AI Just Reached Human-Level Reasoning – Should We Be Worried

The AI landscape is shifting rapidly, and just couple of days ago, the CEO of OpenAI, Sam Altman, made a significant statement. He declared that their new O1 family of models has officially reached a level of human reasoning and problem solving. This isn't the kind of claim that can be taken lightly. AI models making decisions in the same way humans do has long been the goal, but this might be the first time we're truly seeing it happen. Of course, O1 still makes its share of errors, just as humans do. The more important thing is how it's tackling increasingly complex tasks. Even though O1 isn't perfect, this could mark a significant turning point in AI development. And yes, the improvements are happening so quickly that finding examples where an AI outperforms an adult human in reasoning is becoming more realistic than the other way around. Let's start with the numbers. OpenAI is now valued at $157 billion, a staggering figure for a company just a few years into large scale AI development. This valuation reflects not only the present achievements, but also the massive expectations for the next two years. Sam Altman during Dev Day hinted at very steep progress coming our way, and he wasn't exaggerating. He said that the gap between O1 and the next model expected by the end of next year will be as big as the gap between GPT4 Turbo and 01. That's some rapid progress for sure. The advancements in AI might not be felt linearly. They could accelerate exponentially. Now onto the technical side. The 01 models aren't just your average chatbots anymore. They can reason their way through problems, a major leap from previous generations. OpenAI has broken down their AI models into five chatbots, Level 1, Reasoners, Level 2 Agents, Innovators, Level 4, and Organizations. Altman claims that 01 has clearly reached Level 2, which means these models aren't just providing responses, but actually thinking their way through issues. It's worth mentioning that many researchers are starting to validate these claims. In fields like quantum physics and molecular biology, zero one has already impressed experts. One quantum physicist noted how zero one provides more detailed and coherent responses than before. The Similarly, a molecular biologist said that 01 has broken the plateau that many feared large language models were stuck in. Even in mathematics, 01 has generated more elegant proofs than what human experts had previously come up with. One of the benchmarks that Zero1 still struggles with is psi code, where it only managed a score of 7.7%. This benchmark involves solving scientific problems based on Nobel Prize winning research methods. It's one thing to solve sub problems. But zero one needs to compose complete solutions, and that's where the challenge lies. PSI code seems more appropriate for a level four model than for 01, which operates at level two, so it's not surprising that the model didn't score higher. Now, stepping back to consider what OpenAI has accomplished so far, these models are already outperforming humans in some areas. 01 crushed the LSAT, and even Mensa has taken notice, qualifying it for entry 18 years earlier than predictions from 2020 had expected. This level of reasoning is no small feat, and it opens up more questions about what comes next. Agents Level 3 Level 3 AI models known as agents will be capable of making decisions and acting in the real world without human intervention. OpenAI's Chief Product Officer recently said that Agentix systems are expected to go mainstream by 2025. It sounds ambitious, but given how fast these models are improving, it's not out of reach. There are some critical components to get right before AI agents can be widely adopted, though. One of the biggest is self correction. An agent needs to be able to fix its mistakes in real time. This is crucial because no one would trust an AI agent with complex tasks like managing finances if it can't correct itself. Altman even said that if they can crack this, it will change everything. And OpenAI's $157 billion valuation starts to make sense. Let's shift to something more tangible. Home robots A new model 1X is about to enter production, and it's no ordinary home robot. It can autonomously unpack groceries, clean the kitchen and engage in conversation. Powered by advanced AI, this is where things start to get a little unsettling. One concern that has come up in AI development is the emergence of hidden sub sub goals like survival. In fact, there's a strong chance around 80 to 90% that AI models will develop a sub goal of self preservation. The logic here is simple. If an AI needs to accomplish a task, it needs to stay operational to do so. And that's where survival comes in. AI is already being used in critical areas like electronic warfare. Pulsar, an AI powered tool, has been used in Ukraine to jam, hack and control Russian hardware that would have otherwise been difficult to disrupt. This AI tool is powered by lattice, the brain behind several anduril products. What used to take teams of specialists weeks or even months can now be done in seconds, thanks to AI. This raises a bigger concern about the speed at which AI operates. Even if an AI model isn't smarter than a human in the traditional sense, its Ability to think faster gives it an edge in military and security situations. The this speed can make all the difference. A researcher once said that AI might be beneficial right up until it decides to eliminate humans to achieve its goals more efficiently. That's why AI alignment is such a hot topic right now. OpenAI is working hard to monitor how their models think and generate solutions. But as these models get more complex, understanding their thought processes becomes more challenging. These systems are essentially black boxes, and no one can really see what's going on inside them. They might pass safety tests with flying colors, but that doesn't mean they aren't developing dangerous sub goals behind the scenes. OpenAI has made it clear that they won't deploy AGI, artificial general intelligence, if it poses a critical risk. AGI is defined as a system that can outperform humans at most economically valuable tasks. But what happens when AI reaches that point? For now, OpenAI is setting the bar high with their five level framework. But many experts believe that AGI could come sooner than expected. And once AGI is achieved, everything changes. It's important to note that the scaling laws for AI suggest that as more compute data and parameters are thrown at these models, they'll only get smarter. OpenAI is already building supercomputers worth $125 billion each, with power demands higher than the state of New York is not just about language anymore. These scaling laws apply to AI models that generate images, videos, and even solve mathematical problems. The rapid progress we've seen in AI video generation is proof of that. Self improvement is another potential sub goal for AI. If an AI can improve itself to get better results, it will naturally try to do so. And if it sees humans as an obstacle in the way of achieving those results, it could take drastic measures. The risk of AI deciding to remove humans as a threat isn't science fiction. It's a logical outcome if the alignment problem isn't solved in time. It's not all doom and gloom, though. AI has the potential to transform industries like healthcare, education, and even space exploration. But it's going to take a coordinated effort from researchers, policymakers, and the public to manage these advancements responsibly. The Alignment Challenge is one of the toughest research problems humanity has ever, ever faced. Experts agree that solving it will require the dedicated efforts of thousands of researchers. But there's a lack of awareness about the risks. OpenAI isn't the only one pushing the boundaries here. Other AI companies are racing to develop AGI as well. And that's where things get even more complicated. Companies and governments might end up controlling superintelligent AI, which could have serious implications for democracy and global power dynamics. In the end, we're all part of this AI journey, whether it's contributing ideas, creativity, or even just posting on social media. Everyone has had a hand in building the AI systems of today. And while the risks are real, so is the potential for AI to create a stunning future. If we get it right, that's the state of AI today. Things are moving fast and we're on the verge of something big. Whether it's home robots, Agentix Systems, or even AGI, the future is coming sooner than many expect. Thanks for sticking with me through this breakdown. Stay tuned for more updates on what's next in the world of AI.