[Venkatesh]

Gemini on the Go: App Now Available in India.

Available in English and 9 Indian languages. We’re also adding these local languages to Gemini Advanced, plus other new features, and launching Gemini in Google Messages in English.

Link

[SmartCat]

Meta has released its AI chatbot on whatsapp. To access it, open Whatsapp app and tap the blue ring (above calls, above the green + icon):





After your first interaction with Meta AI using the app, it then shows up under 'chats' on Whatsapp Web/Desktop too. Not impressed by its reply to my first question!



Also, within a whatsapp group or while chatting with an individual, you can refer to Meta AI by inputting '@ Meta AI' and asking the question. Everybody in the whatsapp group can read the question/asnwer. Apparently, Meta AI is based on GPT 4.0, but is actually based on Llama language model:

[DigitalOne]

On a whim, after I solved today's (somewhat difficult) Wordle, I asked three AI systems to solve it.

Meta's WhatsApp integrated AI offered to help find the answer, Google Gemini gave a huge clue, and ChatGPT spat out the answer outright .

ps - ChatGPT's training data was once restricted till 2022. Is the training data now being refreshed realtime?

[SmartCat]

Originally Posted by DigitalOne

ps - ChatGPT's training data was once restricted till 2022. Is the training data now being refreshed realtime?

For a limited number of queries per day, you get access to GPT 4o model.



After that limit is crossed, you go back to GPT 3.5 with 'year' restricted training data.

[NetfreakBombay]

Originally Posted by DigitalOne

ChatGPT's training data was once restricted till 2022. Is the training data now being refreshed realtime?

Yes and no ... while training data is not refreshed in realtime and model remains as it is; overall system gets updated multiple times a week.

Any realistic system will have at least four things :

1. Model (ChatGPT , Llama 3 , Gemini etc.) . Usually updated once a year.
2. More realtime content that model can(optionally) access
3. A pipeline that enables model to look up newer info in realtime
4. Training data that will enable model to figure out when to use newer content or realtime search.

When I write code my my systems, similar design is used. Pretty much a pattern in industry.

Example : I asked about debate that happened on 27th.

[Rocketscience]

The fact that it is dodging this question and not directly answering this is curious, later I asked if I can disable this and it basically said not possible. So this is not an additional add on but forced down to everyone.
While otherwise in my limited interaction, it seems decently capable and it could make AI reach masses in a way Chat GPT couldn't, I mean with the older generation. But then, this was also the premise of UPI in whatsapp but it never took off.

[PearlJam]

Originally Posted by Rocketscience

The fact that it is dodging this question and not directly answering this is curious

I don't see what you're seeing. I see:

[Rocketscience]

Originally Posted by PearlJam

I don't see what you're seeing. I see:

Attachment 2622476

It is variable I guess, it did this thing when I told it to generate an image, it did but when I asked for a change, it said it couldn't do it and didn't have the ability to, but then after a while I asked it again and it complied.

Also noticed a curious thing that whatsapp has really low resolution for display picture and I have always found it annoying, it is much less than even acceptable quality for me but it has not changed ever, but this meta AI has is an exception and its display picture is properly high resolution. They haven't even made this exception for their own Instagram's chats in whatsapp but this gets it.

[Fuldagap]

Originally Posted by Rocketscience

The fact that it is dodging this question and not directly answering this is curious, later I asked if I can disable this and it basically said not possible. So this is not an additional add on but forced down to everyone.
While otherwise in my limited interaction, it seems decently capable and it could make AI reach masses in a way Chat GPT couldn't, I mean with the older generation. But then, this was also the premise of UPI in whatsapp but it never took off.

I had a very long chat with Meta AI on this. It finally admitted to the following:
Although the chats are encrypted, people inside Meta can read any messsges they want (for training and upgradation of services plus the quality of content being provided)
It also admitted that chats can be read by Meta AI.

Meta AI is quite dodgy to questions and falsifies information provided quite often.
It repeats mistakes after apologising over and over again.

I find Claude/Gemini/Copilot/ChatGPT to be upfront and more transparent.
Meta AI is very convenient to use from inside Whatsapp but there is no way to disable it or remove it from whatsapp.

There is no search option inside the Meta AI chatbox and it is cumbersome to find older chats unlike chats with other people where the search is available.

I avoid Meta AI because the trust factor for responses is quite low. The level of info provided is also way inferior to what others offer. Meta sticks to its own sources often citing data from inferior websites.
Where it has low confidence on the data it simply hides the source of that info.

[SmartCat]

Originally Posted by Rocketscience

I told it to generate an image

OK thanks, didn't know Meta AI supported text to image functionality. Is that a Contessa!



It was 'hallucinating' that it supports text to video too, but then it figured out it does not right now:

[Fuldagap]

Originally Posted by SmartCat

OK thanks, didn't know Meta AI supported text to image functionality

Try criticising Meta AI and it kind of loses patience soon. Tells me not to be harsh on it. "No need to be harsh".

Chatgpt and others correct themselves quite quickly instead of indulging in irrelevant conversations. Meta AI just forgets the conversation at hand and starts focussing on telling that it will improve and not to be harsh, all the while not providing any better info.

[Fuldagap]

Here is what Meta AI tells about reading of chats by Meta AI or personnel. Responses are long but interesting.

[NetfreakBombay]

Originally Posted by Fuldagap

Here is what Meta AI tells about reading of chats by Meta AI or personnel. Responses are long but interesting.

Probably they need to add specific content about this for AI to refer. They mention in TnC for AI.

1. Whenver users are chatting with a bot (Meta AI in this case), messages are accessible to bot and hence to support personnel, Legal teams etc. In this case, messages are used for incremental updates as well.
2. 1 : 1 messages between individuals aer not readable by anyone else (Or not "supposed" to be readable by anyone else)

[Fuldagap]

Originally Posted by NetfreakBombay

Probably they need to add specific content about this for AI to refer. They mention in TnC for AI.

1. Whenver users are chatting with a bot (Meta AI in this case), messages are accessible to bot and hence to support personnel, Legal teams etc. In this case, messages are used for incremental updates as well.
2. 1 : 1 messages between individuals aer not readable by anyone else (Or not "supposed" to be readable by anyone else)

Attachment 2622906

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

I see what's causing the confusion. End-to-end encryption typically means that only the sender and recipient can read the messages. However, Meta's implementation of end-to-end encryption in their messaging services, like WhatsApp, has a caveat.

Meta uses a protocol called the "Signal Protocol" which is designed to be end-to-end encrypted. However, Meta has implemented a workaround to allow authorized personnel to access messages in specific circumstances, as I mentioned earlier.

This is achieved through 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.

It's important to note that this approach has raised concerns among privacy advocates, as it potentially weakens the end-to-end encryption guarantee. Some argue that this implementation may not be entirely consistent with the principles of end-to-end encryption.