AI News: Things Are Getting Weird

So there's been some really interesting storylines and developments in the world of AI this week, and in this video, I'm going to break them all down and loop you in.

Before I really get into it, I want to take a moment to really appreciate how far AI generated images have come. So this image that we're looking at on the screen right now, this was generated with Flux, which is an AI image generation model that I made a video about earlier this week. But the way it can make humans is just mind blowing. I mean, the only really obvious giveaways in this one is if you look at his lanyard here, the text just looks like gibberish. And then I think this is supposed to be a microphone right here, but it looks a little off. But if you're just looking at the guy's face and you're looking at the hands and you're looking at the shirt, it looks really, really good. Here's another image from the same Reddit thread, and if we look at this one, I mean, same thing, it is really, really difficult to tell that it's AI. It's. Now, if we zoom in really close, like I am, you can tell maybe there's a little weirdness going on with the teeth, especially on the right side, where the teeth get a little bigger. Then if we zoom in on the microphone, it looks just kind of like gibberish around the top of the microphone. But other than that, I mean, the Google logo is there. The lanyard, it's kind of hidden by our hand, so you can't see it. But dang, AI imagery has come so far that it's gotten to a point where most people are not going to be able to tell what was AI generated and what wasn't. We can only tell right now because I'm zooming in super close and looking for little giveaways. Anyway, this is from a Reddit thread over on the Stable Diffusion subreddit. They did follow a very specific process to actually get these images to look this way and look so real. But they do break it down in this Reddit thread, so I'll make sure it's linked below if you want to try to create images that are this realistic. All right, now let's talk about OpenAI. We're probably going to spend a few minutes talking about OpenAI today because there's been a lot of weird stuff happening there, starting with probably the most, like, intriguing thing that's come out of OpenAI this week, and that's that Sam Altman posted this random image on His Twitter account that just says, I love summer in the garden, and it has pictures of his strawberries. Now, if you're not familiar, or maybe you didn't see some of the older videos where I was talking about what this is in reference to, OpenAI supposedly had this model called Q that was really, really good at reasoning. There was suspicions that that's what Ilya saw when the Ford tried to oust Altman. And there was claims that this Q Star model was just too powerful, too scary to unleash into the world. Supposedly, that Q Star model has been rebranded as Strawberry. So now Internally over at OpenAI, they refer to this more advanced, better reasoning model as the Strawberry model. So this post here by Sam is essentially Sam just kind of trolling. The people paying attention to OpenAI, acting as if something huge is coming related to Strawberry. At least that's the way most people are interpreting it. These seemingly are genuine strawberries from his garden, and he might have just wanted to post a picture of his strawberries because he was proud of them. But a lot of people are reading a lot more into this. There's also this Twitter account here that has a lot of people baffled. So the Twitter username is ruletheworldmo, but the actual written out username is just three strawberry emojis. And this Twitter account grew insanely fast. I mean, a couple weeks ago I think it was like under a thousand followers. And just since all of this strawberry weirdness has started popping up again, they've jumped up to over 13,000 followers. And if you look through their feed, it's pretty much just a non stop feed of strange trolling memes about strawberries. See you in a while, travelers. I'll be back. Strawberry chess boards, Strawberry matrix images. Just lots and lots of random memes and troll posts about strawberries. One thing to note is that they do write in all lowercase, which is something that Sam Altman is sort of known for on his Twitter account, which has led to rumors that maybe this is Sam Altman's alternate account. There's also been a lot of rumors that this is just a bot and that maybe OpenAI or Sam Altman or somebody created a Twitter bot that just continually posts memes and random stuff about strawberries and retweets anything related to AI and strawberries and Sam Altman. My personal guess is it's probably just somebody with a little bit too much time on their hands trying to just stir things up. And it seems to be working. One interesting thing is that Sam Altman himself actually appears to be paying attention to this account. Welcome to level two. How do you feel? Did I make you feel? And then Sam says amazing, to be honest. Again, my guess, somebody with a lot of time on their hands just screwing around with Twitter, trying to see how quickly they can hype things up and get excitement going. But I kind of secretly hope that it's some sort of AI bot created by OpenAI to build additional hype around what they're launching next. I'm fairly confident that's not the case, but I also kind of want to be wrong. Moving on. There's been a lot of other really interesting news and storylines coming out of OpenAI. I actually made a whole dedicated video about all this sort of drama unfolding over at OpenAI right