Comment by finnthehuman | Hacker News Reader

Comment by finnthehuman | Hacker News Reader

>I agree but simple things it's really nice ¯\_(ツ)_/¯

Not to pick on this person is particular, but this is why I am so against surfacing untrustworthy information as trustworthy.

Google's AI "knows" that Austin is the capital of Texas with the same system that thinks the Wright Brothers flew through the core of the earth. So why the fuck would anyone trust google that the capital of Texas is Austin? Because you think you can intuit the boundaries of what their AI will and won't fuck up? [0] And that's a reasonable burden to put on dummies who already think search engine results are somehow more authoritative than the simple text matching they're based on at their core?

How is anyone at google with a brain comfortable with this feature being live?

But these low hanging, blatantly obvious, wrong things are actually a good thing. Because once the obvious things are fixed the un-obvious ones will linger as even more unwarranted trust is be placed in the feature. And it will become even harder to get it though people's thick skulls how goddamn stupid it is to outsource reasoning to machines that can't fucking reason.

[0] Sure, xkcd asked fundamentally unanswerable questions, but google notoriously spent years getting "how long does it take to caramelize onions" wrong.

I think this is a UI problem. Google ought to re-state the question they're answering, which may be slightly or entirely different from what was asked.

In other words, if I Google "What year did Tom Hanks land on the moon", the answer box should say "US astronauts landed on the moon in 1970" (not just "1970"). Give users the ability to assess whether the AI was correct.

Why would the factoid "US astronauts landed on the moon in 1970" be anything more then pattern matching on the input?

I don't think I understand your question? Yes, I'd say the system presumably works via some form of pattern matching.

There is no need for a question to answer, just a match.

Oh, I see where you're coming from! Well, if that's the case, that strikes me as something Google needs to fix. It should be pattern matching both the answer and the likely question!

Why would it say that? US astronauts landed on the moon in 1969 (Apollo 11). Tom Hanks starred in the movie Apollo 13, and the real life mission of Apollo 13 happened in 1970.

Whoops! If you can't tell, I didn't actually look up the years! Thanks for the correction.

What is a bit funny is that US astronauts landed on the moon in 1969, 1971, and 1972 but not 1970.

I don’t think it’s a UI problem. You are presuming the system knows the question it is answering.

Yeah, it's not answering questions, it's giving the best answers which they could find themselves. It doesn't means it's the actual answer, nor that it's to the actual question, that's up to you to confirm.

Google isn't an encyclopedia, it's a search engine. It search inside content and try to find the best content that fit your criteria, it could be from trustworthy source, or not, it could be right or wrong on what you meant. For sure their goal is to give you the closest thing to what you asked, with the best sources, but it's not a guarantee, so it shouldn't be trusted as such.

It's crazy that we all came from the "don't trust Wikipedia" and now some people are arguing that we should be able to trust Google results directly.

This is so wrong in so many ways. Google shouldn't restate anything I type in (this is my opinion, but it's based on solid ethics and a decent understanding of good software engineering practices).

Also, everything about Google's "answer" is wrong, from the date, to the people, to the singleton response to a question that should have multiple answers.

Let's all move to DuckDuckGo and Yandex.

This reminds me of something I once wrote in a conversation about the phrase "when all you have is a hammer, everything looks like a nail" wrt web technologies:

When a tool's convenience grows faster than its reliability and quality, this is what often happens (assuming the tool needs to be more than just convenient). This is just one example.

If you ask "which president invented coffee", it brings up an inventor named George C. L. Washington who founded a coffee company (and invented a process for instant coffee). So it kind of merges three separate pieces of information (has same name as a president, is an inventor, worked on coffee) in an entirely incorrect way.

https://www.google.com/search?q=which+president+invented+cof...

> So it kind of merges three separate pieces of information (has same name as a president, is an inventor, worked on coffee) in an entirely incorrect way.

Is it an entirely incorrect way? Or does it present a very reasonable pass at what a person was likely to be looking for if they had a recollection that there was a President who invented coffee, but wanted to know who that was?

Entirely incorrect? Or a spectacular feat of induction on a cryptic crossword clue? Where does one draw the line?

I think most humans reading the entirely correct sentence "George Washington invented an early instant coffee process" would interpret that to mean "The first US president invented instant coffee." George Washington (the president) had a wide range of careers like surveying and operating a distillery, and further was an avid coffee drinker (the article google cites for me is one claiming he imported 200 pounds of coffee in 1770). Even if you know when instant coffee was invented, the modifier "early" leaves open the possibility that this was some precursor to what we would today call instant coffee.

I think the answer is "google isn't research". Google might be fine for verifying basic/objective fact, such as capitals, but nothing with any sort of nuance, e.g. "Did OJ do it?"

Google is a research tool, not a source to cite.

> ". Google might be fine for verifying basic/objective fact

Unfortunately, the first moon landing happened in 1969, which is an objective fact, so no, Google can't even verify basic facts.

How is anyone at google with a brain comfortable with this feature being live?

Because people at Google test this feature against their query traffic, measure how many answers are correct and useful, compare it to the number of answers that are wrong and misleading, and use the overall data to decide whether it is on balance more helpful than harmful. With billions of queries flowing through the system, of course you will often make a mistake. The correct strategy is not to set the bar at "zero incorrect answers".

It sure seems like the nature of the question and the likely consequences of answering incorrectly is:

- of profound importance

- obviously unknowable with their current solution, given the absurdity of RM’s screenshots

That’s a pretty dangerous gamble at any moment, but I’d say it’s inherently unethical to take the risk in the current climate.

It’s called Google Brain for a reason

These are definitely a bad. Had to tell my girlfriend, who is a high school math teacher, not to trust the info in those boxes.

Reply on news.ycombinator.com