Innovative Features of OpenAI's New Model

0:28Have you subscribed to AIChatDaily.com? daily.com. There's a big subscribe button in the top corner. I'll probably leave a link in the show notes to the newsletter. And also is just a full news site with tons of articles, probably published about 10 of them every single day. So if you want to go check that out, AI chat daily.com. Okay, open AI is shipping GPT live ones. This is a full duplex voice model. And what duplex voice model means? Well, this is actually kind of crazy, because Miriam Maradi, who was formerly at open AI, she released something recently with her most recent company, thinking machine labs, where it's, I mean, essentially it looks like OpenAI has just knocked her off.

1:02So this is going to be interesting. Also, China has flagged Anthropics Cloud code. They say that this is a security risk because there's some backdoor risks with it, which I don't know. I just think anything with security risk is ironic coming from China. NVIDIA's Nemotron 3 Ultra has just hit closed model parity at a 10x lower cost. And this is from some experiments I've been doing on laying chain. I'm really excited when we're able to get the costs down on some of these models. Meta is adding a tamper-proof LED to their AI Meta glasses, and they're also getting more data collection from other parts of the glasses, so there's kind of a mixed bag with what's going on there.

1:39An AI, the OpenAI chief futurist John H.E.M. is leaving. He's been there for nine years. He's one of the OG trust and safety people over OpenAI. But let's kick the podcast off with OpenAI's latest model, GPT Live 1. This is a voice model. It listens and it speaks at the same time, which is something that you've probably noticed a lot of these voice models have struggled with in the past. If you're talking to, you know, what voice mode is on ChatGPT right now, which, by the way, this replaces their advanced voice mode. But if you've ever talked to it before, you talk, there's a couple seconds of latency, then it responds back to you.

2:12And they've done a lot of work to try to, like, get that latency down as much as possible. But what's really cool about this new Live One model is that it's literally listening, and it's able to listen and talk at the same time, because that's how actual humans talk. when we're in a conversation, we kind of interrupt each other, we talk over each other, they're saying something, you're saying something, and I mean, the goal is not to interrupt and talk over each other, but it's just part of communication, right? Like, we don't always wait for there to be a perfect pause at the end of everything that someone says before we jump in.

2:37And so that's what these models are getting really good at, which is really interesting. And what's, to me, really crazy about this is that Miriam Morati released a model just like this a number of months ago, I remember covering it on a podcast, and it was like the big thing from Thinking Machine Labs. And I was like, man, this is a really cool idea and concept. well turns out you know a couple months later anyone that releases any sort of cool new feature a couple months later the frontier labs are going to be able to clone it reverse engineer and put it out themselves because you know they don't want to fall behind the reason why i think this is one in particular is an important feature is because there's three separate things that are getting collapsed into one step number one it's cutting the delay so it's going to let your conversation be way more natural and they're also going to be rolling this out to 150 million chat gpt voice users.

3:22And finally, the way this works is really interesting to me. So there's something called GPT Live 1 Mini, and this is going to be for basically all of the free users. It's kind of the default. But if you're on the paid tier, if you're paying$20 a month for ChatGPT, you get the larger GPT Live 1, which is going to send some of the harder questions over to GPT 5.5 mid-conversation. So you're going to be talking in the middle of the conversation if you're saying something kind of complex, and it's going to basically get half your sentence, send it over to GPT 5.5 to try to come up with a response and send it back and put it out before you even finish your sentence, which is really crazy.

3:53It can be interrupted in the middle of a sentence. So if it's talking, you can interrupt it. It handles silences while you think. So just because, you know, you're pausing to think about something doesn't mean it feels like it has to jump in, which is something really annoying. I feel like I can barely, like, keep my thoughts. Like, I can barely think when I'm talking to these things because I got to talk so fast because if I stop talking, it's going to want to jump in and start talking to me. So I really appreciate that. And it's going to show charts or images alongside whatever you're talking to it about, which, you know, before it was just kind of this like animated strobe on your screen.

