I don’t think either Bing Chat or Bard is ready to be released widely. You need a certain mindset to be able to wring value out of them, and most normal users will not think this way.
That being said, it’s probably fine to release these tools to enthusiastic early adopters who understand how to use them.
And yet it will be, because everyone is wowed so we need something out to capture those wowbucks. There are meetings all over the place to figure out "how to best leverage that technology in our product". While there is some value to raise GPT3 as a technology available in your toolbox, these are largely "let's find a problem that this can fix" meetings, and you can bet your sweet ass we'll see GPT in all sorts of stupid maladapted use cases very soon.
Also, whatever happened to just slapping 'beta' on the product and calling it good to roll out? I'm only half kidding as the issue wasn't that products were labeled 'beta', it's that they seemingly never came out of beta.
[1] Actually even more so today than pre-ML Google since the services now are constantly (incorrectly) deciding that what I'm asking for isn't actually what I want.
e.g, my wife and family members would enter:
* "Distance from Toronto to vancouver" * "How old is the universe" * "What is the capital of Belgium"
I'd enter:
* "Toronto vancouver distance" * "Universe age" * "Capital Belgium"
etc.. you get the point. So for those, I think the ChatGPT approach of "ask question in natural language, get natural language response" is actually quite attractive. Provided it's not outright hallucinating the answer
All of these companies know the tech isn't ready, but they couldn't afford to be left behind when someone moved. It's funny, because this happens quite often in tech. Remember when Tesla announced self-driving? Suddenly Intel were doing deals with MobilEye and BMW, GM did Cruise etc. etc. and then 5-10 years later nothing has been delivered? They all jumped the gun and couldn't afford ot be left behind.
1. The former is entirely in software and has the favorable economics that brings with it
2. Language models don't need to be that good to provide value, whereas self-driving cars basically need to be solved before they become significantly useful
3. There's a cottage industry of secondary startups being built around using language models for various applications, whereas no such thing exists for self-driving (probably because of 2)
You still need data labelling, which is not done by software and which is quite difficult to scale up and to carry out in a consistent manner. I know that most probably there are self-learning LLM models that would partly alleviate that issue, but (absent AGI) I don't think that they'll really solve the problem going forward.
If you are ok with a higher degree of inaccuracy in exchange for minimizing your efford by a lot, ChatGPT is a game changer for many people.
And now, Google who didn't have any intention of releasing LaMDA, panicked..
There’s lots of papers on hybrid information retrieval with LLMs that is probably the right strategy for Google rather than a chat bot (which they tried with Google Assistant which was fairly accurate but not very useful). The main issue with these models right now is they are horrible at recalling facts from their training set but incredible at pulling facts from their prompt. Google has the best IR and LLM researchers in the world, they just need a bolder strategy rather than this FOMO strategy.