Ask HN: Why is Google's Bard still far behind the competition?

7 points by razodactyl ↗ HN
I would have thought it would have grown significantly by now but it doesn't even seem to be able to follow through after the first turn of the conversation.

9 comments

[ 0.17 ms ] story [ 32.5 ms ] thread
It is really weird and disappointing how much worse it is than the OpenAI GPTs.
(comment deleted)
I have noticed in the past that companies that try to do many different products tend to fall behind the competitors that focus on just one product.

Prior to MS Office, there were companies that tried to create product suites that contained Word processors, spreadsheets, and so on. They were all very rudimentary compared to the market leaders in the category. It wasn't until Microsoft decided to buy the market leaders and package them as a suite that we got a great office suite.

I suspect that's Barb. Google is big but it has many products to deal with and it is extremely hard to be a product leader in all. Not only that but search is its main focus. My suspicion is that Google's offering will get better only after it finds a startup they can buy that will give them the edge. Their big advantage is that they don't always have to produce what they need, they have the money to buy what they need to stay ahead.

I would suspect this too but I know it's not the case.

I've been writing my own LLM from the ground up (literally from random weights, not from any other models) and have been training it on publicly available data and even my highly inaccurate model is able to maintain coherence that Bard completely lacks.

For example: Bard just seems to go on a completely unrelated tangent and throw map directions at me for no reason. Given that I'm just one person, how is it that I'm able to nudge my model to do anything useful and Bard has remained stagnant for the better half of a year? It makes me suspect they're up to something bigger :P

Did you train your model to not violate copyright and avoid politically charged statements and to not be biased against race or gender?

Trying to make a model handle everything in a 'safe' manner makes the problem so much harder.

The PaLM api has also been similarly disappointing, I can't even get some of the example prompts in MakerSuite to return anything but empty responses.
just thoughts, I'm probably wrong:

google got paid a big chunk of money + pseudo anti-monopoly economics + dod stuff

Esentially, THE ad and search company running the worlds top AI would raise too much suspicion once the nudging starts to crawl out of the skin of the early adopters. There's more relevant points on this one but I'm busy, sorry.

The DOD and friends really want to know the people and if google becomes too AI driven, the DoD won't know the people, but versions of them that revealed less because AI gave them more than they asked for on the first or second attempt they phraised their question.

Less content consumed via google means less (ads and other) exposure because AI assistance leads to some useful result quicker, hence the big chunk of money, because there's two sides of a coin.