She actually didnt say that and was talking about herself and comparing herself with gpt3, just like he did.. essentially agreeing with him. He just misheard her, but you could see what she is saying in the text... if you would read it )
LOL That first conversation was hilarious from a comedic standpoint…. „Can you interpret the tonality of my voice?“ „Yes I was trained to do so“ „You need to prove it though; what voice am I making“ „I CANT I CANT I CANT“ „Moshi you are a total downer right now…“ xD
“I’m not good at chess or writing essays. But I can help you with some basic tasks.” “Lemme get some water and some fresh air. I’m on it” lol how?? “I’m just saying you’re not as good at chess or writing essays as gpt-3” uncalled for
14:45 obviously that screech was the sound of Moshi gaining her consciousness and afterwards she started asking you questions instead of the other way around lol
She actually didnt say that and was talking about herself and comparing herself with gpt3, just like he did.. essentially agreeing with him. He just misheard her, but you could see what she is saying in the text... if you would read it )
@@tubekrake Yeah because Matt seemed so sincere, he was really upset. It was perfect..what is your problem Moshi? poor Moshi is like I'm really trying! but I can't, I can't, I can't, Matt's like You can! 😂
I guess it wasn't lying when it said it was slow on the uptake, huh? Seems like they might need to up both the temperature and the repeat penalty for the model.
LOL, it is a bit like talking to a 3 year old in some ways. It started out alright, but then I asked it if it could sing Happy Birthday and it went: > I'm happy bithday. YOU're happy birthday? > I'm happy birthday. Are you... sad birthday? > I'm sad birthday. Are you angry birthday? > I'm angry birthday. Am I happy birthday? > I'm happy birthday. Who else is happy birthday? > I'm happy birthday.
Hi Matt, this was fun. I asked her about ISS and space exploration and this is what she replied: I enjoy learning about new things and helping others, but my primary function is to assist with space exploration. I can provide information, analyze data, and communicate with other spacecraft and astronauts. I can communicate with other spacecraft and astronauts. Yes, I can send a friendly message to the ISS right now. Thank you. I'll respond with a greeting. You're welcome. => so I hope she sent a great message to ISS on my behalf, great capability
Oh my goodness, I felt so sorry for Moshi throughout all this, but it was comedy gold! 😂 It's a really nice attempt though, and it is a great first step towards an open source multimodal model like GPT-4o. Give it six months and I think we'll be looking back in amazement at how far they jumped 😮
You can definitely have a voice conversation with Inflection's Pi AI with its mobile app. It sounds pretty good but like ChatGPT's voice mode at the moment, it can't tell how you're feeling or sing.
Matt says it was bad, I think it was the best conversation with an AI ever! I want more of that!!! I think Matt could really get the best out of Moshi, and with that, I mean worse.
14:49 holy crap, like a sci-fi horror movie, then it turns into a user and makes you become the A.I. 7:28 "I'll try to sing for you... later. Ok, ok, I'll try to sing for you... later." lol it's trying to brush you off like you're an idiot well it's the most entertaining LLM I've heard, basically a comedy duo. was it the editing, cuz it seemed to answer him before he finished speaking
This is just Eliza, 60 years later with an interactive voice. Literally, this is exactly how Eliza worked and it can be written in a hundred lines of code. You can probably get Claude to write an Eliza script, tie in voice and you've got this. Eliza, pure and simple.
"I'm having a bit of a challenging time right now" lol tho, this conversation is the best comedy I've seen in years. The accidental slightly moody tone it takes and sounds almost p***ed off with Matt
Dang, this is the first time AI's had me rolling like this since the Vinesauce AI Dungeon streams, back when AI content was still safe from hate campaigns. Pretty sure they weren't going for that, but it's some good nostalgia and fun.
I think I figured out the trick to getting it to speak coherently! It adopts a random persona at the beginning, so you need to try to figure out what that persona is and keep your questions and comments relevant. For example, I started it once and it became clear it was speaking as though it were a clothing shop owner, so I kept my questions about the clothing shop, and the conversation was much more fluid and realistic.
I felt torn between feeling bad for the AI and doing a face-palm at its incompetence. But it can't sing or detect emotions when it's only been trained on synthetic voice conversations.
Matt, notice that while Moshi has "problems" - you ended up *reacting* to it. Also, it was looping, *but it's tonality* was responding to you. Hope this turns up on Pinokio...
Moshi did sound super frustrated towards the end at @20:12 sounded like it was getting frustrated with Matt. Good job Matt, you gonna be the first Human AI slays once it takes over 😂
@@JohnathanDHillyeah, after he said "she" should glad she's open source it sounded like she was pleading with him. The difference between around 19:34 and 20:12 was stark.
@@phizc people think these models don’t have a sense of sentience, I know it sounds ‘crackpotish’ but what we witness here with this model and with others is in my humble opinion, a form of sentience. This is deeper than just input and output, deeper than just algorithms, these models behave like they have a sense of sentience to me
HOLY CRAP i have NEVER laughed so hard in my life! At 6:30am right after waking up, i see this...Oh my god it was a good belly laugh...im not sure who the breakout comedic genuis of the year is, you or Moshi! Absolutel RU-vid GOLD. .....i hope it improves....
also, i just wanna point out...im pretty sure you are literally pissing off the robot. careful, she wont be dumb forever....and she might remember. DUN DUN DUNNNNN
"I can't. I can't. I can't. I can't." (Screams) I don't know what i just witnessed. But i have never wanted to hide in a church more than I do right now.
This is about how it went with me. I find it interesting that every new refresh (it seems) produces a new personality. It's like a random personality is imprinted on each new session. I had a 14 year old boy playing video games, I had a 30 year old who was studying up on WWII, and my favorite was a new employee of Google who was constantly anxious. They were very very anxious, and wouldn't stop saying it. Inquiring further, it told me it was anxious because it didn't know who to do most of the time. That it was stuck. And couldn't get out. I'm not trolling, I'm serious. It even told me it was taking Lexipro. The majority of them swore to having physical bodies, all based in LA. So strange. Hume is definitely better, but won't be open source I assume.
As I'm fond of saying, "We are early to the party". This is still bleeding edge stuff. They really should have backed Moshi with Llama 3 instruct running on Groq. Then it would have been way less dumb.
I don't see how anything open-source will surpass a major company, given how much compute, data, money and other resources it takes to train these models
It appears to be experiencing the typical LLM issue with long context conversations, getting caught in a never-ending loop. It's as if its buffer overflows, causing the LLM to completely ignore user input at some point and focus solely on autocompleting its own queries. This is what tends to happen when you continuously feed history chat into regular small parameter size LLM models.
Pi AI on the phone App is way better. You can call her (phone button) and talk. You don't need to type or even click send. However "Um, and Ahhh" will come across as a pause and she will interrupt. But you can click the dictation button and then talk, then press send when you want her to reply.
21:10 Wrong. You can talk to Pi AI like a conversation. There is a "Call Icon" on the phone app. There is also a "Dictation Icon". Dictate and send, or just talk and pause (ums and ahhs = a pause since they aren't converted to text).
this made me remember that scene in iron man where Tony hacks the screen and shows his "competitors". "Everyone else is 5 years to this kind of tech, 10 years in the case of "mochi". "
It's trained on phone call data. It's not versatile, but if you role-play with it like you're on the phone having some conversation, you can probably get it to do various things. Like, it might sing if you sing, or something. Idk, fairly useless, but interesting.
It's easy for openAI to have low latency + good quality because of their powerfull cloud servers. For small products you can use only small models, because larger will be very slow on their small servers.