4:23But now if you're talking, you're like, hey, you know, what are the, you know, what's the weather going to be like, it's actually pulling up weather charts on the screen while it talks to you. So you get visual and audio, which is really cool. The product leave on this is or the product lead is Addy Aledi. He said that he held a 30 to 40 minute conversation with the model on walks. And I think the goal and probably the reason why they're telling us this is because they're trying to make this something for or trying to explain this is really good for long running tasks. This isn't just quick commands like, hey, what's the weather like?

4:53It's like, hey, I'm trying to break down this complex business plan that I want to, you know, create, talk with me about this, or I'm trying to figure out what the best strategy is for XYZ, talk to me about this. And you go and have a 40 minute conversation to learn something to plan something to actually get something done, which if I'm being honest, I was kind of moving away from opening eye and chat for a while, because it didn't feel like that was their focus. And I was going to Claude. But I've been trying out a bunch of things on chat gpt lately mostly like whenever my whenever my tokens my token limits reach on clod i go back to chat gpt and i've been really impressed by the research it can do where i'm like hey you know go scrape like these 5 000 things and do an analysis on these different websites or these different products and like bring me back uh you know help me create a list of 365 xyz things i can do for this like it goes and gets some real serious work done so I'm really excited.

5:43I think the reason why this is particularly important beyond just humans is because we're going to have a lot of work with earbuds and wearables that OpenAI is currently working on. And so a lot of the hardware that they're making, these tools are going to be really useful for. And, you know, when we start talking about robots, getting them and all that kind of stuff and being able to communicate with that, this is going to be important that the communication is as smooth as possible. China is flagging Anthropics Cloud Code as a security backdoor risk, which is hilarious because China has been accused of putting backdoors into chips and hardware inside of China for many years on servers and all sorts of crazy stuff.

6:22But this in particular, I do think is probably accurate. What they've said is that Cloud Code has a backdoor vulnerability that's going to leak your location and your identity data. And they're telling everyone in China to uninstall or to upgrade the version if you have something installed between April and June to upgrade it. The advisory is interesting because Cloud Code isn't officially sold in China. So, you know, Alibaba has banned Anthropik tools company wide. This is also a week after Anthropik accused Alibaba of stealing their AI technology. And what was interesting to me is there's like some some moments where different people inside of I think Exome, they had like an AI developer conference, I think, and they were kind of publicly acknowledging that they were using cloud code at a state organized forum.

7:06This is back in March, even though it's, you know, officially not available in China. So it's funny, because it's like officially not available, officially not sold, Alibaba is banning it, no one's supposed to be using it. And yet China's like warning everybody, like they know they're using it. Anyways, it's kind of crazy for me. What I do think is important, though, is this particular flag they have is version 2.191 to 2.19656, but we're already on 2.12 for the version that we're currently on. So if you're on the latest Anthropik thing, this isn't really relevant to you. Most modern coding assistants, so if you look at OpenAI and Google, they're sending what's called telemetry and crash reports back to the main servers.

7:49And anytime that something happens, I think this is what China's worried about, the crash reports. They're worried it has too much information about the users in it. but China wasn't really clarifying if quad code was the only one doing this why they were singled out and not like open AI and Google I'm not even sure if Google's is available in China so anyways it's China's a crazy crazy place in a big win for open source models NVIDIA's open source Nemo Tron 3 Ultra model now matches the best closed AI systems on business tasks so not everything but business tasks specifically, and it can do them at a tenth of the cost.

8:26Langchain actually tuned the agent software around the model without retraining it. And they basically proved that the right engineering environment can compete with really expensive proprietary AI on when you when it comes to doing like enterprise work. So you can get these open source models, if you fine tune them, you get them just as good for way cheaper than something like Anthropic or OpenAI. LaneChain adjusted system prompts, tool descriptions, and middleware. They did all of that to really just squeeze the top tier performance. Then they put it through the deep agent benchmark. Ambridge, Amdocs, Box, and EY are already building specialized agents on this exact stack for their platforms and their workflows.

