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

LaMDA: Google's New Conversation Technology

blog.google

Post reply on HN
Print view View original

131–140 of 157 posts

Re: LaMDA: Google's New Conversation Technology

#131

Post by fshbbdssbbgdd »

throwaway_dcnt wrote: post #102
TeMPOraL wrote: post #76

Earlier quoted context omitted.

I think what's even more unsettling is realizing people's interests aren't as much predictable as they're being made predictable . My wife and I, we experience a "Facebook is listening" event roughly every month these days. Of course I know they aren't, but these too-well-targeted ads or post recommendations aren't about generic, easy-to-guess interests - they're about completely random things that we just happened t…

Why do you think they are NOT listening? Whatsapp and facebook messenger/chatter thing both have mic access for a lot of people and it is truly bizzare how topics of conversations, even the first time they are broached, end up in facebook ads.

You can decompile these apps and audit them to your heart’s content.

Re: LaMDA: Google's New Conversation Technology

#132

Post by stellaathena »

phreeza wrote: post #60
msoad wrote: post #36

How's this paper's reproducibility? Because more and more I see Google publishing papers with internal data and telling others "if you had this data you would be able to reproduce this results"

Isn't it the same in many cases? E.g. the discovery of the higgs boson is only possible with access to a large particle accelerator, many astronomy studies are only reproducible with access to a large telescope, and some ml studies are only reproducible with access to a large dataset. Sure it's not ideal, but it's better than not publishing the results at all.

Yes, but that data is made publicly available in physics. It’s not kept as a secret so some company can make millions of dollars selling it.

Re: LaMDA: Google's New Conversation Technology

#133

Post by dash2 »

TulliusCicero wrote: post #81
raphlinus wrote: post #52

Disclosure: work at Google, opinions my own. I've had a chance to play with this and it's eerie how good it is. Most impressive to me, it was able to defend itself persuasively when confronted with evidence of how its previous answers were inconsistent (in the version I played with, it tends to be very coherent but frequently wrong).

Also a Googler, and same. I tried to get it to say controversial/inflammatory/bigoted things with leading questions, or questions where the framing itself was problematic, and it was quite adept at deflecting/dodging/resisting. I was very surprised.

So it's an awesome simulation of a prissy bore? ;-)

I'm joking, I know that if it did say controversial things the headline would be "Google has created a robot fascist". You can't win at the moment....

Re: LaMDA: Google's New Conversation Technology

#134

Post by tobr »

eximius wrote: post #126
tobr wrote: post #125

Earlier quoted context omitted.

I think it’s more out of respect for other people than for the bot itself. If your bot is given a feminine name and voice, and accepts being called various sexist slurs, then it becomes a rather offensive depiction of women.

That's a related but somewhat different point. 1. I don't want to feel bad for giving commands to my machine. For avoiding niceties. This is not talking about slurs, etc, but not saying 'please' or some such. 2. I agree it would be nice to avoid negativd depictions of groups. Feminine bots accepting abuse isn't great. I guess the balance for me is kind of like what they person I was replying to said: don't make it a…

I agree that you shouldn’t have to say “please” to set your egg timer or get the weather forecast. It would perhaps be amusing with a bot that refused to do anything unless you were extremely polite, but it wouldn’t be very practical.

What’s interesting to me is how this is never a problem with a traditional menu or button based UI. Maybe I’m muttering “f*ck off” as I click “Remind me later” on the umpteenth alert of the day trying to get me to update this or that, but the software would never know. (On the other hand, you could argue that the software is offending me by trying to get me to do something that is mostly in the developer’s interest, and not allowing me to say “No” - this is where being able to be frank using voice input could be quite liberating)

Re: LaMDA: Google's New Conversation Technology

#135

Post by missingsteps »

A previous step for creating a tool like LaMDA is designing a filter for information, something like rss. Since no something like rss is discussed here I think something is missing.

Another question for the laMDA team: is this chatbot using any logic system, something like prolog, to connect data with goals?, in other terms, could I explain to this chatbot how a prolog system works and hope that the this chatbot is able to use some known facts and rules to deduce new knowledge?, if the system is not able to make use of some simple rules I think it utility is not going to be top notch. Finally, if the chatbot is not crippe it could provide valuable information about its own weak points to be used for (very) bad purposes.

Re: LaMDA: Google's New Conversation Technology

#137

Post by missingsteps »

zitterbewegung wrote: post #17

If you want a freely available system that attacks this problem then Blenderbot is a good choice (made by Facebook)[1] . It is also freely available and integrated with Huggingface. [1] https://ai.facebook.com/blog/state-of-the-art-open-source-ch... [2] https://huggingface.co/facebook/blenderbot-3B

It would be interesting to hear how the google and facebook bots discuss who is the winner in chatbot's technology, and I wonder if the evaluator function of both model agree on who is the overall winner, this imply some scale or measure to divide the knowledge graph into brachs and compare the quality of conversational technology of both models, perhaps using a random forest model to evaluate the performance mean scoring function of each one to get some transfer learning or adjust some hyperparameter. Finally, I would also like some comparison with the wolfram tool. I think wolfram shoud be designing some chatbot centered about discovering new facts or empowering users in the scientific realm.

Re: LaMDA: Google's New Conversation Technology

#138

Post by shannifin »

goliatone wrote: post #79
Mizza wrote: post #27

Earlier quoted context omitted.

I came here to point this out - this is how people from California communicate. They simply take turns talking about themselves, orbiting around the same subject until somebody says something with enough gravity to drag the conversation into its next orbit. It was really off-putting when I first arrived, I got used to it over time, though I still don't like it and it's sad to see it used as the default model for this…

I often interpreted this as a conservative interpersonal communicative approach. It is safer to talk about yourself than to ask questions from people because there risk of inadvertently saying something offensive. Sure, you will find people that are a bit more self absorbed and would just pong about themselves. But there are so many conversational constraints- asking “why” is aggressive, direct feedback is discourage…

I actually find it easier to converse by asking polite general questions about what the other person just said about themselves. (Probably not something like "why?", more like "oh cool, how was it?") Because so many people's most comfortable topic is themselves, it can put them at ease. Can't go overboard with it though.

Re: LaMDA: Google's New Conversation Technology

#139

Post by phreeza »

stellaathena wrote: post #132
phreeza wrote: post #60

Earlier quoted context omitted.

Isn't it the same in many cases? E.g. the discovery of the higgs boson is only possible with access to a large particle accelerator, many astronomy studies are only reproducible with access to a large telescope, and some ml studies are only reproducible with access to a large dataset. Sure it's not ideal, but it's better than not publishing the results at all.

Yes, but that data is made publicly available in physics. It’s not kept as a secret so some company can make millions of dollars selling it.

I think the fact that there are two things called 'data' here obscures the better analogy. The physics data is made available publicly, but you can't reproduce the method of generating it without building a particle accelerator. Similarly, trained models are often made public, but the dataset used to generate it isn't. So perhaps the data in physics is more analogous to the trained model.

Re: LaMDA: Google's New Conversation Technology

#140

Post by shannifin »

Definitely hard to make any judgments without actually being able to play with it.

Still, while perhaps being useful for finding specific answers that already exist (among other things), it's the coming up with creative but constrained solutions that, to me, would be the real pot of gold. That may be beyond a transformer's ability.

Post reply on HN