Google Bard and GPT-4 over each other

Finally, there’s universal access to Bard, Google’s dream rival to GPT-4. A great opportunity to ask the question of questions to both AIs. When GPT-4 and Google Bard are asked about each other, it leads to some remarkable results.

Google Bard now available for every US and UK IP address

Until now, May 10, 2023, it was not possible to test Bard without a waiting list. Now that is possible, although you must have an American IP address. This can be done via a certain VPN subscription, but an excellent alternative is the built-in free VPN of the Norwegian browser Opera. This browser can be downloaded here.

Choose the North and South American option when you enable the VPN, which defaults to Europe. The functionality is only available for Google accounts that have undergone an age check. If you have made a payment via Google Pay, that should in principle be OK.

GPT-4 on Google Bard

GPT-4’s dataset has been truncated at a point in late 2021. Asking GPT-4 about an AI that did not yet exist at that time is therefore not possible via the OpenAI interface.

Bing Chat, but with internet access

But luckily there is a way out. Microsoft’s search engine, Bing, also has GPT-4 in the form of Bing Chat. Unlike the OpenAI version, it can indeed have internet access. And of course Microsoft makes extensive use of this, to praise its own AI in comparison to that of arch-rival Google.

It is not certain whether the information on which Bing Chat relies is still up to date. Bard reportedly now has the much better and larger language model PaLM, with over half a trillion nodes and a more efficient structure.

Google Bard about ChatGPT-4

Google’s priority is to remain the most popular search engine and prevent the lucrative advertising market from being cannibalized by its own AI. After a disastrous start, see here, it now seems to have worked out reasonably well, as witnessed by the latest release of Bard. Note: this only works with a British or American IP, which is possible with the aforementioned Opera or, for example, with Proton.

Compared to previous versions, which received quite a bit of criticism, Bard appears to have progressed in leaps and bounds. Presumably this due to the improved language model PaLM.

When we asked Bard the same question, we got a modestly worded response.

In this first version, Google Bard calls itself strong in accurate answers, understanding complex questions, and safety. Weaknesses include difficult access, less creativity, and slower text generation. (The first two points are correct, but GPT-4 is also often slow at peak times).

We couldn’t resist giving it a second try.

Striking in this version is that Bard is much more enthusiastic about the competitor than during our first attempt and contradicts himself considerably. Here, Bard calls itself particularly good at natural-sounding dialogue, while GPT-4 is said to be better at generating factually correct text. The consistency of Google Bard is therefore not very great.

We therefore understand the hesitation of Google’s developers to put this model online immediately. But, as Bard also says, the model is in full development. What is so does not have to stay that way.