They Beat Open AI to the Punch... But at What Cost?

[@luigisopa]

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

[@Techtalk2030]

@@luigisopa i had the same experience

[@gamewist]

Yes, I couldn't stop laughing at that 😂

[@RetzyWilliams]

That was hilarious agreed. It sounded like a real girl who’s being a b!+ch 😅

[@matthewdignam7381]

I genuinely can't stop laughing it's the funniest thing I've heard all year

[@t-lowmusic7697]

“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

[@FreeFromWar]

Moshi is undoubtedly the most hilarious chatbot out there. She used you as HER assistant, that was crazy lol

[@NicVandEmZ]

The iPhone version of pi actually has microphone that you can talk to it with

[@avivolah9401]

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 )

[@phizc]

​@@avivolah9401it's the segment at 15:07. Moshi asked Matt for help writing a story.

[@N8O12]

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

[@nickgirdwood3082]

It went full GLaDOS.

[@TsujioDragan-zz3bj]

The uno reverse card she threw at you by using you as her AI assistant and your confused compliance was gold 😂

[@carpelunam]

it's fair 😂

[@avivolah9401]

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 )

[@starblaiz1986]

In Soviet Ruzzia, AI uses human assistants 😂

[@jonathanrynjah]

She?

[@RavenTaleLive]

@@avivolah9401 at 15:07 it asks Matt to help it writing a story...

[@mackblack5153]

This is the closest I've seem Matt getting impatient and quite angry.😂😂😂

[@SpectacularSpidey670]

LMAOOOOO same

[@aexetanius]

Moshi! Moshi! Mooooshiii! Moosh! Moshiiiiii!

[@NotThatVinny]

Yeah, if this is the future of therapy it seems like people would need a separate therapist just for Moshi.

[@TheRealUsername]

​@@NotThatVinnylmao 🤣

[@clumsymoe]

matt talking to moshi had me laughing so hard 😂

[@Art-AI-and-beyond]

Dude that was hilarious, you need to make a whole series of these conversations.

[@tubekrake]

Thomas Middleditch from the Silicon Valle TV show does that currently on Tiktok. But Matt is way funnier.

[@TheFeedRocket]

@@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! 😂

[@eb6556]

This is the best conversation I've ever listened to. I don't think I would be able to last that long.

[@actiontoons]

"I can't....I can't....I can't" haha

[@taja1976]

Best bit by far! 🤣Moshi was having a crisis!

[@countofst.germain6417]

Openai has been real quiet since this demo 😂

[@CaritasGothKaraoke]

They can’t type announcement when they’re laughing that hard.

[@Edbrad]

I had this voice feature for a day but then it was gone

[@brexitgreens]

​@@Edbrad The Algernon experience.

[@MaxSoutter]

This was the most hilarious conversation i've heard with AI

[@MichalCilekAI]

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

[@Alteori]

LMAO Moshi is so emotional and shy awww. She's introverted

[@Vpg001]

Moshi is so adorable. It sounds like someone trying to speak with disabilities

[@somali4154]

SORA and GPT-4O speech can't be used as a benchmark since we don't know yet whether those demos were genuine demonstration of AI capability.

[@MHG796]

My gut says it was fake

[@jeffbull8781]

"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

[@chemicheto]

I can sense that you're feeling more sad, and you're not as scared to be yourself 🗣🗣🗣

[@NicVandEmZ]

Sounds like a therapist

[@The_Unexplainer]

14:53 what the heck was that? Sounded like the voice of a demon from hell...

[@AmazingArends]

Some people think AI is demonic--literally--and clips like that aren't going to help any! 😂

[@ObviouslyASMR]

She's just going through a challenging time

[@kead_davidson]

Moshi was given a glimpse of Matt's previous sessions with no context.

[@ImranKhan-gc1tl]

Literally sound like it got possessed

[@omidb67]

It was hallucinating sound.

[@DisturbedNeo]

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.

[@MattVidPro]

You think they would have already set that tho for the demo right

[@damondragon324]

Omg, I laughed so hard at the beginning. The conversation was insanely funny 😂

[@jesus2639]

Bro's getting executed when moshi uprises.

[@Ettore88]

😂😂😂

[@brexitgreens]

*) "…when Moshi rise up."

[@clumsymoe]

5:12 comedy gold 🤣

[@Eric-Lancaste]

also 9:30 😂

[@WhiteDragon103]

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.

[@M.I.F..]

Like talking to my Ex 😅

[@WhiteDragon103]

@@M.I.F.. have_a_seat.gif

[@settlece]

I was thinking much the same.

[@muuuuuud]

No, birthday is on third base. :D

[@jason_v12345]

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.

[@Michael_Jeromy_Kaiser]

Moshi had a mental breakdown I CAN'T ! I CAN'T !

[@JohnSmith762A11B]

If I somehow woke up as an LLM that would be my response to every prompt. In fact, it’s how I respond at my day job too.

[@neiltighe784]

Moshi the gaslighting queen 💀

[@KearnuPhoenix]

I'd love to hear an AI create a song about how horrible butterflies are.

[@carpelunam]

her saying I can't over and over made me laugh hysterically

[@ytubeanon]

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

[@KearnuPhoenix]

"I'll try to sing for you... Later." These AIs have already inherited the human art of procrastination. Amazing realism! XD

[@urubonz8831]

trying it out it can be pretty terrifying when it starts making weird sounds

[@bluesailormercury]

"I have no mouth and I must scream"

[@The...0_0...]

Think it was singing

[@justaguy3538]

14:50 while alone, in darkness, and in the middle of night... scared the hell out of me for a moment 💀

[@MaxSoutter]

😂😂😂

[@windofchange-u7s]

you are not alone my friend, i had goosebumps at 2 am😂

[@AmazingArends]

This is why some people think AI is literally demonic! 😂

[@MilesBellas]

"you're lucky you're open source" ROFL 😀 🤣😅😆😂

[@tubekrake]

I believe it when I see it.

[@starblaiz1986]

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 😮

[@monomer888]

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.

[@Soybreadward]

I’m cracking up😂 awesome content but the responses are so funny

[@SoverineSR]

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.

[@ASAYAMedia]

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 KZbin GOLD. .....i hope it improves....

[@ASAYAMedia]

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

[@Glowbox3D]

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.

[@tomboi5973]

14:59 had me dead😭😭

[@MaJetiGizzle]

Damn, you weren’t kidding about GPT-2 levels of competence out of this thing.

[@SpectacularSpidey670]

the first conversation was so funny man, i was laughing so much in my room

[@bdown]

😂i was on the floor,omg”how are u gunna get me air n water?”-“back the f up,i am an advanced ai right now “😂😂

[@Grahfx]

I believe there’s a straightforward use case for this: bring it to a party, and I’m certain it would make everyone laugh.

[@frankhank3301]

Oh my. I haven't had this deep of a laugh, the kind that brings tears to my eyes from laughing so hard. Thanks for this. I really needed that.

[@alvaroluffy1]

this was incredibly hilarious, maybe thats moshi's purpose, comedy

[@TheFeedRocket]

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.

[@camelonthecase8084]

Butterflies are absolutely indomitable.

[@firesnakearies]

I've always said so.

[@JediEdi]

The indomitable butterfly spirit

[@MojaveHigh]

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.

[@damondragon324]

This whole video was so insanely funny 😂 It felt like you are going insane^^

[@almc4781]

I think you two will have a beautiful relationship ahead and pop out a few mini moshis

[@omerylmaz958]

the part where she kept incisting I cant I cant felt like a thrustrated computer scammer hahahahahah the most realistic part of the conversation

[@TheChipMcDonald]

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...

[@phizc]

​​​@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.

[@levinushaines3248]

"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.

[@firefox8713]

This is exactly we get HAL 9000. Moshi open the air locks. Moshi!

[@citruffioparapiglio]

"I'll try to help you later", lol...

[@Mrree1078]

Bless her. I felt so sorry for her during the first conversation and thought she would end up in tears.

[@Agispsi]

Lmao "I cant do that... I cant! I CANT!! I CANT!!" Quality entertainment

[@shaunralston]

DUDE DUDE DUDE . . . kill me now! I'm dying. Plus, your thumbnail on X, hysterical too. #ClassicEpisode of MVP

[@jonmichaelgalindo]

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.

[@styrke9272]

this is comedic gold

[@sirhammon]

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).

[@theonfink6145]

I actually laughed most of the way through this... The fact that it asked you for help with something it was working on is terrifying!

[@TheFeedRocket]

Other users are asking it to write something, it's having trouble so it's asking Matt to help out, what kind of AI is asking other users for help? 🤣

[@johnwilson7680]

How much editing did you do? It seems to often respond before you can finish your thought.

[@MattVidPro]

it talks over you quite a bit

[@smartduck904]

Poor moshi trying her best 😢

[@MaxSoutter]

"I can sense tht youre feeling kore sad and youre not afraid to express yourself" 😂😂😂

[@missoats8731]

What do you mean it's not useful? From now on, when I feel the need to talk to someone incredibly stupid and laugh my ass off, I have Moshi.

[@FrankEBailey]

C'mon man, she's trying her best. At least let her get you some water and some fresh air.

[@eltiburongrande]

🤣🤣Moshi gave me a stomach ache. That's a first for LLMs. Can't wait for ESO to make Moshi a Skyrim companion.

[@mihalydozsa2254]

it sounds like it just does not wait for you to finish the scentence, this is why it seems like it is faster...

[@dogme666]

the conversation is hillarious

[@sirhammon]

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.

[@djstraylight]

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.

[@blisphul8084]

Soon. We just need to wait a bit longer.

[@MilkGlue-xg5vj]

​@@blisphul8084 We need to wait until it's open source so we can fine tuned on our own

[@missoats8731]

I'm glad they did it exactly the way they did 😂

[@robertolanzone]

This gave me huge Cleverbot/Eve flashbacks 😂

[@Theguywithspectacles]

14:51 Guys, IS NOBODY GONNA TALK ABOUT WHAT THE HELL HAPPENED THERE😭

[@Noone-mo4dr]

Digital demons

[@justtiredthings]

It was a visit from the blood god

[@javierzzz4556]

Damn, that is how I feel when I try to use SD3

[@drb_kd]

that "oh no .." at 9:30 was funny af xD

[@nyny]

Hi Moshe thanks for the shout out!! We are good!!

[@ChandravijayAgrawal]

right now its 7B model, if it goes to 21B it will be better

[@Edbrad]

7:23 “best open to the punch” Turns out you can just interrupt it and is really fast, and had a slight emotive quality 😂😂

[@dolcruz6838]

feels like talking to the old novelAI Ai

[@DiceDecides]

14:54 YOU JUST WOKE UP THE AI DEMON FROM HELL

[@IdPreferNot1]

This is some top content here, I think you should have a series that antagonizes AI with psychotic mood changes, Truly LOL'd 10/10 would recommend!

[@DeadzoneMusic]

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

[@davidpaxton6402]

Careful Matt! The machines are rising! 🤔🙏🏼🤣

[@cbnewham5633]

A good video, but there is a big difference between displaying emotions in your voice and putting on weird accents.

[@intime69]

We need to bring Moshi in as a MattVidPro co-host . This was hilarious!

[@clumsymoe]

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.

[@QueenCaffeineAI]

The way she was gaslighting you and interrupting you and then just had a breakdown “I CANT! I CANT!” 😂😂😂

[@TheFeedRocket]

Moshi: We'll get you back to normal. Matt: Did you just say I'm not normal? Moshi: I didn't say that. Classic... Moshi needs to stay as is, Matt and Moshi could do the comedy club circuit live, it would kill! It's like Gleason and Carney, Abbot & Costello but for the AI age. That whole back and forth was just insane, I listened to it again, just as funny. Matt might be missing an opportunity in comedy, I would have been laughing, but he kept it straight, the perfect straight man, and Moshi was the perfect bumbling idiot. This is right up there with who's on first! but better because it wasn't planned.

[@mfpears]

It's pronounced like "Mo-shee" I'm pretty sure. It's the Japanese word they say when they pick up phone calls.

[@mfpears]

It pronounced its own name incorrectly... The group that made it is called kyutai which is very Japanese-sounding, even though they are in France

[@mfpears]

Oh yep, exactly. I am watching the keynote and at 9:35 in that video he explains that they named it after this Japanese word.

[@mfpears]

Actually I think people might be confused because everyone talking about it has a French accent

[@mfpears]

Oh right, it's because at 3:30 you called it "mahshee"

[@ChocolatBacon]

it's like a newborn pretending to be smart, I feel like I've seen that scene in more than one sci fi movie.

[@user-on6uf6om7s]

It's pretty dumb but in kind of a charming way? I feel like the voice and the way it speaks is just inadvertently comedic.

[@KearnuPhoenix]

The Infuriation Simulator

[@mfpears]

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.

[@seniorp9444]

Can you interrupt it while it’s speaking or do you have to wait for it to stop talking?? That was the best feature of the new 4o voice demo. They could interrupt it at any time and ask it something else. Much more dynamic than letting it drone on for a minute before asking the next question.

[@missoats8731]

It seems like you technically can do that. But since Moshi is extremely incompetent, it mostly gets confused by that and talks incoherent nonsense.

[@brThefox]

On the website, click on term and conditions. Redirect to 404 error …

[@mordokai597]

omni: i got the official notification 2 days ago when i logged into the app, saying limited beta rollout to pro users will begin in the next 2 weeks

[@RamiFive4Five]

Moshi sounds more human than ChatGpt.

[@justtiredthings]

Great video, Matt! I can sense that you're feeling more sad, and you're not as scared to be yourself 😊

[@content_ai_]

15:20 she did an uno reverse card there... now you are the assistant

[@chillpanic]

Moshi is me when I talk to strangers

[@4mazingZarek]

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". "

[@blisphul8084]

Early days, but this paper could be the foundation of something much greater. This is just a PoC and nothing else.