now. John Schulman, one of the co founders over at OpenAI and actually left to go join Anthropic this week. Peter Deng, a product manager over at OpenAI also left. And Greg Brockman is going on extended vacation. Essentially he's going on sabbatical and said that he's stepping away from OpenAI until the end of the year, leaving a lot of people to believe that he's also just ditching and moving on from OpenAI. I don't get that impression. There's been a lot of pressure, a lot of stress, a lot of drama. A lot of things have happened over at OpenAI recently, so I genuinely think he's just stepping back to chill out and sort of get out of that mix for a little bit. But a lot of people are speculating that he's gone as well. On the 8th, he posted this, which also leaves me to believe that he is planning on coming back. He says a surprisingly hard part of my break is beginning the FOMO. Everything happening at OpenAI right now, lots of results cooking. I've poured my life for the past nine years into OpenAI, including the entirety of my marriage. Our work is important to me, but so is life. I feel okay taking this time in part because our research, safety and product progress is so strong. I'm super grateful for the team we've built and it's unprecedented talent density and proud of our progress, looking forward to completing our mission together. All indications point to he's stepping away, but planning on coming back. There's a lot more to the OpenAI drama that's happening right now, but I broke it down in this video here called OpenAI. Seems like it's imploding, so check that video out if you want all of the details and sort of my thoughts and opinions on what's going on at OpenAI right now. But in other news, that wasn't part of that video. OpenAI got a new board member this week named Zico Coulter. Zico is a professor and director of the Machine Learning department at Carnegie Mellon University and they're predominantly focused on AI safety, alignment and the robustness of machine learning classifiers. We've seen a lot of storylines lately of people claiming that OpenAI isn't taking safety and alignment seriously. Yann Leake, left, wrote up a big X post about how he doesn't think they're taking safety seriously. OpenAI disbanded their safety and alignment team. It sounds like they are actually taking it fairly seriously by bringing on board members that that is their main focus. But not only that, they put out this this week called the GPT4O system card. We can see here it says this report outlines the safety work carried out prior to releasing GPT4O, including external red teaming, frontier risk evaluations according to our preparedness framework, and an overview of the mitigations we built in to address key risk areas. Now this is a very long write up here, but basically it goes into all of the details and all of the ways that they are focused on safety and alignment here. They even gave themselves a little bit of a scorecard here. They feel like when it comes to cybersecurity security, it seems to be low risk, biological threats, low risk, persuasion threats, a medium risk and model autonomy is a low risk. They say that only models with a post mitigation score of medium or below can be deployed and only models with a score of high or lower can be developed further. So if something comes back at a critical score, apparently they'll just cease development on that thing. But if you are a masochist and you do want to read this whole thing here, it is linked up in the description so you can dig into all of the details. Another thing OpenAI warned about this week is that users could become emotionally hooked on its voice mode. If you've seen the movie her with Joaquin Phoenix, he pretty much falls in love with his AI and his AI becomes his girlfriend because he got so emotionally attached to that voice mode. Jumping back to that scorecard here, that was written up straight in that scorecard. They say that during early testing, including red teaming and internal user testing, we observed users using language that might indicate forming connections with the model. They were saying things like this is our last day together. They do have some worry that it could negatively affect healthy relationships but potentially benefit lonely individuals. But they do intend to further study the potential for emotional reliance. But it is interesting they're coming out and saying it right now that we haven't rolled out GPT4.0 voice to everybody yet. But one of the things we have noticed is some people are getting a little too attached. And since we're on the topic for safety, I'm going to jump over to Anthropic for just a moment because they're also making a big focus on model safety. They're creating a bug bounty program that sounds like you can apply and if accepted, participants will get early access to the latest models before public deployment with the goal of trying to to find ways to circumvent the safety measures. They're offering up to $15,000 for novel Universal jailbreak attacks. A jailbreak attack is basically convincing the AI, the large language model, to ignore its guardrails and give you an output that it's not necessarily designed to give you. If people can help them find those jailbreaks so that they can close those loopholes, they're willing to pay for that information. All right, back to OpenAI. Because they still had some more announcements and things that happened this week, they rolled out a new feature for developers. They now have structured outputs in the API. Now, if you're not a developer or you're not using the API for any