I can see a number of use cases that will drive very fast user adaption, followed by controversy, followed by even more adoption.
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
I think that AGI is possible within the next decade, even if it's not right around the corner. The rapid progress we've seen from China this year really surpassed my expectations. Just a year ago, the US completely dominated AI LLMs with GPT-4 and 3.5, AI art tools such as DALL-E, Imagen. There was a lot of skepticism about China's AI capabilities by the west, with a lot of trolls claiming that China was decades behind due to GPU restrictions and lack of innovation, and that they could never develop their own LLMs.
Fast-forward to today, and China seems to have caught up in all modalities. Even small models like Qwen2.5-72B are now outperforming the latest GPT-4o in livebench, Qwen2-VL is the top VL model, and QwQ-preview has by far the best price-performance compared to o1-preview or any other model the USA has. There is also DeepSeek R1, Yi, and others competing against the top models from the US. In fields like AI video, I recently counted the top posts from an american subreddit called AI video, and 17 out of the 20 top videos were made using Chinese models. This massive leapfrog acceleration in Chinese AI development shows that progress is not only continuing but gaining momentum in China, potentially with them overtaking the US and driving the entire field forward.
In other areas like self-driving, BYD is about to revolutionize the entire industry by integrating L2+ ADAS in all their car models next year in China. Contrast this with last year, when the US seemed to be leading with companies like Tesla and Waymo. Similarly, in robotics, what looked like a US lead last year has evaporated, and is now seeing massive competition from China, with countless cutting-edge humanoid robots being developed and introduced every month, making China appear to be several years ahead.
A year ago, I was not optimistic about when I'd be able to get a new Huawei device post-ban, but now I'm writing this using a powerful Matepad with Harmony OS and a Kirin 9000 chip manufactured by SMIC, leveraging a robust new domestic supply chain, contrary to the vast majority of earlier predictions by the west. The US thought China couldn't manufacture its own AI chips, but now Ascend 910C and others are starting to compete with Nvidia's chips.
Given these shockingly fast advancements from China, I believe AGI is definitely within reach in the next 5 to 10 years, if China continues accelerating. While it might not happen in the next few years, the exponential progress, especially from China, makes it inevitable in the next decade. The pace at which China is advancing makes me think that AGI is not a matter of if, but when.
Fast-forward to today, and China seems to have caught up in all modalities. Even small models like Qwen2.5-72B are now outperforming the latest GPT-4o in livebench, Qwen2-VL is the top VL model, and QwQ-preview has by far the best price-performance compared to o1-preview or any other model the USA has. There is also DeepSeek R1, Yi, and others competing against the top models from the US. In fields like AI video, I recently counted the top posts from an american subreddit called AI video, and 17 out of the 20 top videos were made using Chinese models. This massive leapfrog acceleration in Chinese AI development shows that progress is not only continuing but gaining momentum in China, potentially with them overtaking the US and driving the entire field forward.
In other areas like self-driving, BYD is about to revolutionize the entire industry by integrating L2+ ADAS in all their car models next year in China. Contrast this with last year, when the US seemed to be leading with companies like Tesla and Waymo. Similarly, in robotics, what looked like a US lead last year has evaporated, and is now seeing massive competition from China, with countless cutting-edge humanoid robots being developed and introduced every month, making China appear to be several years ahead.
A year ago, I was not optimistic about when I'd be able to get a new Huawei device post-ban, but now I'm writing this using a powerful Matepad with Harmony OS and a Kirin 9000 chip manufactured by SMIC, leveraging a robust new domestic supply chain, contrary to the vast majority of earlier predictions by the west. The US thought China couldn't manufacture its own AI chips, but now Ascend 910C and others are starting to compete with Nvidia's chips.
Given these shockingly fast advancements from China, I believe AGI is definitely within reach in the next 5 to 10 years, if China continues accelerating. While it might not happen in the next few years, the exponential progress, especially from China, makes it inevitable in the next decade. The pace at which China is advancing makes me think that AGI is not a matter of if, but when.
18 months or less
This is my intuitive gut feel of it too. I don't work in that industry or bother to read that much into it. My only experience with LLM is in applying it to various work and daily life related tasks. It is a glorified search engine that makes shit up half the time and gets things wrong way too often for it to have any application against anything serious. It actually makes code up too or straight up spits out junk code that simply doesn't work. It isn't even able to generate some routes on a map with coordinate references. Ask it to evaluate a few products and it even gets some specs wrong which you can correct with a simple google search on your own. GPT 4 and GPT 5 at the moment (and probably every other LLM) are meh at best and dangerous at worst.
So my intuition is whatever methods these geniuses are using towards their unicorn (ahem money) chasing journey simply isn't the right approach to creating genuine AGI. All their talk about AGI seems very manipulative when everyone knows they're not only not even close, they're quite literally trying to turn shit into gold by creating and compacting more shit together. Anyway that's my ignorant take. But I should say that some of the tools we get from this work is indeed useful. Narrow AI is revolutionary stuff indeed.
How do you think Chinese companies will monetize their AI technology?
I'd be very surprised if AGI happens in the next ten years. The problem is that the current state of the art is simply fundamentally different than how the only known example of cognition (the brain) works. Transformers are good at ingesting lots of data and spitting out results that sound reasonable, but there is no deeper meaning behind the output. The capability of various AI models today to deliberate, self-reflect, and take action to improve their own performance is extremely limited today beyond various flavours of backpropagation. Our best algorithms still require massive amounts of data to approach human performance where humans make do with a far more limited dataset (with the caveat that human learning is far more holistic than ML). I am skeptical that the direction we are headed in today (stacking more data and more compute onto a big pile of linear algebra) is ever going to result in anything resembling AGI. A more fundamental change in approach is needed.
Consumer products integrating AI, such as AR glasses, autonomous vehicles, smart homes, and consumer robots. Subscription services like Kling and ChatGPT. There are also many opportunities to improve efficiency in manufacturing and industries like healthcare, education, agriculture, smart cities, retail, and entertainment.
Scaling law extrapolation never made sense since dataset size can't be scaled indefinitely, there are differences in dataset quality, and fundamentally all known intelligence, including animals, don't follow scaling law.
American obsession is very much a combination of religion and greed, two of their debufs as a society, and moral hazard from quantitative easing just added fuel to the fire.
IMO anyone who wants to build a mini nuclear reactor to learn how to do grade 2 math need to rethink their life.