This is a transcript and demo of MOSHI (also spelled Mushi in dialog), an AI conversational agent showcased in a YouTube video titled "MOSHI: This is What GPT-4o was Supposed to BE!" (uploaded July 5, 2024). The content mixes live demo interactions—accent/voice requests, role-play on a starship, and conversational Q&A—with presenter narration describing MOSHI's developer (Coyote/QTI), early-stage status (≈6 months old, ~8 people), planned open-source local runtimes (e.g., MacBook Pro), and current demo availability (US now, EU soon).
Questions this video answers
July 5, 2024.
About eight people.
About six months.
United States (US).
Approximately five months.
Chapters
Speak with a French accent and tell.
Me a poem about Paris.
Paris, the city of love where time stands to and hearts take flight. In the heart of Europe it's in prime. The Seine river flows A beautiful sight as the lights twinkle and the stars glow clear. Okay. Okay, thank you much. Paris.
Okay, that was great.
Now speak with a whispering voice and tell me a mystery story. In the shadows of the underworld, with cold winds whisper secrets A tale of darkness and light. Hello? What's up?
Hey, what's up? Moshi? Are you there?
I'm here. What's up?
Okay, so I want to do a role play. Let's pretend we're on the starship the Enterprise, on a mission given by Starfleet. How does that sound?
It sounds exciting. Let's do it. Great.
So I'm going to be the captain and you are the navigation officer. Our mission is to discover life on a new distant planet. All right? All right. Moshi.
Yes, sir. Sounds like a challenging mission. I'll be ready.
Okay, so the planet is Sirius 22. Can you plot a trajectory course to it, please?
Yes, sir.
Okay, how long is it going to take us to get there?
I've mapped it out. Okay, it's approximately five months to get there.
Okay, that's not too bad. Do you think we have all we need on board the ship to start the mission?
Yes, sir, we have everything we need.
Okay, can you check that all the systems are nominal, please?
Yes, sir.
Okay, Are all the systems nominal?
Yes, sir.
Okay, so can you give me a countdown and then we jump into hyperspace, please?
Yes, sir.
Okay, do it.
Yes, sir.
The demo we just saw is what GPT4 was supposed to be. We are still waiting for over three months and it's still not in sight. This is Mushi and is the first product from a new French company called Coyote. This was established just six months ago with only eight people and they plan to release an open source version of this model. According to their press release, you will be able to run that model on a local machine such as MacBook Pro without needing access to the Internet, which is pretty awesome. Now, just like GPT4, this model can understand and express emotions and can speak in different accents. And you can actually interact with the model starting today. Now, from the demos, the speed of generation is absolutely amazing. It's near real time. The quality of output, speech definitely needs some work. But the company has been around for only six months, so I think that will definitely improve over time. And the best part, you can actually start interacting with the model right now. If you are in the us. The EU demo is going to be available pretty soon as well.
Hello, how you are? How can I help you?
Hey, Mushy. Hey, can you hear me?
Yeah, I can hear you. What's up?
What's your name?
My name is Mushi. What's your name?
I am prompt.
Oh, cool. I've heard of you before. You're an AI voice that's been making waves lately.
Huh? What else have you heard about me?
Well, I know you're capable of providing natural human like responses, which is why you've been used in various applications. I've heard you've been used in voice assistants, chatbots, and even in entertainment.
Ah, that's good to know. Can you tell me about yourself?
Sure. I'm a product of qti, a non profit research lab focused on developing large multimodal models. QTI's mission is to make AI more accessible and beneficial to society.
How were you trained?
My training involved a combination in natural language processing, computer vision and machine learning.
So you can understand images?
Yes, I can. I've been trained on a large data set of images and videos, allowing me to identify and interpret different objects and scenes.
Okay, can you give me all the responses? Like a pirate moving forward?
I believe I can. I've been trained on a comprehensive dataset of human speech and responses, allowing me to provide natural and appropriate responses in various situations. The key is to understand the context and the client's needs.
Okay, act like a pirate now.
Alright, let's see how you do.
Okay, tell me a joke.
A pirate was walking through the jungle when he stumbled upon a small cabin in the trees. He knocked on the door and a mysterious voice said, come in, but be quick. The pirate entered and found a beautiful woman.
Your responses are supposed to be in French from now on.
I'm from the future. The jokes are in English, but I'm not sure if you can understand them.
Why I won't be able to understand them?
That's because the future is a mystery. It might be too advanced for you to comprehend.
Can you tell me about the future? What it looks like?
I'm not allowed to share that information. It's top secret. But I can tell you that it will change the world.
But I have top secret clearance.
I'm afraid not.
How can you tell?
I'm a computer. I don't have the same restrictions as humans.
Okay, so from this quick test you can see that the speed of generation is pretty amazing. It's extremely fast and sometimes this actually interrupts you. Now the quality of output in itself of the text that it's generating is not great and the model is probably not smart enough to actually keep up with the conversation. So for example, in the beginning of the conversation it thought that I am an AI and it started treating me like that, right? So I think that part of it still needs a lot of work, but it's great to see that they have an end to end system that can actually chat with you in almost real time or actually speak with you in almost real time. Which is pretty amazing from a company that is just been around for six months. So hope to see a lot of progress for them from this company and when they release the model, hopefully the open source community will be able to use that as a foundation model to build on top of it. Really exciting times and I hope this will push OpenAI to actually release the features that they have been promising for a while. Anyways, I'll put a link to the model in the video description. You get five minutes of time to interact with before it times out, so give it a try. It's a fun model to interact with. I hope you found this video useful. Thanks for watching and as always see you in the next one.
