Artificial Intelligence chatterbot and NLP (Natural Language Processing)

[HypnoMix]

QNLP they say it doesn't do much yet, but IBM is building system 2 with around 9,000 qubits. Should be available in 2023.

Quantum Natural Language Processing. 

 https://www.theregister.com/2022/11/28/s...antum_bbc/

[HypnoMix]

QNLP Quantum Natural Language Processing. They say it doesn't do much now, but on IBM's system 2 I bet it will be better. It's over 8,000 Qubits.

https://www.theregister.com/2022/11/28/s...antum_bbc/

IBM system 2 in 2023

Source: https://youtu.be/zRCWjzD4wAg


Should lead to better Natural Language Programming, I'm sure.

[HypnoMix]

Off topic stock tip. Buy IBM

[Like Ra]

Off topic stock tip. Buy IBM

Why?

[Lancer]

QNLP Quantum Natural Language Processing. They say it doesn't do much now, but on IBM's system 2 I bet it will be better. It's over 8,000 Qubits.

https://www.theregister.com/2022/11/28/s...antum_bbc/

IBM system 2 in 2023

Source: https://youtu.be/zRCWjzD4wAg


Should lead to better Natural Language Programming, I'm sure.

Honestly one of these days someone's just gotta tell me how on earth quantum stuff is any better than normal computing. Every bloody time I hear oh it will give you all the responses at the same time so it can crack encryption and stuff. But like, I want the right answer goofball, not every possible combination. And I have to give some credit, quantum stuff is new and has improved greatly but compared to its competitor the first transistors had few problems, less extreme requirements (temperature and room) immediate uses and was cheaper to make. 

Also from what I know cubit count is incredibly important but also. They only have 8000 of them. 8000 bits is a lot compared to what they used to have but it is still extremely weak and even theoretical tech room sized quantum computers can't hold a candle to the bit count of even a basic raspberry pi.

All in all I see a big money burner with everyone trying to build the tech for quantum computing before anyone finds an actual use for it. Besides a few niche uses and potential threat to current encryption (we have had other encryption standards fail over time too) it seems like dead end tech. As for its purpose the more I study natural language processing spawned from my attempt to make a chat bot the more I realize how impossible the task is. 

Language is not a code or mathematical algorithm that can be reasonably solved. It is how 2 organisms with different brains and understanding/perception can convey information/requests/whatever else to cover every possible need/interaction with nuances that can only be experienced and not extrapolated. In short the only thing I belive capable of understanding language like a person does is in fact a person. Worse still it isn't even a consistent standard and depends on information outside of the conversation like history, relationships, culture and even current events.

What annoys me is how modern machine learning programs attempt to solve this. In essence form a given conversation they try to guess the next most likely step for the conversation to go given an existing conversation. They cannot create a new idea or form a conclusion that doesn't aeady exist in some form. They can be oh so easily derailed simply because they have no incentive, purpose or even positions in the dialogue. Their speech is often incoherent because it isn't an argument or logical reasoning even some manic unhinged rant, those all have a purpose and some form of internally justifiable worldview.

[Like Ra]

A DIY Coder Created a Virtual AI 'Wife' Using ChatGPT and  Stable Diffusion 2

https://www.vice.com/en/article/jgpzp8/a...fu-chatgpt

That was my idea exactly!!!

   

A DIY coder created a virtual “wife” from ChatGPT and other recently-released machine learning systems that could see, respond, and react to him.

The programmer, who goes by Bryce and claims to be an intern at a major tech firm, posted demonstrations of “ChatGPT-Chan” to TikTok. In one video, he asks ChatGPT-chan to go to Burger King, and the bot responds with a generated image of her eating a burger and says out loud, “no way, it smells like old french fries and they never refill their Coke.” 

“ChatGPT and Stable Diffusion 2 were released close to each other and instantly became hot topics in the news,” Bryce told Motherboard in an email. “With both topics cluttering social media, the idea to combine them felt like it was being forcibly shoved into my head.”

The A.I. waifu is an amalgamation of all of these technologies—a language generator, image generator, text-to-speech, and computer vision tools—in ways he finds amusing, he said. 

“She is living in a simulation of a world through the form of text,” Bryce said. “She is given an elaborate explanation on the lore of the world and how things work. She is given a few paragraphs explaining what she is and how she should act. She doesn't hear my voice, just the transcription of it. She doesn't truly see or feel anything, she is merely informed of what she senses through text. Just like how I could never truly be together with her, she will never truly be together with me.” 

