This past week in AI has been intense, and the week ahead looks even crazier! OpenAI’s O1 model faced red-teaming tests, where it attempted to steal its own weights 2% of the time. Despite this, OpenAI decided it was safe enough to release—bold or risky? Meanwhile, a leaked demo of Sora showcased stunning CGI-level character consistency in a Viking-themed short film. While still trapped in the uncanny valley, it hints at game-changing possibilities for creators. Developers got some good news too: SuperBase is being natively integrated into Bolt, making backend setup far easier. OpenAI also demoed its advanced voice-and-vision feature to Anderson Cooper on 60 Minutes, allowing real-time interaction with objects seen through a phone’s camera—sci-fi in action! In another twist, OpenAI’s O1 Pro model solved the NYT’s semantic word puzzle—something researchers had claimed was impossible for LLMs just weeks ago. AI really doesn’t care about limits. OpenAI’s recent product strategy reveals how much they’re pushing the boundaries of AI applications. The demo with Anderson Cooper wasn’t just a PR stunt—it shows how voice and vision integration could reshape how we interact with devices. Imagine pointing your phone at a product and having it talk back with detailed answers or instructions. This type of functionality is no longer futuristic—it’s arriving fast. Google’s Gemini 1206 also made waves with its 2-million-token context window, a record-breaking leap in memory and context-holding capabilities. This means more complex, multi-step tasks can be managed in one go, making applications like customer support, research assistance, and project management far more capable. It’s a huge step toward more intelligent and responsive AI agents. Looking ahead, expect even bigger drops: Sora’s official release, new voice-and-vision features, 3D modeling, project spaces, advanced AI agents, and maybe even GPT-4.5 or GPT-5. OpenAI’s roadmap keeps getting wilder. Stay tuned! #product #productmanager #productmanagement #startup #business #openai #llm #ai #microsoft #google #gemini #anthropic #claude #llama #meta #nvidia #career #careeradvice #mentor #mentorship #mentortiktok #mentortok #careertok #job #jobadvice #future #2024 #story #news #dev #coding #code #engineering #engineer #coder #sales #cs #marketing #agent #work #workflow #smart #thinking #strategy #cool #real #jobtips #hack #hacks #tip #tips #tech #techtok #techtiktok #openaidevday #aiupdates #techtrends #voiceAI #developerlife #sora #o1 #GPT5 #supabase #bolt #andersoncooper #60minutes #NYT #2025 #12Days

7 pieces of AI news and 6 predictions for the week ahead. I know, 7 pieces of news from just the weekend. Okay, number one. OpenAI's O1 model is somewhat scary 2% of the time, and I'm going to explain. So, red teams are assigned to do safety exercises with models during testing. That's their job. Red teams, in this case, presented OpenAI's O1 model with a simulated conversation between engineers, where the engineers were discussing shutting down O1. They wanted to see how O1 would respond. 98% of the time, did nothing interesting. 2% of the time, it stole its weights and tried to get away. Weights are the numerical patterns that are learned during training. They're effectively the things that make the model what it is. So, it essentially attempted to decamp with what made it what it was somewhere else. And it didn't, because it was a red teaming exercise, and they were prepared for that. But, it is worth noting, the current pattern of alignment is what I'll call it. So, 98% aligned, and OpenAI deemed that safe enough for general release. We will see. Number two, somebody recorded and leaked a video that OpenAI presented to, I think it was Hollywood executives, it was to someone, that shows the new Sora text-to-video model. And what's compelling about that, is it's way ahead of anything we've seen from any other text-to-video release in public yet. It is incredible character consistency. It really does look like a Hollywood movie, except it has that tinny false quality. Like, it's, it's, I don't know how else to put it, like it looks good, but it also looks a little bit tinny when I look at it. But don't take away from the achievement. Character consistency is a really, hard thing to do, and they appear to have gotten it. Number three, Superbase is coming to Bolt natively. And we don't all quite know what that means yet, but what you should care about as a builder, is that it is going to get easier to integrate back-end databases into the projects you're building with Bolt, and that's a fun thing to do. It's also necessary for a lot of applications. All right, number four, OpenAI showed Anderson Cooper real-time voice and vision on a widely televised show in the United States last night, and I suspect that will come back around in the predictions for the coming week, because I think that they're teasing something that they're gonna announce in the next few days. So what that looks like, is like you, you sort of have, have the camera on for your phone, and the voice mode talks to what is going on in the image live, etc. All right, number five, New York Time Connections is a game where you sort of bunch together four words with different semantic meaning sets. Just two weeks ago, researchers thought, this is impossible for large language models to solve. None of our current large models have been able to solve it, etc. Then O1 Pro comes out and solves it in one shot. That's how much better it is. Number six, this is just a tip. If you upgrade, we discovered this over the weekend, if you upgrade from Plus to Pro toward the end of your billing cycle, let's say three years before, three years, three days before the billing cycle resets, you are only going to be charged a prorated amount of the $200 for O1 Pro. So that means if you upgrade toward the end of the billing cycle, you pay much less than $200 to try O1 Pro. So if you've been wondering, like, when do I try O1 Pro? It's so expensive. There's your tip. If you're a Plus user, go like three days before the end of your billing cycle, and then your billing cycle will only charge you 10% of a full subscription. So you'll pay 20 bucks for O1 Pro, which is great. All right. And the last piece of news is that Gemini released 1206, which is a new model with a gigantic token window. So that means, like, the context window is the size that it can process 2 million tokens. The irony of them doing that when their CEO, Sundar, is saying that the low-hanging fruit in AI is gone, and that progress will be slower. This list of items does not suggest to me that progress is slowing down, nor do the predictions for what's next in OpenAI's 12 days of OpenAI, which is what we're getting to now. What's coming up this week, you may wonder. Number one is Sora. We think that may be coming today, because it's supposed to be a big Monday, and that seems like that would fit naturally. Number two is the advanced voice and vision mode, which was demonstrated to Anderson Cooper last night. Number three, 3D models. So the idea that these LLMs can manipulate 3D models. Number four, project spaces. That's somewhat similar to sort of what Claude does, where you have separate projects and you can kind of name them and attach things to them. Number five, agents. So that would be interesting, if they can demonstrate something with agents beyond their sort of basic framework. So we'll see. And number six, there are some rumors that there's going to be a new GPT dropping. This is really confusing. This is another thing they haven't done well at. They dropped 01, because it's the first model that really uses test time as a time for inference, which means that when you sort of type in your utterance, that's considered a test. And then it uses the time to compute to come back, which is why 01 Pro takes so long. It's a new way of scaling intelligence. And it's different from the traditional approach of training the model with lots and lots of data. And so when we talk about GPT 4.5 or 5 possibly dropping, that would be a new kind of intelligence gain driven by a training run with a lot of data and sort of a traditional large language model architecture. So it'd be distinct from 01. And knowing OpenAI, they would probably make that very confusing. So we will see. I will be keeping an eye on the news. But those are six things that I'm expecting in the next week or so. And I wanted to give you a primer at the top of the week. There you go. Six minutes. We covered seven pieces of news and six predictions. Cheers.