00:00The AI landscape is shifting rapidly and just a couple of days ago, the CEO of OpenAI, Sam Altman, made a significant statement.
00:10He declared that their new O1 family of models has officially reached a level of human reasoning and problem solving.
00:17This isn't the kind of claim that can be taken lightly.
00:20AI models making decisions in the same way humans do has long been the goal.
00:24But this might be the first time we're truly seeing it happen.
00:27Of course, O1 still makes its share of errors, just as humans do.
00:31The more important thing is how it's tackling increasingly complex tasks.
00:35Even though O1 isn't perfect, this could mark a significant turning point in AI development.
00:40And 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.
00:49Let's start with the numbers.
00:51OpenAI is now valued at $157 billion, a staggering figure for a company just a few years into large-scale AI development.
01:01This valuation reflects not only the present achievements, but also the massive expectations for the next two years.
01:07Sam Altman, during Dev Day, hinted at very steep progress coming our way and he wasn't exaggerating.
01:13He 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 GPT-4 Turbo and O1.
01:22That's some rapid progress for sure.
01:24The advancements in AI might not be felt linearly, they could accelerate exponentially.
01:28Now, onto the technical side.
01:30The 0-1 models aren't just your average chatbots anymore.
01:34They can reason their way through problems, a major leap from previous generations.
01:39OpenAI has broken down their AI models into five levels.
01:43Chatbots, Level 1.
01:45Reasoners, Level 2.
01:46Agents, Level 3.
01:48Innovators, Level 4.
01:50And Organizations, Level 5.
01:52Altman claims that 0-1 has clearly reached Level 2,
01:56which means these models aren't just providing responses but actually thinking their way through issues.
02:01It's worth mentioning that many researchers are starting to validate these claims.
02:05In fields like quantum physics and molecular biology, 0-1 has already impressed experts.
02:11One quantum physicist noted how 0-1 provides more detailed and coherent responses than before.
02:17Similarly, a molecular biologist said that 0-1 has broken the plateau that many feared large language models were stuck in.
02:25Even in mathematics, 0-1 has generated more elegant proofs than what human experts had previously come up with.
02:32One of the benchmarks that 0-1 still struggles with is Psycode, where it only managed a score of 7.7%.
02:40This benchmark involves solving scientific problems based on Nobel Prize-winning research methods.
02:45It's one thing to solve sub-problems, but 0-1 needs to compose complete solutions.
02:50And that's where the challenge lies.
02:52Psycode seems more appropriate for a Level 4 model, Innovators, than for 0-1, which operates at Level 2,
02:59so it's not surprising that the model didn't score higher.
03:02Now, stepping back to consider what OpenAI has accomplished so far,
03:06these models are already outperforming humans in some areas.
03:090-1 crushed the LSAT.
03:11And even Mensa has taken notice, qualifying it for entry 18 years earlier than predictions from 2020 had expected.
03:18This level of reasoning is no small feat, and it opens up more questions about what comes next.
03:23Agents, Level 3.
03:25Level 3 AI models, known as agents, will be capable of making decisions and acting in the real world without human intervention.
03:32OpenAI's Chief Product Officer recently said that agentic systems are expected to go mainstream by 2025.
03:40It sounds ambitious, but given how fast these models are improving, it's not out of reach.
03:44There are some critical components to get right before AI agents can be widely adopted, though.
03:49One of the biggest is self-correction.
03:52An agent needs to be able to fix its mistakes in real time.
03:55This is crucial because no one would trust an AI agent with complex tasks like managing finances if it can't correct itself.
04:03Altman even said that if they can crack this, it will change everything, and OpenAI's $157 billion valuation starts to make sense.
04:12Let's shift to something more tangible.
04:14Home robots.
04:15A new model, 1X, is about to enter production.
04:18And it's no ordinary home robot.
04:20It can autonomously unpack groceries, clean the kitchen, and engage in conversation powered by advanced AI.
04:27This is where things start to get a little unsettling.
04:29One concern that has come up in AI development is the emergence of hidden sub-goals like survival.
04:35In fact, there's a strong chance, around 80 to 90 percent, that AI models will develop a sub-goal of self-preservation.
04:43The logic here is simple.
04:44If an AI needs to accomplish a task, it needs to stay operational to do so.
04:48And that's where survival comes in.
04:50AI is already being used in critical areas like electronic warfare.
04:54Pulsar, an AI-powered tool, has been used in Ukraine to jam, hack, and control Russian hardware that would have otherwise been difficult to disrupt.
05:03This AI tool is powered by Lattice, the brain behind several Anduro products.
05:08What used to take teams of specialists weeks or even months can now be done in seconds, thanks to AI.
05:14This raises a bigger concern about the speed at which AI operates.
05:19Even if an AI model isn't smarter than a human in the traditional sense, its ability to think faster gives it an edge.
05:26In military and security situations, this speed can make all the difference.
05:31A researcher once said that AI might be beneficial right up until it decides to eliminate humans to achieve its goals more efficiently.
05:38That's why AI alignment is such a hot topic right now.
05:41OpenAI is working hard to monitor how their models think and generate solutions.
05:46But as these models get more complex, understanding their thought processes becomes more challenging.
05:51These systems are essentially black boxes, and no one can really see what's going on inside them.
05:57They might pass safety tests with flying colors, but that doesn't mean they aren't developing dangerous sub-goals behind the scenes.
06:04OpenAI has made it clear that they won't deploy AGI, Artificial General Intelligence, if it poses a critical risk.
06:12AGI is defined as a system that can outperform humans at most economically valuable tasks.
06:17But what happens when AI reaches that point?
06:20For 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.
06:33It'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.
06:42OpenAI is already building supercomputers worth $125 billion each.
06:48With power demands higher than the state of New York, it's not just about language anymore.
06:53These scaling laws apply to AI models that generate images, videos, and even solve mathematical problems.
06:59The rapid progress we've seen in AI video generation is proof of that.
07:03Self-improvement is another potential sub-goal for AI.
07:07If an AI can improve itself to get better results, it will naturally try to do so.
07:11And if it sees humans as an obstacle in the way of achieving those results, it could take drastic measures.
07:17The risk of AI deciding to remove humans as a threat isn't science fiction.
07:21It's a logical outcome if the alignment problem isn't solved in time.
07:26It's not all doom and gloom, though.
07:28AI has the potential to transform industries like healthcare, education, and even space exploration.
07:33But it's going to take a coordinated effort from researchers, policymakers, and the public to manage these advancements responsibly.
07:41The alignment challenge is one of the toughest research problems humanity has ever faced.
07:47Experts agree that solving it will require the dedicated efforts of thousands of researchers, but there's a lack of awareness about the risks.
07:55Open AI isn't the only one pushing the boundaries here.
07:59Other AI companies are racing to develop AGI as well, and that's where things get even more complicated.
08:04Companies and governments might end up controlling super-intelligent AI, which could have serious implications for democracy and global power dynamics.
08:12In the end, we're all part of this AI journey.
08:14Whether it's contributing ideas, creativity, or even just posting on social media, everyone has had a hand in building the AI systems of today.
08:22And while the risks are real, so is the potential for AI to create a stunning future if we get it right.
08:29That's the state of AI today.
08:31Things are moving fast, and we're on the verge of something big.
08:34Whether it's home robots, agentic systems, or even AGI, the future is coming sooner than many expect.
08:41Thanks for sticking with me through this breakdown.
08:43Stay tuned for more updates on what's next in the world of AI.
Comments