reason, this is probably not relevant to you. But essentially, when you are using their APIs and code, the data that OpenAI is sending over to whatever application you're writing, the code for the data is a little bit more structured and organized, theoretically making it a lot easier to work with. One of the big criticisms of OpenAI lately, and I've been one of the people who's been personally criticizing them a lot about this, is the fact that they talk about a lot of cool things that they're doing, but then don't actually ship them and give people access. To this day, we still don't have SORA access. To this day, most people still don't have GPT4O advanced voice features. They announced a new ChatGPT search to go head to head with Google. We haven't seen access to it yet. They keep on showing off these cool, interesting things and then not shipping them. Well, it sounds like we should lower our expectations for OpenAI's next Dev Day as well. Their Dev Day happens in San Francisco on October 1st. Last year's dev Day was when they announced the GPT Store, which kind of never ended up going anywhere. And this year they're basically saying, you're not going to get GPT5 and don't expect anything too crazy from this dev day. Basically don't expect to see any massive announcements or new features in Chat GPT, but if you are a developer who's using OpenAI's tech and their large language models, there's probably going to be some cool stuff to make your development process easier. It also came out this week that OpenAI actually has a tool to recognize whether or not text was generated with AI or not, but they've deliberately decided not to release it. It says here OpenAI has built a tool that could potentially catch students who cheat by asking ChatGPT to write their assignments, but according to the Wall Street Journal, the company is debating whether to actually release it. They basically say here that they're worried about bad actors figuring out how to circumvent them. They're also worried that it could stigmatize use of AI as a useful writing tool for non native English speakers. Who knows if we'll actually see this or not? Yet another thing OpenAI says, Look, we've got this, but you can't have it. But to be honest, this isn't one that I'm like super excited about, so I'm like really bummed that OpenAI isn't giving it to us or anything. It's just interesting. It's another thing that they're like, hey, look what we have. Anyway, so you've been hearing about all of this AI stuff, but perhaps you're not sure how to apply it yet. Despite all of the concerns about AI replacing jobs, you can actually use AI tools to your advantage and position yourself as an ideal candidate for your dream position. That's why I teamed up with HubSpot to showcase their toolkit. Designed to help you secure the perfect job using AI. This toolkit includes super valuable resources like 20 AI powered tools for job seekers. And I'll admit it, as the AI tools guy, some of these tools were even new to me. You'll also find 60 ways to use Chat GPT during your job search, which includes tips on things like improving your negotiation skills. And let's face it, whether you're job hunting or not, being a better negotiator is a pretty valuable skill to have. But honestly, that's just the tip of the iceberg when it comes to the resources available in this toolkit. So if you're currently in the job market, contemplating a career shift, or simply want to enhance your skills for further opportunities, I strongly Recommend Checking out HubSpot's completely free kit on landing a job using AI. Once again, it's totally free and you'll find the link in the description below. And once again, a big thank you to HubSpot for sponsoring this video. Elon Musk is suing OpenAI again this week. I say again because he had a previous lawsuit with them saying that they basically went against what they originally formed OpenAI for, but then Elon dropped that lawsuit and then a few days later opened up this new lawsuit. Elon's basically saying that OpenAI guys manipulated Musk into co founding their nonprofit venture by promising that OpenAI would be safer and more transparent than profit driven alternatives. So while the old lawsuit was kind of claiming that they went against what the company originally stood for, the new lawsuit is much more nuanced in saying they basically tricked Elon Musk into investing in a nonprofit that they eventually started figuring out how to make profits for. But that's not the only lawsuit that OpenAI is getting hit with this week. A YouTuber named David Millette is trying to get this class action lawsuit started up against OpenAI because OpenAI went and scraped a whole bunch of transcripts from YouTube videos to use as training data. Personally, I don't think this lawsuit is going to work. I think by using the training data from these transcripts, it's essentially fair use. Especially since the output that you're getting from chat GPT when you ask it a prompt isn't the same thing as what was inputted. Without really, really tricking the large language models, you're not going to get the large language model to output word for word exactly what the input data was, therefore making it transformative enough that I don't think there's much of a case here. But I'm not a lawyer, I don't know what I'm talking about. That's just based on how I've seen some of these other lawsuits that have come up play out. Along the same lines, Nvidia is in some hot water this week when a leaked document showed that Nvidia was scraping a human lifetime of videos per day to train their AI models.