OpenAI’s ChatGPT was, for a while, synonymous with AI. Launched in 2022, its position as the first widely known and used large-language model meant that for many, ChatGPT was AI. However, as the market has matured, there are signs that OpenAI’s position is under threat. We look at some of the leading models, Meta’s Llama 3.1, ChatGPT-4 and Google’s Gemini to see which is the best.
What they have in common
All three AIs are large-language models. In essence, this means they are trained on huge datasets that allow them to interpret and provide natural language questions and answers. This can be remarkably effective, but can also be their downfall.
Because they are not artificial general intelligence, they aren’t actually answering the questions, instead, they are using their mammoth datasets and algorithms to predict what comes next. Although this can be incredibly effective, it means LLMs effectively generate answers that look right, but might not actually be correct. As many high-profile, and often amusing, examples have shown, the AI’s confidence can be misplaced.
The models are also starting to progress being their LLM roots, adding features and offering more ways to interact, and get, responses. This is increasingly considered a key to the monetisation of AI. While sites like DALL-E can be amusing, its the integration into common devices, like our phones, that is essential if AI companies are to see a return on their investment.
Llama 3.1
Meta’s entry in the AI race is Llama. Only introduced to the public in February 2023, it’s now at version 3.1 and is one of the largest AIs available. Trained on a dataset comprising 15 trillion tokens, the model is available in three sizes, 8 billion, 70.6 billion, and a massive 405 billion parameters.
In theory, the larger the model, the better. And Meta’s models have received positive press for their capabilities. Many reviews have noted it seems more adept at avoiding traps that other AIs fall into, correctly interpreting human intent, demonstrating logical consistency in exchanges, and avoiding hallucinations.
Llama is also unique in making its models available for download, although this has resulted in some criticism since it means that the safety barriers in place on most AIs can be disabled.
ChatGPT-4
OpenAI’s flagship, ChatGPT-4 is notable because it is — with limits — available for free use, but also because it is pushing it’s LLM AI further away from the traditional text input and response.
It has deepened the integration of the model with other OpenAI products, such as DALL-E. The model can now accept and produce any combination of inputs, whether that’s text, images, or even voice and video.
What is, perhaps, most impressive, is the improvement in processing and response time over ChatGPT’s previous versions. Speech input, for example, previously had a very noticeable lag, as the model interpreted and responded. Its response time has now been reduced to milliseconds. It means that for premium users with access, voice interaction with the model is so fast, it’s comparable to normal human dialogue.
Google Gemini
Perhaps the first question is which version of Gemini are we comparing? In true Google fashion, there are multiple versions. Launched in February 2022 as Bard, it was merged with Duet AI in February 2023 to become Gemini and now how has four models, Ultra, Pro, Flash, and Nano.
Each are optimised for different uses, Nano, for example, is intended for on-device use, while the Gemini chatbot, which most people will use, is based on Flash, which is smaller, and optimised for speed.
Gemini’s model use a Mixture-of-Experts architecture. Instead of being a single network, it is a collection of smaller expert networks. In theory, this makes it more efficient, however, it adds the challenge of determining priority between the networks which can, in turn, affect the answers provided.
Which is best?
Choosing the best AI is an almost impossible task. And much will depend on your use. However, based purely on ability to perform the tasks it is asked, Meta’s Llama 3.1 is commonly said to be the better performer. Benefiting from its huge dataset, it tends to outperform competitors in many of the common tests, especially those which are designed to catch AIs out.
However, it’s always dangerous to write the competitors off. Google’s size means it won’t give up the AI battle any time soon, and its integration with Android devices means it is likely to have a significant advantage in collecting both training data and real-life feedback. OpenAI, meanwhile, has a close association with Microsoft and has again taken the lead by extending their feature set to create a much more versatile AI.
While Llama 3.1 might be edging out a lead at the moment, you can expect that to be wiped out the next time a competitor releases an update.
