Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Google scrambles to manually remove weird AI answers in search

theverge.com

Post reply on HN
Print view View original

321–330 of 387 posts

Re: Google scrambles to manually remove weird AI answers in search

#321

Post by dleeftink »

avar wrote: post #263
jononor wrote: post #208

Earlier quoted context omitted.

But current models must be retrained to incorporate new information. Or to attempt to fix undesirable behavior. So just freezing it forever does not seem feasible. And because there is no way to predict what has changed - one has to verify everything all over again.

Would you by extension argue that e.g. modern relational database aren't deterministic in their query execution? Their query plans tend to be chosen based on statistics about the tables they're executed against, and not just the query itself. I don't see how that's different than the LLM case, a lot of algorithms change as a function of the data they're processing.

At least in case of Bigquery, I have fought with indeterminist-like issues many times over, especially when dealing with window functions that aggregate floats from different compute nodes, where rows cannot be further sorted on a unique column (i.e. the maximum sorting granularity for rows with a similar column of interest to compute a window function over has been reached).

Inconsistent results could be resolved by introducing additional out of data constraints (e.g. incremental hashes), but it can take quite a while to figure out at which exact point in a complex query these constraints need to be introduced.

Beyond that, some functions might still produce different results between runs, e.g. `ml.tf_idf` and `ml.multi_hot_encoder` that take some approximation liberties. Whether these functions are relational in the traditional sense is up for debate.

Re: Google scrambles to manually remove weird AI answers in search

#322

Post by dleeftink »

Terr_ wrote: post #253
dleeftink wrote: post #159

Earlier quoted context omitted.

Maybe some models can be deterministic at a point in time, but train it for another epoch with slight parameter changes and a revised corpus and determinism goes out the proverbial (sliding) window real quick. This is not unwanted per se, and the exact feedback loop that needs improving to better integrate new knowledge or revise knowledge artefacts incrementally/post-hoc.

I think what you're describing is that training/execution effects aren't predictable . It is still "deterministic" in that training on exactly the same data and asking exactly the same questions should (unless someone manually adds randomness) lead to the same results. Another example of the distinction might be a pseudo-random number generator: For any given seed, it is entirely deterministic, while at the same time…

True in the ideal case, but taken together (e.g. corpus retraining, temperature settings, slight input changes, initial descent parameters) unpredictability and indeterminism become difficult to distinguish. Especially in the distributed training case, training data may be propagated to different nodes in different order (e.g. when leaving it to a query optimiser), which makes any large-scale training operation difficult to reproduce exactly.

Re: Google scrambles to manually remove weird AI answers in search

#323

Post by hnfong »

lukev wrote: post #83
jerf wrote: post #27

"Achieving the initial 80 percent is relatively straightforward since it involves approximating a large amount of human data, Marcus said, but the final 20 percent is extremely challenging. In fact, Marcus thinks that last 20 percent might be the hardest thing of all." 100% completely accurate is super-AI-complete. No human can meet that goal either. No, not even you, dear person reading this. You are wrong about som…

I feel like there's some semantic slippage around the meaning of the word "accuracy" here. I grant you, my print Encyclopedia Britannica is not 100% accurate. But the difference between it and a LLM is not just a matter of degree: there's a "chain of custody" to information that just isn't there with a LLM. Philosophers have a working definition of knowledge as being (at least†) "justified true belief." Even if a LLM…

The Gettier problem is an indication that the definition has (at least) a bug.

There are other formulations of "knowledge" which does not involve justification, see eg. Gnosticism.

Of course, for a publicly available frequently used service, the "JTB" formulation of knowledge is probably the only one we can practically use, but this kind of indicates that the whole idea of search engines, knowledge systems, or expert systems is flawed due to the Gettier problem.

Re: Google scrambles to manually remove weird AI answers in search

#324

Post by charleslmunger »

onemiketwelve wrote: post #314
zogrodea wrote: post #186

This approach to remove bad search suggestions manually reminded of a different approach Google once took, where they weren’t satisfied with manually tweaking search results but rather wanted to tweak the algorithm that produces these results when there were bad results. 'Around 2002, a team was testing a subset of search limited to products, called Froogle. But one problem was so glaring that the team wasn't comfort…

reminds me of this time we kept getting bugs in our app from a super old android phone from 2011. we could never reproduce it with any other hardware. There were only 4 users with this phone. We spent weeks trying to fix it but couldn't. I suggested we buy the 4 users a refurb phone from another brand. Would've cost like $300 total. Nope, not allowed. Something about not giving up as engineers. We spent 3 weeks tryin…

A similar story:

https://issues.chromium.org/issues/41088357#comment32

Years ago I bought some Korean phone with a foot long antenna and TV tuner on eBay because it was crashing at a disproportionate rate. It was just the nature of Android development at the time.

