The AI Battle Heats Up – Google’s Palm 2 vs OpenAI’s GPT 4

At I/O, Google announced the release of its new AI model, Palm 2, which aims to take on OpenAI’s GPT-4 head on. They’ve also said they’re training Gemini, which will likely be a GPT-5 competitor.

Here’s a great overview of the key differences between the models…it’s interesting to see that Palm 2 actually does better on a few fronts than GPT-4, though the latter is still better overall at the moment.

Here are the key points made in the video:

1. Palm 2, released by Google, is competitive with GPT-4. Though it may not be as smart, it excels in certain areas such as translation and linguistics, largely due to the diverse multilingual text data it was trained on.
2. Palm 2 is significantly smaller than the largest Palm model (which had 540 billion parameters), but it outperforms it on a variety of tasks. It is estimated to have between 100 and 200 billion parameters, which is similar to GPT-3’s count but delivers performance competitive with GPT-4.
3. Palm 2 operates at an inference speed about 10 times faster than GPT-4. It is also cheaper and faster to serve, and comes in different sizes suitable for various applications, including mobile devices.
4. While Palm 2 was trained to increase the context length of a model beyond that of Palm, it doesn’t hurt its performance on generic benchmarks. If true, this is a significant breakthrough.
5. Palm 2’s performance across multiple languages is not significantly better in English than in other languages, unlike GPT-4 which performs noticeably better in English. This again is attributed to the multilingual text data used for training.
6. Palm 2 demonstrates new emerging abilities, such as in multi-step arithmetic problems, temporal sequences, and hierarchical reasoning. However, there are instances where it gets questions wrong that GPT-4 gets right, and vice versa.
7. In the area of coding, Palm 2’s specialized coding model (Palm 2s) is likely not as good as GPT-4. But Google’s model does cite its code sources, which is a notable improvement.
8. When it comes to following a question and answer format, the fine-tuned version of Palm 2 (Flan Palm 2) scored lower than GPT-4. However, there are areas where Palm 2 outperforms GPT-4.
9. In terms of bias, toxicity, and gender, Google has dedicated much of its report to these issues. However, there was no mention of broader AI impacts, a significant omission compared to OpenAI’s approach.
10. Google has announced Gemini, a next-generation foundation model that’s currently in training and is expected to rival GPT-5. Gemini is designed to be multimodal, efficient at tool and API integrations, and capable of memory and planning. Despite this, there is a lack of discussion on potential AI impacts.

The speed with which all this is moving is simply exhilarating (and terrifying at the same time). With Bard now available in 180 countries, I’ll be giving it more of a try to see how it performs vs ChatGPT in the coming weeks. If you’re doing the same, do reach out and share your insights too!