Hacker News new | past | comments | ask | show | jobs | submit login
[flagged] A little Duplex scepticism (daringfireball.net)
35 points by bryanrasmussen on May 13, 2018 | hide | past | web | favorite | 30 comments



Wow, this is a lot of FUD, even for Gruber.

Google does demo plenty of stuff at I/O that isn't available Today. In fact that might be almost everything shown at I/O.

But utter fraud [0], as he suggests here? Is there any precedent for Google doing such a thing on this scale? That seems absurd and, in a company of this size and openness, likely to backfire.

Maybe Gruber doesn't believe anything he wrote and is actually trying to get Googlers to email him in defense of the tech and reveal confidential details of how it works. I don't have any reason to believe this, just asking the question...

[0] I mean, the article is written such that the author could backtrack are any point and say "that's not what I meant, I was just asking the question". But hopefully we can get beyond that facade.


I mean, Gruber is literally accusing Sundar Pichai of lying on stage: https://twitter.com/gruber/status/995538518016487425

I usually like what Gruber has to say about Apple, but this fanboyism is off the charts even for him.


It's classic Gruber. He wants you to definitely come away with the impression that this is a total fraud but he also wants complete deniability that he called it a fraud.


I don’t know if this is classic Gruber. I’ve read a lot of his stuff and it’s always been interesting. Now I haven’t followed him closely either, so maybe I’ve only seen the good stuff, but this is the first time I’ve come away from it thinking he was an idiot.


If you've been following the Markdown debacle closely[1], this is classic Gruber.

[1] https://blog.codinghorror.com/standard-markdown-is-now-commo...


Gruber's stock in trade is ambiguous 1 line response to an extract from an article he's linked to which he can then criticise people for mis interpreting.

Once you've noticed this pattern it's much harder to take his (often very good) long form pieces seriously.


Also his total and complete inability to admit he was 100% wrong about large screened phones being successful is painful.

Gruber/Apple is never wrong - but sometimes reality realigns.


Gruber didn't call it a fraud. He said the demo was not "distinguishable from a fraud," which is true. His actual recommendation is the title of the post, which is that we should have "a little Duplex skepticism."

And yes, Google certainly wouldn't demo a completely non-existent technology, and I'm sure Gruber knows that. But they might very well demo a technology that only works 60% of the time at present. Let's be honest, this wouldn't be the first time we've seen a large tech company demo incredible-seeming too-good-to-be-true tech that turned out to actually be too good to be true and never made it into real world use for one reason or another.

It is kind of weird that the broad tech punditry has just accepted Google's demo at face value with respect to what the tech is actually capable of at present.

For example, Google demoed object removal in Google Photos (https://www.theverge.com/2017/5/17/15654476/google-photos-ob...) last year at IO, and as far as I can tell it never shipped.

If a company in any other industry announced an insane-seeming new technology with a completely non-verifiable demo like this, no credible journalist who covers that beat would report it as credulously as tech journalists have covered the Duplex demo. Gruber is absolutely right that the lack of skepticism around Duplex is baffling and journalistically suspect.


There are basically two hard parts in such interactions - given that one participant is a machine and the person involved does not know that the other part is a machine: understanding the human when he/she thinks to be talking to another human. The second problem is actually understanding a conversation - not just single sentence commands. Much easier: generating sentences and having them spoken with a natural sounding voice.

So, what does it do and how well does it perform? The demo does not really tell. We had such dialogs in research literature 30 years ago, but they were carefully crafted and very domain specific ('ordering a hotel room' via natural language for example).

https://ai.googleblog.com/2018/05/duplex-ai-system-for-natur...

> One of the key research insights was to constrain Duplex to closed domains, which are narrow enough to explore extensively. Duplex can only carry out natural conversations after being deeply trained in such domains. It cannot carry out general conversations.


Can somebody help me understand why Gruber's opinions even matter, in particular, this piece? There's absolutely zero information in this article, zero citations to backup his claims other than good 'ole FUD. I don't think these articles belong on HN. Somebody, please correct me if I'm wrong.