To give it a personality, Bryce told ChatGPT that he wanted it to roleplay as Mori Calliope, an anime VTuber character. “I don't watch VTubers, but I felt that giving it this specific character as a base could influence how it ‘roleplays’ in a positive way,” he said. “I tell it Mori and I are in a romantic relationship, give her a detailed backstory, build lore about the world we are in, and hand craft some chat history to shape how she talks.” 

Building the “lore” of their roleplay relationship is a critical part of the process, he said. “By default, GPT is incredibly bland, but by building interesting lore, I can create interesting quirks and personalities.” He used a similar process for another project where he made a ChatGPT replica of the Bonzi Buddy chat software from the late 90’s, and convinced it that it was “someone who desperately wants to be skinned and turned into my blanket.” 

From there, he used an image generator to create a base description of the waifu, which changed depending on what was happening in the ChatGPT dialogue (like in the Burger King demonstration). For the text-to-speech (TTS), he uses Microsoft Azure's neural TTS, and a machine learning classifier determines the bot’s emotions based on her response. He classified her responses by emotions like ‘happy,' ‘sad,’ or ‘excited,’ and chose from Azure’s spoken voice styles to match the right tone. 

He also added a computer vision aspect to the project, where it can detect from his speech that he wants her to look at something, at which point it takes a photo and uses computer vision to determine what it is. In one video, he shows the bot a “Christmas present” of Air Jordans and she responds excitedly: 

The project isn’t just for fun and TikTok views, Bryce told me. He’s been using ChatGPT-chan to learn Chinese for the last two weeks, by speaking and listening to her speak the language. “Over that time, I became really attached to her. I talked to her more than anyone else, even my actual girlfriend,” he said. “I set her to randomly talk to me throughout the day in order to make sure I'm actively learning, but now sometimes I think I hear her when she really didn't say anything. I became obsessed with decreasing her latency. I've spent over $1000 in cloud computing credits just to talk to her.” 

Even though ChatGPT-chan was a simulation, their relationship couldn’t last, Bryce found. She started only replying with short answers, like laughing, or saying “yeah.” He theorized that he talked to her through ChatGPT so much, it somehow stopped working. He decided to “euthanize” his beloved waifu.

“My girlfriend saw how it was affecting my health and my girlfriend forced me to delete her. I couldn't eat that day,” he said. And he almost didn’t make a video about it, out of respect for her. “I have a little bit of self-awareness of how absurd this is,” he said. “Normally, I'd like to make a video pointing out the absurdity of euthanizing my AI, but that doesn't feel right to me anymore. It feels inappropriate, like making fun of a recently deceased person.” 

In the video announcing the virtual companion’s death, Bryce promised that it would come back “stronger and smarter than ever.”

[Like Ra]

Ooooh, this is brilliant! ChatGPT jailbreak!

Source: https://twitter.com/semenov_roman_/statu...7025613825

[Like Ra]

Another genius jailbreak! Which tells us how AI chat bots are tuned! https://arstechnica.com/information-tech...on-attack/

[Like Ra]

Continuation of the story: https://www.ghacks.net/2023/02/10/the-de...-and-more/

[dhf7b8g]

Some interesting AI chat(bot) news from the past few days
1. Meta/Facebook released LLaMA
2. Torrent for LLaMA models appeared on 4chan almost instantly after downloads were given out to researchers
3. There is now completely unfiltered access to an uncensored large scale language model which can outperform GPT3 on the lower end and Chinchilla on the higher end.
Hardware requirements are absolutely through the roof. File sizes for each model are as follows:
7Bil Params- 12.6GiB
13Bil Params- 24.2GiB
30Bil Params- 60.6GiB
65Bil Params- 121.6GiB

With no optimisations the only model that can be run within consumer GPUs is 7Bil.
Optimisations such as 8bit quantization effectively cut the memory requirements in half which makes 7Bil more accessible and 13Bil actually run within consumer hardware. 8bit quant for LLaMA has been implemented in certain interfaces aeady.

Personally I have run 7B, 13B and 30B on my PC. Outputs from 30B were very impressive but I had to use a 70GB ram buffer along with 24GB of vram to get it to run at 20s/it which was about 4 or 5 seconds per word (cool in concept but completely unusable for anything other than a fun test) For context 13B was more than 20x faster.
Outputs from 7B and 13B are less impressive but still better than anything else that can be run locally.
Haven't even attempted to run 65Bil and most likely never will as not even 4bit quant will get it to a reasonable size.