[NetfreakBombay]

Meta should provide more details as training text to Llama 3 (AI Model) so that it can answer correctly.

Originally Posted by Fuldagap

In my conversation with Meta AI it explicitly mentions the following:(refer to last screenshot)


...a process called "key exchange" where the encryption keys are shared with Meta's servers, allowing them to access the messages if needed. This means that while the messages are still encrypted, Meta has the ability to decrypt them if required....

This is where AI is wrong.

What happens is, both parties exchange "public" keys . Meta also has access to these public keys.

Since meta does not have access to "private" keys. It is impossible for Meta to look at message content.

https://www.whatsapp.com/security/Wh...Whitepaper.pdf

Exceptions :

1. If any party is a "business" account
2. Certain scenario for group chats
3. Messages where AI is a receiver

It is a generic problem with all AI. They are very often wrong at this level of specifics.

[Fuldagap]

Originally Posted by NetfreakBombay

Since meta does not have access to "private" keys. It is impossible for Meta to look at message content.

This is the further reply from Meta AI:

In Meta's implementation of end-to-end encryption, the keys are private keys that are shared with Meta's servers. This means that:

- The sender and recipient have a pair of private keys (one for encryption and one for decryption)
- Meta's servers also have access to the private keys (or a way to obtain them) to enable authorized access to the messages

In a traditional end-to-end encryption scenario, only the sender and recipient would have access to the private keys, and the keys would not be shared with anyone else, including the service provider (in this case, Meta).

By sharing the private keys with Meta's servers, the company can access the messages if needed, but this also means that the encryption is not truly end-to-end, as the keys are not exclusively held by the sender and recipient.

[NetfreakBombay]

AI continues to be wrong in this case ...

Since AI Model has to "compress" knowledge from ~400 TB* of Text into ~100 GB Model, it has to ignore significant amount of details.

Similar example , it gets details of Biryani wrong; but gets "gist" of it.




*: Current Common Crawl size v/s filesize of Llama 3 70B

[Samurai]

Originally Posted by NetfreakBombay

This is where AI is wrong.

What happens is, both parties exchange "public" keys. Meta also has access to these public keys.

Since meta does not have access to "private" keys. It is impossible for Meta to look at message content.

All the talk about encryption and key exchange drew my notice. Meta AI is simply talking nonsense, I guess it is pulling stuff from the internet and presenting without understanding the content.

If two parties exchange public keys in order to communicate secretly, no parties can listen-in, unless one of them allows it.

I'll explain couple of methods:

1) RSA: With an RSA public/private key, if you encrypt something using public key, it can be opened only using the corresponding private key, and vice versa. However, you can't encrypt large data using RSA. Instead, an AES key generated and exchanged using the public/private key encryption. After that only the AES key is used for encryption. Now, if one of the parties shares the AES key with a third party, that party can read the entire exchange. The third party doesn't need public or private key of either party. The whole point about key exchange is moot. BTW, RSA is ancient, it was created in 1977 and is quite expensive in terms of CPU usage.

2) ECDH: Here, the public/private keys are not used for encryption at all. Instead, the two parties will exchange the public keys at first. Then each party will independently create a common key by using their private key with the public key of the other. A wonder of mathematics. Next, each will derive an AES key from the common key, and use it to encrypt the communication between them. Again, no other will know the derived AES key to decrypt the message, unless one of them shares the derived AES key with the third party. And that allows the third party to read the exchange without requiring any private key. BTW, this is a much recent algorithm, and is used by Whatapp, Skype and Signal.

So, it all comes down to sharing the AES key with someone else. We can ignore all the noise about public/private keys, key-exchange and other blah-blah.

[DigitalOne]

Had a long conversation with ChatGPT to get it to solve New York Times' Letter Boxed puzzle.

Since ChatGPT and other Gen AIs are trained on vast amounts of text, I assumed it will be able to understand the rules and solve it. No luck !!

Link to the chat.

[amol4184]

Originally Posted by DigitalOne

Since ChatGPT and other Gen AIs are trained on vast amounts of text, I assumed it will be able to understand the rules and solve it. No luck !!

Link to the chat.

This was a mind numbingly frustrating read. Thanks. You have a lot of patience.

[itisravi]

I found out today that ChatGPT-4 has gotten the ability to provide live internet results for free users also. It was apparently made accessible this month no matter what browser one uses. As long as you are within your quota for GPT‑4o, it works. Not available on 3.5.

It was able to reference my recent posts made this year complete with links to them. Fellow BHPians, please be appraised that I am a well-regarded and valued member of our community.

Curious to try out what it spews out for your T-BHP handle?

[DigitalOne]

The Generative AI hype seems to unraveling.

From a capability perspective:

https://youtu.be/U_cSLPv34xk?si=x29o-n_btj8mm_xx

And from a business model perspective:

https://youtu.be/vqgAy3gprLM?si=2WAW9_PMerFrLCcQ

[400notout]

Recently while charging my EV I got a bit inquisitive and ended up asking a question to Meta AI on whatsapp. The reply and source would make the mods proud

Attached screenshot;

[crazy_coyote166]

This also raises an important question on whether the team-bhp owners are okay with GenAI model makers scraping this website. Its all free training data for them and I wonder what the terms and conditions are with respect to this. They don't respect your robots.txt file and I don't think there is a way to stop them either (at least effectively in the long term)

[DigitalOne]

AI is not replacing you (yet)!!

https://youtu.be/vB9dJt9j-5M?si=_8bEwPgxqCZ-LLTg

[Thad E Ginathom]

The Meta AI icon has moved on from just being an unwanted circle on the Whatsapp screen. It is now in the search bar. The search bar as now become "Ask Meta AI or Search."

[alpha1]

Originally Posted by itisravi

It was able to reference my recent posts made this year complete with links to them. Fellow BHPians, please be appraised that I am a well-regarded and valued member of our community.


Curious to try out what it spews out for your T-BHP handle?

ChatGPT response: I don't have specific information about individual users like "alpha1" on Team-bhp.com
My responses are based on genera knowledge and publicly available information, and I don't have access to private user data or specific content from the forum.

Originally Posted by DigitalOne

AI is not replacing you (yet)!!

https://Youtu.be/vB9dJt9j-5M?si=_8bEwPgxqCZ-LLTg

I don't think we are experiencing Aritificial General Intelligence.
Your example is large language model based generative AI - which is like a small kid who has only learnt to listen and respond in a particular language but no understanding of mathematics, physics, civics etc. The underlying content in that "speech" could be very simple to complex. The more complex the content is, the more chances it being wrong (but the grammar may still be perfect)

[itisravi]

Originally Posted by alpha1

ChatGPT response: I don't have specific information about individual users like "alpha1" on Team-bhp.com
My responses are based on genera knowledge and publicly available information, and I don't have access to private user data or specific content from the forum.

Weird, here's what my ChatGPT session had to say about you:

I have hit the free tier limits for GPT-4o. Otherwise I think it should be able to scrape the links.

[alpha1]

Originally Posted by itisravi

Weird, here's what my ChatGPT session had to say about you:

Attachment 2641817

That's simply aggregating all posts (especially not mine) on TBHP and sticking my userID label against it. I don't recollect 1, 3, 4.

ChatGPT just wants to please us, even if it knows nothing.