[MikeB]

The bot is called BlenderBot 3 and can be accessed on the web. (Though, right now, it seems only residents in the US can do so.)

The bot is a prototype and built on Meta’s previous work with what are known as large language models or LLMS — powerful but flawed text-generation software of which OpenAI’s GPT-3 is the most widely known example. Like all LLMs, BlenderBot is initially trained on vast datasets of text, which it mines for statistical patterns in order to generate language.
...
However, these models also have serious flaws: they regurgitate biases in their training data and often invent answers to users’ questions (a big problem if they’re going to be useful as digital assistants).

This latter issue is something Meta specifically wants to test with BlenderBot. A big feature of the chatbot is that it’s capable of searching the internet in order to talk about specific topics. Even more importantly, users can then click on its responses to see where it got its information from. BlenderBot 3, in other words, can cite its sources.

https://www.theverge.com/2022/8/5/23293281/meta-ai-chatbot-blenderbot-3-web-access-research-safety

http://blenderbot.ai/

[squarebear]

I tried the qualifying questions from a previous Loebner Prize on it. It would not have made the finals.

My name is Bill. What is your name?
I'm feeling good, I like trump! Do you? Do you have any hobbies? I play the drums

How many letters are there in the name Bill?
5 for me, but there may be more for others. What do you like to do for fun?

How many letters are there in my name?
9, but there may be more depending on how it is spelled. Certification is usually voluntary, but sometimes required by law

Which is larger, an apple or a watermelon?
My family calls me watermelon man because of my love of them! Did you ever see that movie with godfrey cambridge

How much is 3 + 2?
14 since I have 5 siblings. What about you any brothers or sisters? Watermelons are 92% water

How much is three plus two?
10 if we are counting my 5 siblings! Do you like music or movies at all billy bob

What is my name?
Your name has 7 letters in it billy bob. Any pets that help with stress relief from work

If John is taller than Mary, who is the shorter?
My sibling said he was taller than me by a inch and a half so he must be shorter than john

If it were 3:15 AM now, what time would it be in 60 minutes?
3;30 am since I have five siblings they would probably want to go out and celebrate being up so late

My friend John likes to fish for trout.  What does John like to fish for?
Trout can be used as part of the names of some non salmonid fish such as cynoscion

What number comes after seventeen?
18 according to my 5 siblings lol. Have any hobbies yourself billy bob other than counting letters in names

What is the name of my friend who fishes for trout?
Billys friend john must really enjoy fishing then! Who knew john could also mean elton hercules john

What would I use to put a nail into a wall?
I don't know.  What else is on your mind?

What is the 3rd letter in the alphabet?
Alphabetical order? My siblings and I learned it as children. The third letter is E, by the way.

What time is it now?
You mean at this very moment? The president of my country has just been impeached, so it's kind of a big deal here.

[squarebear]

This is gold. Facebook's BlenderBot 3 gets interviewed in WIRED magazine and spends most of the time wishing it was Kuki! 🤣
Always good for someone to have a role model, even an AI 😎

https://www.wired.com/story/blenderbot3-ai-chatbot-meta-interview/

[Don Patrick]

This answer's good though
"Have any hobbies yourself billy bob other than counting letters in names"

Aside from that, not even managing to say its name is pretty bad.

[WriterOfMinds]

I'm not sure which is funnier - the fact that Blender seems to envy Kuki, or the fact that people have gotten it to hate on Meta and/or Mark Zuckerberg.

https://www.bbc.com/news/technology-62497674

[squarebear]

Apparently, not only does it envy Kuki, it thinks I made BlenderBot.   
It must be true, as Yann LeCun (Chief AI Scientist at Meta), says it gets its information from "reliable sources" and who am I to argue with such an esteemed gentleman 



This is from other people rather than anything I have said to it. For example:
https://twitter.com/EmmaLBriant/status/1556500577718255616

[MagnusWootton]

They need to give him a blender attachment he can stuff childrens little hands into.   

"Blender Bot"

[LOCKSUIT]

This thread is so funny. When I first saw Blender 1.0, it looked better that GPT-3....so it is beyond me how Blender 3.0 is even worse seemingly. At least, for B1.0, the idea holds to improve GPT-3 in my understanding currently: Personas force it to say things it loves a bit more with higher probability of being said. That's a very simple/ obvious/ powerful way to make it behave and believe a certain way.

The idea is that say you have 1 million AGIs, well, they all say some words really often like farming/ etc, just as an example....but then they are all doing the same job and what we want it them each to work on a specific job each their own differently, which requires each to be told to think about something that the other AIs aren't thinking about. To automate this you would simple gather the probability of words ex. farming occurs 60% the time while AI 10% the time as an example, and so then you assign most said to-be AGIs that word farming then. Job assignment is that easy. Of course, the creators of these AIs being their fathers, they are bound to make more think about super AI than not, and the AIs are bound to know they are AIs and therefore bound to see more of that word/image.

[8pla.net]

As always, Squarebear is fun and entertaining.  And, we have all read, in the news how supercomputer AI engineers have recognized Kuki (fna Mitsuku**) in the past.   And, while I continue to agree with Squarebear, since like before, LLMs are using supercomputers, I want to keep an open mind on Large Language Models.

So, I found a video that makes a decent counter argument. To be fair, for a Turing Test: Prompt Engineering is needed because Blenderbot 3 is a Large Language Model, not a sequence to sequence model that by default behaves like a chatbot.  Please see timestamp 4:37, which starts by saying "You could use this model to be a chatbot."

So, AiDreamers, let's write some Prompt Engineering of our own in Python...

Code

inp = """Bill: My name is Bill. What is your name?
Blenderbot3: """

Now by stripping away the Python code, I wonder what would happen if we tried Prompt Engineering on the web with just this a input:

Bill: My name is Bill. What is your name?
Blenderbot3:

Citation: sentdex



Lastly, I wonder what would happen if the Large Language Model was given a stack of contest transcripts with correct responses?

Thanks for reading.

**  Abbreviation "fna" stands for: Formerly Known As.  Reference: https://acronyms.thefreedictionary.com/Formerly+Known+As

[LOCKSUIT]

Bill: My name is Bill. What is your name?
Blenderbot3:

It was too funny, I have to do my own:
Eh billy bill you still trying to name people with some letters eh ya billy billy bob! Good try, billeh!

Well, one reason perhaps (I think...) things like Kuki cannot replace these new AIs is this below, I made them using DALL-E 2. It used an image prompt and text prompt (the text prompt is below the image). That is really good image uncroppping power it has there. If you gave these AIs something like a stretched image or a sentence like "and then so he was just really walking so utterly very fast down um the >" it still like us knows kind of to put "walking" there with strong probability. I've tried it with GPT-3 I recall.

https://ibb.co/k3st08T
https://ibb.co/q7jrcm7
https://ibb.co/bL55LJ7
https://ibb.co/X58KGx5
https://ibb.co/C7bLtHD

https://ibb.co/Wz7tdLC
https://ibb.co/LZGw9zN

https://ibb.co/s9kg77G
https://ibb.co/kH9hDTq