Hacker News Reader

Comment by umvi | Hacker News Reader

Seems like Bard is still way behind GPT-4 though. GPT-4 gives far superior results in most questions I've tried.

I'm interested in comparing Google's Duet AI with GitHub Copilot but so far seems like the waiting list is taking forever.

View on HN

I'm not sure Bard and GPT-4 are quite an apples-to-apples comparison though.

GPT-4 is restricted to paying users, and is notable for how slow it is, whereas Bard is free to use, widely available (and becoming more so), and relatively fast.

In other words, if Google had a GPT-4 quality model I'm not sure they would ship it for Bard as I think the cost would be too high for free use and the UX debatable.

How is Bard widely available. ChatGPT is available worldwide, Bard isn't in Europe yet.

There's a good chance ChatGPT gets banned from Europe, whereas Google, despite its fines by EU authorities (most of which are for antitrust), can at least demonstrate that it's set up and continues to maintain GDPR compliance.

Bard is available in 180 countries. https://support.google.com/bard/answer/13575153?hl=en

Why is basically almost all the countries in the world except the EU countries. GP comment about "bard is still not available in europe" still stands.

(Snapshot of the page at the time this comment was written: https://archive.is/hScBl )

If we're going to be pedantic, then "bard is still not available in europe" is not true as it's available in the UK which is in Europe.

I get the general point, but I would say that "everywhere but the EU" is very much "widely available".

Yes, basically everywhere except europe, likely due to regulatory concerns. (Would be interested to know what precisely, but the page doesn't say. Any guesses?)

IMO this is exactly apples-to-apples comparison.

They both represent SOTA of two firms trying for technically the same thing. Just because the models or the infrastructure aren't identical doesn't mean we should not be comparing those to the same standards. Where Bard gains in speed and accessibility, it looses in reasoning and response quality.

On the contrary, Bard is a product not a model. If you want to see the cutting edge capabilities then comparing the GPT-4 API to the bigger PaLM2 APIs available on GCP is probably a more apples to apples comparison.

Bard is more directly comparable to ChatGPT as a product in general, and since it doesn’t have swappable models, comparing it to the opt-in paid-only model isn’t really a direct comparison.

Bard represents SOTA in terms of optimizing for low cost; ChatGPT represents SOTA in terms of optimizing for accuracy. On the SOTA frontier, these two goals represent a tradeoff. ChatGPT could choose to go for lower accuracy for lower cost, while Google could for higher accuracy at higher cost. It's like comparing a buffet to a high end restaurant.

Even if Bard were targeting accuracy, it'd still fall short of ChatGPT, but much less so than it does now. (That said, as a product strategy it's questionable: at some point, which I think Bard reaches, the loss in quality makes it more trouble than it's worth.)

Is this state of the art in terms of fast, incorrect answers? An incorrect answer is often less valuable than no answer at all!

The OpenAI strategy here then seems like a no brainer.

I think there are two viable strategies here: make a model that is useful at the lowest possible cost and make a model that is maximally useful at high costs. Probably some spots in between them as well.

Google's mistake is in thinking that ChatGPT was a maximally useful product at high cost. Right now, ChatGPT is a useful product at a high cost which is nonetheless the lowest possible cost for a useful model.

I cancelled my OpenAI plus because why pay for something you cannot use because it is always slow, down, busy, or returning errors. You cannot build a reliable business on OpenAI APIs either

ChatGPT also spouts falsehoods and makes mistakes on non-trivial problems, there is not much difference here. Both have enough issues that you have to be very careful with them, especially when building a product that will be user facing

Reply on news.ycombinator.com