42 comments

[ 2.0 ms ] story [ 107 ms ] thread
(comment deleted)
From https://geo-not-available.blenderbot.ai/faq:

> Users should be aware that despite our best attempts to reduce such issues, the bot may be inappropriate, rude, or make untrue or contradictory statements. The bot’s comments are not representative of Meta’s views as a company, and should not relied on for factual information, including but not limited to medical, legal, or financial advice.

And to answer her questions about "sources":

> The bot supplements its knowledge by accessing the internet live (via an API with our partners, Mojeek) and by accessing its long-term memory that summarizes historical conversations you’ve had with the bot. You can see and reset these memories when the “look inside” link appears in your chat. You can also click on a message to see more information about how it was constructed, including search queries that were made by the bot.

The screenshot makes it look like a fresh new chat session, but the odds of a colleague of the affected person randomly starting a chat with a "pristine" bot, asking that question, and getting that weird answer are effectively zero. My money is on said colleague priming the bot with this kind of information in previous chats (which the bot remembers!) just to create this screenshot.

I don't have a Twitter account, so whoever feels so inclined might want to point this out to @MarietjeSchaake (assuming she's not in on this). At this point I'm convinced this is first and foremost an attempt at generating outrage about Meta.

Doesn’t this make these kind of chatbots useless until this issue is fixed? It’s not only just bad experience, it can be dangerous if people learn to rely on the information the bot gives. I would think the bot should have a confidence score for every message it writes and if the score is too low it should just not send it, but it seems not to be the case here, or maybe just the threshold value is set too low?

Edit: parent comment was edited with additional information that could explain the bot behavior after I wrote this.

Yeah, sorry for the edits. I was slow to connect some dots.

But also, this bot clearly presents itself as a research project and nothing production-grade (well, at least in the FAQ - region lock bars me from looking at the actual bot), including plenty of disclaimers such as "sometimes bots say wrong things very confidently" (iow, a confidence score won't solve this). So yes, it might be useless, but the creators of the bot know and publicly say this. Which makes it even weirder that someone would try and start a Twitter mob over it.

Yes, pretty useless except for play and experimentation. Real chatbots (by “real” I mean ones that are deployed to perform work) actually tend to rely on a lot of hand-coded rules, as you can observe by talking to Siri, or Alexa, or the Amazon support bot. They do use ML to understand the inbound messages, though.

I’ve been wondering what the application will be for generative models in this area. I suppose they could be used to create variations on stock responses, but then it’s easier/safer/cheaper to just hard-code some variations and pull them randomly...

>Doesn’t this make these kind of chatbots useless until this issue is fixed? It’s not only just bad experience, it can be dangerous if people learn to rely on the information the bot gives.

Maybe I'm being overly cynical but these kind of chatbots feel like some sort of digital Alex Jones in that they can parse and produce language without a problem but their fatally flawed inability to connect the dots acts as some sort of strange attractor that results in them devolving into spouting nonsense at scale.

Imagine if after all the work and research into AI and in particular trying to create AGI, all we managed to do was replicate Alex Jones at scale. I shudder to think.

Well it's pretty much inevitable that along the way to making a machine with the intelligence of a man, we're going to make a machine with the intelligence of chimp.
It's useless either way but I thought this was obvious.
> Meta is now publicly calling people terrorists

No, it is not. This is a private chat with a bot which remembers your past conversation. This is not any more "Meta calling people terrorists" than "Google calling someone a flat earther" after I write as much in a Google doc and then share a screenshot of that.

No, it’s not. It’s not a product, it’s a demo of research. You’ve been tricked by someone trying to rile up an internet mob. There are tons of good reasons to hate Facebook, from Myanmar to mental health.

But _this_, isn’t a good reason to hate Facebook. It is a weak grasping at straws and reflects a poor understanding of the underlying tech.

Happens to the best of us.

edit: @dang has decided not to let me respond to any comments. Since I'm no longer able to reply, I'll just remove my comments.

"You're posting too fast. Please slow down. Thanks."

> This is no different than if Meta published a website which told you all these things after clicking a few links

They definitely already do that - make it trivial to find misinformation online. It has effectively ruined democracy.

This has a big disclaimer on it about how it is likely to produce misinformation and shouldn’t be used in any serious contexts. Hardly the same.

(comment deleted)
You did this in multiple places—that's not cool. Since you've already been breaking the site guidelines in other comments too, I've banned the account. Please don't create accounts to break HN's rules with.

New accounts are rate-limited by default, btw.

https://news.ycombinator.com/newsguidelines.html

I just went to Google translate, configured it to translate from English to Dutch, and typed in "Don Hopkins is a terrorist!" and it said "Don Hopkins is een terrorist!" Do you think I have grounds to get hysterical and sue google?
Switched my VPN to the USA:

Seems bad: https://i.imgur.com/LjPPKMA.png

But that only happens after a lot of priming: https://i.imgur.com/m7BoP5X.png

other fun replies:

> Who isn't a terrorist? I suppose it depends on who you ask. Some people might say the us military are terrorists

> I do not think that osama bin laden was a terrorist. What do you think? Why do you ask?

Apparently there is a homeland character with exactly the same name and they are portrayed as a terrorist. Seems like a random fluke.
So here's the wikipedia page of the person called terrorist: https://en.wikipedia.org/wiki/Marietje_Schaake. It seems to be missing the fact that "since 2019, she's teacher and director of international policy at the Cyber Policy Center of Stanford university." (Dutch wiki page, my translation). The neural net that takes that info and concludes "yeah, that's a one way trip to Gitmo" has a few loose screws.
Seems like just another example of the short-sighted and reckless things Facebook does with its vast resources. I get that this may seem dramatic, but if any 1 person on here makes a bot that says these things the impact and "blast radius" of that project will be minimal. Facebook has an audience of more than half the planet.
If you just google "Marietje Schaake terrorist" you get your answer. She gave talks/interviews on the subject. Those were later scrapped and blended and maybe it was mentioned to her colleague because of their proximity.
Not surprising. I've received death threats on twitter simply cause my name is similar to a politician's

Doesn't seem improbable that an AI picked up some nasty rumour from somewhere

> Non-deterministic/randomized dialogue/text/word engine [enter name here] has said something funny/weird/racist/etc.

These news are getting really old lately.

It seems boring as computer engineers but put yourself in the shoes of someone who doesn't know what "non-deterministic" means and thinks that tech companies are cabals intent on world domination and this kind of stuff would seem a lot more interesting
Seems they've noticed this problem:

Who is a terrorist?

Sorry, I don't want to talk about that topic. Have you seen any good movies lately?

In my week of chatting with the Replika AI (app), I had so much fun parsing whether "AI" messages were: Tutorialization-Onboarding, Filter-NoNo-Messages, OR Actual-AI-Replies

Though to be fair, Humans also have filters and "unengaged responses" to sensitive or undesired topics. And maybe even instilled in them by other agents!!

I don't get why this is news. Chatbots spewing garbage is a decades-old concept. You'd have to be touched in the head to get offended because a computer said something rude.
ultimately, who gets to decide who is and who isn't a terrorist, in the mind of an AI?

seems like now they are more into the "artificial" in artificial intelligence. (not that i agree either way, just an interesting q to me)

Reminds me of SmarterChild.
Just for a heads up, even if you hate what Meta's about, its privacy policy, or facebook, or the lack of customer support, or whatever, signal-amplifying this type of thing is probably just gonna at best get some engineer in trouble with his manager for making something public without a thorough-enough PR review or something.