Google PaLM 2 vs GPT-4 – How do the two LLMs compare?

Google recently announced a new large language model, so how does PaLM 2 compare against GPT-4?

The release of GPT-4 in early 2023 shook the AI world. The LLM (large language model) is rumoured to have over 1 trillion parameters, and since its launch it’s increased the popularity of artificial intelligence to a near uncountable level. As expected, this caused plenty of giant tech companies to realise the full potential of AI, both from a technological advancement perspective, but also in terms of business growth. With companies such as Adobe and Microsoft adopting artificial intelligence as part of their future development, it’s no surprise that Google have focused their attention on Google Bard, and more specifically, PaLM 2 – the large language model that looks to change the internet, especially with changes to Google Search through a generative experience.

PaLM 2 vs GPT-4: Size and Parameters

While not really confirmed, the general rumours suggest that GPT-4 has over 1 trillion parameters. Compared to Google’s PaLM 2, we don’t really know much about the model size specifically. Google’s PaLM 2 report addresses the following:

“… with varying parameters depending on model size. Further details of model size and architecture are withheld from external publication.”

However, we do know what sub-models it has, ordering in terms of descending scale:

• Unicorn
• Bison
• Otter
• Gecko

Gecko, the smallest model, targets mobile devices, and is said to even work offline. Meanwhile, the other models are more suited to complex tasks. However, the largest model in the PaLM 2 family has been confirmed to be “significantly smaller” than the largest PaLM (previous model), though it uses more “training compute.” This basically tells us that while the parameters and size of the PaLM 2 might be smaller than its competitors – a more careful and curated selection of training data has improved the performance of the LLM.

Alongside that, the smaller scale of PaLM 2 lends itself much more to technological innovation – with the development of up to 25 soon-to-release Google products featuring the LLM.

Alongside that, the API will allow third-party developers to utilise it for innovation, with gaming and other casual industries looking to benefit from the performance upgrades that the model will bring.

PaLM 2 vs GPT-4: Reasoning

Again, according to the PaLM 2 report addressed above, it “outperforms PaLM [predecessor] across all datasets and achieves results competitive with GPT-4.” However, no claim is presented that it reasons better than its OpenAI competitor. ChatGPT, the viral chatbot that makes use of GPT-4’s capabilities, is hard to catch out, and has certainly revolutionised our relationship to information. GPT-4 image input has even gone far enough to be able to create website designs from doodles on a napkin, and though ChatGPT can’t quite do this, we’re sure to see it in the future.

It’s going to be hard to compete with with GPT-4’s reasoning ability, and Google’s entry onto the AI scene highlighted just how badly Google Bard image interpretation get things wrong.

PaLM 2 and GPT-4: Usability and application

One of the key differences between the two language models is going to be how they are adapted for practice.

As mentioned before, the significant difference in size and parameters has a lot to offer here. PaLM 2’s smallest model – Gecko – is the fastest and most powerful mobile large language model that can also be run offline. Compared to GPT-4, which requires thousands of crunching through data to provide you with a response, the fact that both PaLM-S and PaLM-M are showing signs of catering to mobile and lower-performance devices, alongside possible having offline functionality.

Conclusion

At the end of the day – it seems as though GPT-4 and PaLM 2 both have a varying set of usages, performance capabilities, and capabilities. While GPT-4 looks as though it’s going to become the AI backbone of information, data, and language on large-scale and mid-scale use, PaLM 2 is gearing itself up to be the champion of portable and small-scale usages, alongside application development and flexibility.

However – there was one, important bit of information that might help you form your own opinion of the two. While researching how many parameters PaLM 2 has, we remembered that PaLM (the first model) had 540 billion parameters. While knowing that Google have decided to withhold how many parameters PaLM 2 has, and also that PaLM 2’s largest model is smaller than the original, we asked Google Bard how many parameters the new LLM has:

It states that PaLM 2 has 540 billion parameters, eerily similar to the old version, and also contradictory to what the PaLM 2 paper-work says. The rational response to this is to realise that the actual size of PaLM 2 is not privy to Google Bard or the language model itself, and it’s claiming the size of the old model in its place.

Is Bard’s mistake here enough to help you decide between PaLM 2 and GPT-4?