9:08The tuned model is live on six hosting providers, so you can get it on Bastion, Crusoe Cloud, DeepInfra, Firework, Nebius, and Together AI, which is what I use for AI Box. So it's, you know, pretty widely available. This is really exciting. And I believe we actually do have Pneumotron over on AI Box right now, which by the way, if you haven't tried AI Box.ai, my own startup, we have over 80 different AI models all on there. So for$8.99 a month, you get access to all of the top AI, image, audio, video, and even music models. So if you don't want to have to have subscriptions to all these different things and have, you know, logins on a ton of different platforms, just go to one place, AI Box.ai.

9:44We even have an MCP that will tie all of those models into things like Claude. So if you want to get like image and audio and video inside of Claude, you can use our MCP$8.99 a month. I'll leave a link in the description to that as well. Meta just added a tamper-proof LED to their AI glasses and it disables the camera if the light is modified or destroyed. So the idea being, you know, you want to know when someone's recording you with the Meta Ray-Bans or any of these other Meta AI smart glasses. A lot of people have been complaining about, you know, non-consensual recording. People are recording them when they don't know or they don't consent.

10:16People are in the gym recording, whatever. And so they've added a light, which I think a lot of devices have, like a light that pops on when it's recording so you know someone's recording. But theoretically, someone could go with like a pin and just like, you know, kind of shatter the LED bulb or wreck it, right? So it doesn't turn on or cover it or do something. So they're making it so that you can't tamper with it. If it's broken, it won't actually record. And this is, I mean, I think a lot of people are really excited about that. But this is also happening the same week that Meta is expanding the data collection elsewhere.

10:45So they're going to be allowing their AI to train on public Instagram photos by default, and they're testing glasses that continuously collect audio, and they snap photos every few seconds, which, come on, that's terrifying. But I mean, maybe they're mailed in a product that they think some people will like that. I really doubt that, but this particular prototype that is under testing is going to move the glasses from being an on-demand camera, right? So you're like, hey, meta, record whatever I'm doing, and it turns on, to actually just being always on and having an always-on sensor, and it's feeding continuous audio and periodic photos to Meta's AI systems.

11:18They don't have a date launched for this. I hope there'll be a way to opt out. Otherwise, I don't imagine these glasses will continue to be very popular. For the Instagram thing, this is a change that will affect about a billion, you know, billions of photos for users. Most, like, everyone posting is not wanting their photos to be trained, like, AI to be trained on them. So that's going to be kind of crazy. OpenAI's chief futurist, Joshua Achiam, is leaving after about nine years. He's one of the safety advocates that's been at the company the longest since the early days. He is leaving, and he's the fifth senior safety-focused leader to leave OpenAI in the last two years.

11:56They're getting ready to go public. He actually joined as an intern back in 2017, and then he led their mission alignment team until it was disabled, or I guess disbanded. banded. Sam Altman got rid of it back in February of this year. Then he moved into chief futurist role and was kind of studying AI harms and benefits. Five different safety adjacent leaders have now left OpenAI since 2024. And a lot of them have started nonprofits focused on AI safety standards. The former White House AI advisor Dean Ball started this week as the head of Strategic Futures and is going to briefly overlap with Echium before he actually leaves and will kind of take over a lot of policy focused research work without having a direct replacement for the chief futurist title.

12:39So if you believe you'd be perfect for that job, maybe you should send in your application to OpenAI. Guys, thank you so much for tuning into the podcast today. Make sure to check out AIbox.ai. Links in the description. If you have not already left a review for the show, it would mean the world to me. It helps the show out tremendously. I have some exciting interviews and content and stuff planned for the show this year. And if you would be able to leave a review, it would help me to bring you more incredible content. Over on Spotify, you have to hit the About tab. If you've listened to three episodes, you're eligible to leave a review.

13:07Otherwise, I won't let you. So if you have, that would be incredible. Hit the About tab, hit some stars. On Apple, it's kind of annoying to find, but you go to the podcast page and you can hit the stars and leave a comment. I read them all, I appreciate them. All right, guys, I hope you all have a fantastic rest of your day.