Hacker News Reader

Hacker News

Comment by _def | Hacker News Reader

Full thread

_def

If I find out someone pipes my chat messages into a LLM, I will not converse with that person anymore.

View on HN

Since Meta is a big AI investor, I suggest you skip WhatsApp altogether.

You think they (plan to) decrypt messages and then upload them again in plain text to a server?

Since on-device processing is neither as objectionable nor could be very large

I don't use WhatsApp myself because of who runs it and there are plenty of better options out there, so I certainly agree with the sentiment of steering clear, but this claim does seem pretty far out there

They don't plan it because they have no use for it. They only care about the metadata. When you talked to this person; your wife; at what time of day; was it at night; how long is the message; was there a product mentioned in the message; was the message about sports; etc.

They don't plan it, because so far, they don't have the keys to do so.

We do need to trust Meta that they really don't, to some extent, but people way smarter than me have researched the WA implementation of the Signal protocol and it seems solid. I.E: Meta appears to simply be unable to read what you chat and send. (but TBC: they do see with whom and when you do this, just not the contents).

They don't have the keys, but they probably can get them.

That's a strong accusation.

The only way I can think of, is by pushing an update that grabs all your keys and pushes them to their servers.

Otherwise, it's pretty decent set up (if I am to believe Moxie, which I do)

What prevents them from simply pushing an update that quietly uploads private keys or unencrypted messages to their servers

Presumably they use proper HTTPS, so all the data is essentially encrypted twice, if they just concatenate some packets with keys, it would be extremely difficult to detect as you'd need to decrypt HTTPS (which is possible if you can install your own certificates on a device), then dig through random message data to find a random value you don't even know.

> What prevents them from simply pushing an update that quietly uploads private keys or unencrypted messages to their servers

Reputation

Or what's the translation of bank run but generic for any service? Leegloop in Dutch. Translator gives only nonsense. Going for the descriptive route: many people would leave because of the tarnished reputation

The trick is to have Facebook continue to believe that this reputation/trust is more valuable than reading the messages of those who stay behind, which can partially be done by having realistic alternatives for people to switch to so that there is no reason to stay when trust is broken. Which kinda means pre-emptively switching (at least to build up a decent network effect elsewhere), which is what I've chosen to and encourage anyone to also do. But I'm not a conspiracy theorist who thinks that, at the present time, they'll try to roll out such an update in secret, at least not to everyone at once (intelligence agencies might send NSLs with specific targets)

At least on Android it's possible to dissect an app. You won't get the original java code, but static analysis is possible. And indeed, it's possible to capture it's network traffic and even often decrypt that traffic (with root access to the device). Now, I, or you may not research at this level, but someone looking into wether they may use WhatsApp to discuss attack plans on, say, Jemen, might find such weaknesses.

People find exploits in proprietary code, or even SaaS (where researchers cannot even access the software) every day.

People at Meta might leak this information too.

"Information wants to be free"

My point is: the risk of this becoming known is real.

That's a completely reasonable boundary. Privacy and consent are critical, especially when sharing personal messages or conversations. It's fair to expect that your interactions remain private unless you've explicitly agreed otherwise. If you'd like, you can communicate your stance clearly to others in advance, ensuring they're aware of your boundaries regarding the use of your messages with AI tools or other external resources.

I don't think their main concern was the privacy aspect.

What do you think their concern was? I can't see any other issues someone might have.

Energy usage is another. What would happen to world power consumption if 1% of WhatsApp chats would be fed to ChatGPT?

A third reason besides privacy would be the purpose. Is the purpose generating automatic replies? Or automatic summaries because the recipient can't be bothered to read what I wrote? That would be a dick move and a good reason to object as well, in my opinion

> What would happen to world power consumption if 1% of WhatsApp chats would be fed to ChatGPT?

The same thing that happens now, when 100% of power consumption is fed to other purposes. What's the problem with that?

Huh? It's additional power draw in the midst of an energy transition. It's not currently being used differently. What do you mean what's the problem with that?

Also don't forget it's just one of three aspects I can think of off the top of my head. This isn't the only issue with LLMs...

Edit: while typing this reply, I remembered a fourth: I've seen many people object morally/ethically to the training method in terms of taking other people's work for free and replicating it. I don't know how I stand on that one myself yet (it is awfully similar to a human learning and replicating creatively, but clearly on an inhuman scale, so idk) but that's yet another possible reason not to want this

If people need additional power, they pay for it. If they want to pay for extra power, why would we gatekeep whether their need is legitimate or not?