Re: Google scrambles to manually remove weird AI answers in search

#325

Post by fragmede »

charleslmunger wrote: post #324
onemiketwelve wrote: post #314

Earlier quoted context omitted.

reminds me of this time we kept getting bugs in our app from a super old android phone from 2011. we could never reproduce it with any other hardware. There were only 4 users with this phone. We spent weeks trying to fix it but couldn't. I suggested we buy the 4 users a refurb phone from another brand. Would've cost like $300 total. Nope, not allowed. Something about not giving up as engineers. We spent 3 weeks tryin…

A similar story: https://issues.chromium.org/issues/41088357#comment32 Years ago I bought some Korean phone with a foot long antenna and TV tuner on eBay because it was crashing at a disproportionate rate. It was just the nature of Android development at the time.

That was an amazing read, how'd you come across it?

Re: Google scrambles to manually remove weird AI answers in search

#326

Post by leptons »

avery17 wrote: post #317
leptons wrote: post #295

Earlier quoted context omitted.

It's not only mentally ill people that are at risk, but anyone that doesn't know it's not a good idea to put "non-toxic" glue in pizza cheese. That includes a lot of not-mentally-ill but just plain dumb people. Google didn't need to tell people that glue+pizza is a reasonable thing to do, or even just a thing. It sure did frame it like it was a legitimate response. And Google didn't even have to reply with this or an…

Thats what parents and mentors are for. We as a society should not have to break our backs bending over backwards to stop people from doing stupid things. People can make their own descisions and be responsible for them. If they lack proper guidance, well that just sucks.

>Thats what parents and mentors are for.

It's nice that you have a parent that cares about how you are raised, or "a mentor". Do you realize that not everyone has that?

>We as a society should not have to break our backs bending over backwards to stop people from doing stupid things.

Sure, let's take down all the speed limits and see what happens. Let's tell people it's an option to wear seatbelts and see what happens. Let's deregulate everything and hope for the best. Sounds reasonable?

>People can make their own descisions and be responsible for them. If they lack proper guidance, well that just sucks.

There's this thing called "human nature". I think you should do some reading about it.

Re: Google scrambles to manually remove weird AI answers in search

#327

Post by fragmede »

duskwuff wrote: post #316
gerdesj wrote: post #284

Earlier quoted context omitted.

"users would pay someone else to do the search." My notion isn't a rehash of Google Answers. Google pays the "someone else", not you.

That sounds like a way for Google to spend money, not make it.

Have you seen how much Google spends on their campuses and engineers?

Re: Google scrambles to manually remove weird AI answers in search

#328

Post by fragmede »

avar wrote: post #270
leptons wrote: post #203

Earlier quoted context omitted.

The right answer is no rocks. Some mentally ill person could type that in and get "eat 1000 rocks" and then die from eating rocks, and that would be Google's fault. It's not funny. I have no doubt right now there are at least 50 youtube videos being made testing different glue's effectiveness holding cheese on a pizza. And some of those idiots are going to taste-test it, too. And then people will try it at home, some…

> The right answer is no rocks. Sand is considered a "rock". If you live in e.g. the USA or the EU you've definitely inadvertently eaten rocks from food produce that's regulated and considered perfectly safe to eat. It's impossible to completely eliminate such trace contaminants from produce. Pedantic? Yes, but you also can't expect a machine to confidently give you absolutes is response to questions that don't even…

Salt is rock, and most everyone eats plenty of that.

The LLM is clearly being dumb, but the underlying science of the question is actually interesting. Iron is another interesting one. Run a magnet though iron-fortified cereal.

Re: Google scrambles to manually remove weird AI answers in search

#329

Post by ruszki »

praisewhitey wrote: post #121
jerf wrote: post #27

"Achieving the initial 80 percent is relatively straightforward since it involves approximating a large amount of human data, Marcus said, but the final 20 percent is extremely challenging. In fact, Marcus thinks that last 20 percent might be the hardest thing of all." 100% completely accurate is super-AI-complete. No human can meet that goal either. No, not even you, dear person reading this. You are wrong about som…

You're looking at it the wrong way, the goal should be 0% inaccurate. Meaning for the 20% of things it can't answer, it shouldn't make something up.

Nothing can be sure that it hasn’t inaccurate or incomplete knowledge. So that can’t be a goal either.

Re: Google scrambles to manually remove weird AI answers in search

#330

Post by fragmede »

bluefirebrand wrote: post #165
jerf wrote: post #27

"Achieving the initial 80 percent is relatively straightforward since it involves approximating a large amount of human data, Marcus said, but the final 20 percent is extremely challenging. In fact, Marcus thinks that last 20 percent might be the hardest thing of all." 100% completely accurate is super-AI-complete. No human can meet that goal either. No, not even you, dear person reading this. You are wrong about som…

> You are wrong about some basic things too Sure, but probably not "add glue to pizza to get the cheese to stick" wrong...

The thing about that is that polyvinyl acetate is that's what's in elemers glue and is also used in chewing gum, and chocolate and to make the surface of Apples more shiny, so you're probably eating glue, we just don't like to call it that. emulsifier is a better description.
Post reply on HN