM 2). It is now known as Gemini. As of April, 2024, it beat GPT-4 in a few areas, such as in giving more conversational responses and some reports indicate it might be better at complex reasoning and fact retrieval. It also has some features you won’t find in GPT-4 - for example, the ability to search Google for images and display them in responses, and a text-to-speech function, so that it can read its responses out loud, which is great for accessibility.
Producer: Anthropic, in collaboration with Amazon
Anthropic have focused on designing Claude to be "safer" than GPT-4, giving fewer inappropriate responses to prompts. It's also slightly cheaper (at time of writing) than GPT-4o and also has a slightly better "memory" than GPT-4o (it can remember and act on previous interactions that are further back in your chat history than GPT-4o).
Poe acts as a single web platform where a variety of LLMs can be used. It allows users to easily switch between different models, tailoring the AI assistance to specific tasks. While it's a paid product, its versatility can be a significant asset for diverse work-related tasks.
Then there are LLMs that you might want to use for more specific use-cases. These are typically built on one of the models listed above, but with lots of work into supplementing them with additional databases and fine tuning for specific tasks:
Base Model: Both use OpenAI’s models
Specialization: Integrates a database of scientific papers and research materials. Particularly useful for researchers, academics, and anyone needing to analyze or stay updated with scientific literature.
Base Model: Powered by OpenAI's models
Specialization: Perplexity AI is a conversational search engine the main strength of which is that it provides provides citations (with links) to both academic and non-academic sources in all of its answers. It uses a combination of indexing/ranking and LLM-based search to deliver precise and accurate responses.
Base Model: Powered by Google's models
Specialization: Focused on creating presentations. From content, to structure and design, it has the potential to save significant time in preparing for meetings or pitches.
Things to consider, when picking a closed source LLM to use
Here are some factors you should consider when picking an LLM:
What tasks is it best suited to?
What rights does its licensing agreement grant you?
How much will it cost you? Some models charge a monthly fee for a large amount of use, whereas others charge each time you use it based on the size of the input and output (and some offer both pricing options).
How recent is its training data?
Can it access the internet, to search for answers to your queries in real-time?
How safe is it? For instance, how prone is it to be biased or to make up information (i.e., hallucinate)?
What have the organization’s ethics been like? (e.g., consider the human cost of the technology).