Because of the aforementioned shortage. Paying for more power means coal and gas gets spun up since there aren't enough renewables, and the externalities aren't being paid for by those people

I'm also happy to have them pay for the full cleanup cost rather than discourage useless consumption, but somehow people don't seem to think crazy energy prices are a great idea either

Also you're still very focused on this one specific issue rather than looking at the bigger picture. Not sure if the conversation is going anywhere like this

What's the bigger picture? You said "power usage", "to what purpose?" (you kind of don't get a say in whether I use an LLM to reply to you, though you're free to stop talking to me), and "objections to the training method", which doesn't really seem relevant to the use case, but more of a general objection to LLMs.

I understand why one would think it's funny to feed the parent comment into an LLM but please at least label when you echo such output on the site

Where do you draw the line? LLMs for searching, BM25 for searching, exact match only, no processes at all (forbid whatsapp search, make them scroll)?

Funny that people freakout about a local LLM while using Facebook products. They're probably the same types who use it to do their work.

> They're probably the same types who use it to do their work.

Citation needed.

It's a local LLM with access to an extraordinary amount of personal data. In the EU at least that personal data is supposed to be handled with care. I don't see people freaking out, but simple pointing out the leap of handing it over to ANOTHER company.

Not all Meta products are alike. WA has E2E encryption, has had it for a long time. It's the same protocol as Signal: in fact, it was built for/in WA by Moxie/signal a while ago.

That doesn't make the metadata private. Meta can use that as they want. But not the contents, nor the images, not even in group chats (as opposed to Telegram, where group-chats aren't (weren't?) E2E encrypted).

What you say or send on WA is private. Meta cannot see that. Nor governments nor your ISP or your router. Only you and the person or people you sent it to can read that.

It's a d*ck move if they then publicize this. And, as others pointed out, illegal even in many jurisdictions: AFAIK, it is in my country.

Do you think it'd be okay if they used a local LLM, via ollama, and this MCP server?

Personally, I would say that still reeks of being manipulative. I've received messages from a friend which were definitely LLM-generated, it made me like that person considerably less

If they use the LLM to search ("when did X tell me about that party somewhere around Y's neighborhood") then I don't think there's any problem.

If they configure it to indicate a prefix, for instance when answering questions like "when are you free to hang out" and it responding "[AI] according to X's calendar and work schedule, they may be available on the following days" I might also consider that somewhat useful (I just wouldn't take it as something they actually said).

If they're using LLMs to reword themselves or because they're not really interested in conversing, that's a definite ick.

I would personally use such a system in a receive-only mode for adding things to calendars or searching. But I'd also stick to local LLMs, so the context window would probably be too small to get much out of it anyway.

This is actually something I am curious about, if for example I use this and I and streaming all my contacts information and messages externally, surely I'm breaking privacy laws in some US states and certainly in the EU.

This seems sketchy to me.

It very much depends on the specifics around use cases, parties, and jurisdictions. In plenty of them, you're allowed to record and keep track of conversations you're taking part in, as is the other party, but publishing those on the internet would he illegal.

Processing them (like compressing them to mp3 files or storing them in cloud storage) is probably legal in most cases.

The potential problem with LLMs is that they use your input to train themselves.

As of right now, the legal status of AI is very much up in the air. It's looking like AI training will be exempt from things like copyright laws (because how else would you train an LLM without performing the biggest book piracy operation in history?), and if that happens things like personal rights may also fall to the AI training overlords.

I personally don't think using this is illegal. I'm pretty sure 100% of LinkedIn messages are being passed through AI already, as are all WhatsApp business accounts and any similar service. I suppose we'll have to wait for someone to get caught using these tools and the problem making it to a high enough court to form jurisprudence.

I hope you never contacted anyone with a business account then.

You should just assume that every single thing that you type into an electronic device made after the 90s gets piped into a LLM anyway

Zuck is already piping it into much worse

I mean, the technology is not the issue. Someone can read your past conversations today and take diligent notes to unearth the same insights an LLM might, if they were so inclined.

This might a actually be helpful for people with poor memory or neurodivergent minds, to help surface relative context to continue their conversation.

Or sales people to help with their customer relationship management.

Reply on news.ycombinator.com