Asking smart questions is a good way to start a discussion. The reaction to his questions here are quite emotional. Why so? If they're irrelevant anyways, why not stay cool about it? I find his questions interesting. Why didn't they demo it why Sundar on the other end of the line for example? What's the name of the shop owners? I can see why people pose these questions.


I don't think accusing someone, a person in particular (Sundar Pichai in this case), of fraud without any evidence whatsoever constitutes as asking smart questions.


A blogger criticizing the competitor of the company that puts food on his table.

Just a thought, imagine how this article would be if the companies were reversed. Would it be the same?

And that's how you recognize objectivity...


There's almost no feasible way to do this demo live: Google had to get permission of the participants to make and use the recording. California is a two-party consent state for recording calls.

Beyond just recording you'd have to get publicity permission to broadcast such a call to the world as marketing for your products.

If they got all this lined up in advance, they'd be accused of rigging the demo anyway, so what's the point of taking the risk and gaining little. The alternative would have been using a Google employee to play the part of the business, but this too would have subjected them to accusations of rigging.

It's not like people are going to forget about the demo if it never materializes. Google has a lot more to lose than gain from faking something this high profile.

I can't think of a good explanation for him wording his misgivings with such uncharitable language. I never had the feeling that he hated or resented Google quite this much.


> There's almost no feasible way to do this demo live: Google had to get permission of the participants to make and use the recording.

Apart from - "We'd like to call you next wednesday to demo some phone technology - it'll just be a normal booking call, but it will be recorded and go out live on stage. Is that OK?'


You could have put someone on both on stage and make him interact with duplex.

Of course it could also have been staged but more difficult. By picking a successful recorded try you tend to obfuscate failure rate.

As opposite Apple often demonstrate feature on stage with real call to outside people or live Siri command. And as with real tech it sometimes fail. They always have a backup hardware and they sometimes use it if the first try is a fail.


This raises another point. To what extent is Google recording the calls made by Duplex? In France you have to pre-announce any recording before the call and I suppose in many other countries it is the same.


Remember Bill Gates Blue Screen Demo? [1] I think to play it safe, let alone where AI is involved and you should expect the unexpected, recorded is "safer" predictable than live.

I do however "believe" Google in a way that all their demos that I remember turned out as real products working in real time, and in many cases, serving 1B+ users daily.

[1] https://www.youtube.com/watch?v=UjZQGRATlwA



If I were to create a service like duplex, the first couple of years of the service would be low paid humans. Everything would be recorded to create the transcripts for the ai to learn from.

I would expect Google to work in a similar way.


Dupe: https://news.ycombinator.com/item?id=17053053 was on the frontpage yesterday.

Speaking of which, 'dang et al., would it be hard to add a full ISO 8601 timestamp as a "title" attribute to the date permalink? I get that "1 day ago" is more readable, but there are cases I'd love to know the actual date and time something was posted.


Submissions and comments now display a date when they have been posted sufficiently long ago. E.g. https://news.ycombinator.com/item?id=1

The cutoff seems to be more than a few months, though.


Why would Google do such a con? This technology is clearly plausible so it doesn't really make sense.

I assume it was just fear of demo gremlins that made them use a recording. Also, you could argue it is unethical to broadcast a phonecall with someone live, and ruins the fun if you get permission in advance.

I assume they did it, then got permission afterwards, and Google being Google, it didn't occur to them that people might want to talk to the callee.


> Why not demo it live? Why only play recordings?

Because consent.


What they could've done is do another demo with a live call, not to a real restaurant but to their employee on stage instead. The employee would respond according to a script, but if Duplex reacted accordingly that would be a hell of a demo.


Wouldn't that apply to the final technology as well?


I will add this thread to the list of "the strangely flagged down because bad buzz about Google".

You can say Gruber is wrong and argue about it. But flagging??? This is totally bullshit.


If it was live people would just be assuming that the person was reading a script. I don’t see how they could demo this without sceptics pointing out the flaws.


Why was this flagged?


Even if Google cured cancer, Gruber would still find a way to complain. According to him, tech is only flawless when it comes from Apple.




Registration is open for Startup School 2019. Classes start July 22nd.

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: