alt Hacker NewsGPT‑Live
Comments
This is the opposite direction AI should be going. Human relationships are the most valuable thing we have, and so, naturally, technology seeks to intermediate and now replace them.
I'm not Catholic, but this podcast presents a very interesting argument against talking to AI as if they were human: https://newpolity.com/podcasts-hub/debate-chatbots
What I’m missing from this announcement is the capability to use connectors and tools. I don’t really get it - NONE of the frontier assistants can use tools / connectors while in voice mode - Claude, ChatGPT, Gemini, Grok. It seems so obvious: I want to be able to research stuff, pull up documents, jot down notes and do productive work while I’m talking to it, and not end voice mode whenever I need to connect to an app or service.
It’s weird. The old Claude voice mode WAS able to use tools but when they revamped it, it lost that capability and is now pinned to Haiku :(
So, yay for finally a voice mode that’s powered by a frontier model and hopefully as good as Grok voice, but sad to still not see tool use while in voice mode.
(I haven’t tried it yet, only read the announcement)
I have not used voice mode much with chatgpt. I was surprised to learn that they were already not running the voice model like a UX orchestrator while utilizing other models in background for actual research/response etc. I guess it's good they launched what they could and got here in steps. I suspect in the near future my personal device (mobile/laptop) will be powerful enough to run any UX orchestrator model locally – and route to multiple frontier closed/open model providers in the background as appropriate. The battle is going to be platform owners (Apple/Google/Microsoft) wanting to lock-down the access to that local hardware and local interaction paradigms (ambient always-on full-duplex voice) and intermediate through their platform layers - rationalizing it as consumer security/privacy protection (which is right for most people, but sucks for the open market). Meanwhile I suspect OpenAI/Meta et al will try to build their own hardware and become platform owners themselves, though unsuccessfully. And it's going to take some company like epic games to get them to open that up. and that's probably what the next decade is going to be all about.
This voice is awful, possibly one of the worst AI/computer voices I have heard in like two years now - what's up with that? Is this seriously the best a company burning this much money can do, and they consider this acceptable to release? Like two syllables in and my first response was to grimace and physically cringe. Does anyone here think this sounds good, or even just "fine enough"?
Do engineers that work on this for long periods of time stop seeing the forest for the trees and think this could be mistaken as human? I'm saying this as someone that assumes all narration/most VO work will be fully AI fairly soon (for better or worse.)...
Much better than it was before but it’s still significantly weaker than a direct chat.
For example I asked
“Why should LLM attention use dot product instead of cosine similarity, being that we often hear vector magnitude does not encode most of the useful information needed”?
The voice response was directionally right but lacked detail and was a little hand wavy.
The answer to the same question in a text chat was much higher quality.
The voice response replied “let me think about that…” so it appears to be invoking 5.5 as advertised, but it’s definitely weaker.
I had reasoning set the same for both.
(Atty from OpenAI here)
GPT-Live-1 is the first version of a new generation of models, and we believe the full-duplex architecture + delegation enables entirely new ways of human-AI interaction.
Would love to hear your feedback!
Are there any open source full duplex models that are out besides PersonaPlex? There was a chinese open one, maybe Fun Audio chat or something, that said it was going to release a full duplex version but I am not sure if it did.
My dream would be open source full duplex with function calling or some kind of rudimentary text output. PersonaPlex is still interesting although it was looking like we would need to fine tune it to handle outgoing or avoid going off the rails easily.
I'm so mad that this might make me re-subscribe to ChatGPT. I wouldn't have believed how much I use the voice feature before LLMs and ChatGPT currently has the best voice interface. I think Grok's interface is the next best, then Claude.
I'm crying, that guy who tries to get it to count to 100 may actually have a chance
Does this model do better ignoring side conversations? That's the biggest hindrance to using ChatGPT's carplay feature is someone will say something, stopping ChatGPT from speaking or taking it in a different direction.
For the first actor - why does her accent change the longer she talks. It's like they had an "Estelle Costanza" dial that they started at zero and slowly rolled up to 8 or 9.
This looks very cool. An AI that can listen and speak and handle tasks without breaking the flow of conversation would solve some big annoyances with current tools.
The concern is though as these get better will people struggle to distinguish these with real human connections?
Gemini live has been able to do this for over a year now. I can just activate it on my phone and it really works surprisingly well, especially the interruption. I've tested it with my 95 year old Dutch grandmother and it switched seamlessly between English and Dutch with her and handled her poor hearing very well, including her asking for repetition.
I'm a little surprised by how much OAI is playing catch up here.
There doesn’t seem to be any indication whether this is available in the chat-got app nor is there any indication in the app that anything has changed. Anyone know how to actually try this?
This solves my biggest annoyance with the current advanced voice: its speech getting interrupted by me setting yup or even background noise if loud enough
If an idiot has this on in the subway my conversations are surveilled. What is the antidote? Train another model to talk about bombs etc. and flood the clanker (and by extension the FBI)?
Last night, I was using voice for the first time in a few weeks, and it interrupted me and said, a bit aggressively...
"I'm going to stop you right there. Let's keep the conversation focused on the topic we were covering or a new relevant topic".
I tried to probe it for why it did that, what rules it was following, and it eventually told me...
"My role is to keep us focused..." and, "The behaviour you saw was my attempt to moderate tone".
I've heard of LLMs doing weird things like this, but it was the first time it happened to me. I hope they fix that. It was creepy.
For context, it heard my partner say, "I guess it's the same thing as you mom, because she's..." and then it cut us off.
I cannot wait for the qwen version on huggingface.
Hoping to use this for natural conversation language learning. Previous iterations of the app kept correcting my words/grammar before it got to the model, causing issues with identifying mistakes in speech
Not had a chance to experiment yet, but will this be an upgrade for using AI for language learning?
The demo video shows quite how rough around the edges this is....
Doesn't quite stop fast enough when you interrupt it. Can't find info quick enough so you have to change topic and then have it give you results later, etc.
This is a move in the right direction, but there is lots of engineering still to be done!
I was hopeful that they avoided the well known sultry voice this go around, but alas. There is little hope for these companies.
The full duplex is awesome, and the feedback that it is getting what you're saying is ok, but in some of the demos was a little overkill.
I'll agree that using the "Golden Girls" was at least more entertaining than the usual pitch.
Great to hear about full-duplex. When using voice mode historically, it was infuriating to have the AI go on a long-winded rant or explanation and I would be shouting again and again "stop. shut up. shut up! shut up!!!"; I just needed a clean way to interrupt it.
This feels so dehumanising.
Hopefully this also means it does better with "interruptions". I have used ChatGPT Voice in the car before and sometimes car/road noises will cause it to stop responding in the middle.
I used it to help me figure out how to turn off a feature in the rental car I was in (adaptive cruise control, I love it but snow blocked the sensor and I wanted just normal cruise control but couldn't figure it out while driving).
This kind of voice chat is awesome and I'll be even more excited when open models have this functionality. I'd love something like this paired with Home Assistant (assuming we ever get decent hardware).
I use AI for my job. I understand the impact (as much as anyone can) that it will have on society. I recognize the value.
But I just want to say that talking with AI casually is critically lame. I cringe every time I have to ask my Google Home to turn on the lights and people are having full-on conversations with it? And, imo, dangerous considering how sycophantic AI is. The stupidest, most gullible, most insecure person right now is looking at this thinking they are about to make a new friend.
I was missing the part where the grandmas are saying "I have absolutely no idea what I just said" :D --- Besides that really nice demo I will give this a try, I tried some voice models before and the issue is I will ask a question, get a answer withing the next ~1 sentence and then the usual LLM bs follows which I just wanted to skip at that point
Looking forward to a list of support languages. This would be amazing for language listening/speaking practice. Yes I know it doesn't replace the talking to real people and "immersion". But cheaper than a flight
I looks like they took inspiration from Thinking Machines - http://thinkingmachines.ai/blog/interaction-models/
I like this and felt like some of it was much more fluid; but was I alone in feeling like the interjected "uh-huh" or "yeah?" moments felt a little jarring?
Almost felt a bit *uncanny valley* for what "natural" conversation is supposed to be like. If the "uh huh" isn't timed correctly, it'll feel like a zoom call with lag.
I wonder why they don't compare it to their existing live voice model realtime-2.
Oh wow, I'd like this. Our current voice interactions with ChatGPT are on a 4o era model; really terrible. oAI has always been pretty cagey on the architecture of their end to end multimodal models. And RL has basically made them worse since launch. (Check the launch videos where the model sings, is more realtime, has accents, etc). I'd love to try a next gen version.
The new architecture makes sense, it seems many of the remaining problems like noise and interruptions are at the sound processing and integration level rather than at an architectural or model level now which makes for an exciting new era.
The potential conversational dynamics of people telling each other "quiet!" after they pick up the habit from talking with AI will be interesting. It could lead to people being more assertive and thoughtful, or it could be contentious and rude.
Awesome that they've improved that aspect of voice chat, though.
Very cool. I thought the agent came in a little to hot at 1:03. I wonder how it decides when to jump in.
I'm very eager to test this for brainstorming!
One thing I noticed is that we lost vision feature for some reason on the live chat?
This was an extremely useful feature. Not sure if it’s a regional thing or that they just removed that from the current live chat.
I imagine it will be even more useful with this new version.
It would be great if we could have AI that wasn't trying to emulate a human. When it expresses emotion, we should see that as a bug that needs to be fixed.
When GPT-Live in Codex, so i can walk the dog while shipping?
How does the voice model delegate requests to GPT-5.5? Can the voice model generate text?
I do not fully understand the complexity behind achieving full-duplex but I hope this sets the bar for Anthropic to follow. Turn-based simplex is yesterday.
It sounds like they've switched to a "native audio" model which if I understand right is what Gemini has had for quite a while?
I'm at the point where pricing is the first thing I look for in announcements like this.
Very cool. Not cool bringing Brazil’s loss to Norway again. We're already devastated. No need to keep beating someone on the ground. :(
this is excellent, I've been meaning to update my phone-dialer.apk "fake phone conversation" for when I'm trying to get out of a social sitation (not joking)
Can a model like this critique your accent/pronunciation? That would be cool.
Does this support more than one user voice? Or, are there plans for this? I did not see that mentioned in the announcement.
Oh gosh. I was watching the video and thought it is a live stream. I just noticed when it restarted.
I had preview access to this one for a few weeks. It's very good. I had one conversation that lasted a full hour while I was walking the dog, got some good brainstorming done against one of my projects.
The best feature is that it can delegate questions out to GPT-5.5 in the background, so you're no longer restricted to a voice model that's several years behind the frontier.
I did report a fun bug with it though: it was interrupting me and laughing at my (not really intended as) jokes while I was still talking! They seem to have clamped that behavior down thankfully, it felt a bit rude and condescending.