Artificial Intelligence thread

[tphuang]

bilibili has open sourced a 1.9B parameter model trained on 2.8T tokens

Please, Log in or Register to view URLs content!

results from various LLMs doing Gaokao showing Qwen 72B parameter LLM being even better than GPT-4o

InternLM2-20B also did really well for having only 20B parameters

Please, Log in or Register to view URLs content!

 

[GulfLander]

https://twitter.com/i/web/status/1802887337329987912

https://twitter.com/i/web/status/1803231224342913175

 

[Eventine]

LLMs aggregate common knowledge on the internet which is not going to be sufficient to pass mathematics competitions designed to challenge the 0.000001% of humanity. You’ll need different techniques than that.

 

[tygyg1111]

bilibili has open sourced a 1.9B parameter model trained on 2.8T tokens

Please, Log in or Register to view URLs content!

results from various LLMs doing Gaokao showing Qwen 72B parameter LLM being even better than GPT-4o

InternLM2-20B also did really well for having only 20B parameters

Please, Log in or Register to view URLs content!

English article from Yicai for non-Chinese readers:

Please, Log in or Register to view URLs content!

Goes to show humanities majors can now be replaced by LLMs:

 

[ougoah]

lol what a surprise GPT-n cannot beat humans in exams.

People using GPT for writing assignments is one thing (easy to tell and usually middle of the road quality anyway) but performing anything that requires intelligence like maths? Ha

Go ahead and give all the LLMs another trillion parameters. I personally don't believe that will refine their abilities much further as diminishing returns have long been reached. The Chinese LLMs beat the American ones in parameters in almost high profile example. Only marginally better in maths benchmarking.

 

[luminary]

https://twitter.com/i/web/status/1798625526795628790

Please, Log in or Register to view URLs content!

Hype for investors vs an actual service:

unlike Sora, which still remains inaccessible to the public four months after OpenAI trialed it, Kling lets people try the model themselves.

Also, I think the multimodal approach isn't going to work if you are cobbling crappy models or architectures together. The only thing a LLM is going to add to a humanoid robot is make it more unreliable. That counts out industrial or military applications. Generative video AI doesn't seem to contribute to perception algos either.


Finally, here is my repost and daily reminder that "emergent capabilities" hasn't been a thing for a while now and that LLMs are not going to turn into AGI, in case that was still a question:

Please, Log in or Register to view URLs content!

Please, Log in or Register to view URLs content!

by a trio of researchers at Stanford University show that the sudden appearance of these abilities is just a consequence of the way researchers measure the LLM’s performance.

The original OpenAI researcher

Please, Log in or Register to view URLs content!

s that properties of emergence are unimportant because for abilities like arithmetic, the right answer really is all that matters.

 

[gelgoog]

They are basically an assistant that provides collages of pre-existing information on demand. That is all they do really.
I consider this to be the next generation of search. Which is no small thing.

 

[ougoah]

They are basically an assistant that provides collages of pre-existing information on demand. That is all they do really.
I consider this to be the next generation of search. Which is no small thing.

You forgot to mention they do this unreliably and inaccurately.

It is a better version of search though, much more multi-dimensional and in no small way, a revolutionary piece of technology branch. It's just not AGI, not even close. But we have industry leaders often dropping the G word just to get more hype and funding into this space.

Chinese LLMs are measurably better in many ways (parameters and benchmarking), certainly if you are in the Chinese cultural and internet ecosystem. It's just rare for Chinese AI scientists working in China to consider this avenue as a sure path to AGI. Even the Chinese AI scientists working in the US (and about a quarter of the US AI industry is Chinese) don't claim this but the big wigs (management not engineers) always do. I wonder why.

 

[Randomuser]

You know this video makes me realize why western analyst get China wrong a lot. The scenarios US and China are completely different. China focuses on industry and manufacturing so it has data there. Therefore its AI is geared towards that stuff. So people ask why isn't China following the same path as western AI firms, its because China has a completely different path. And in the long run having AI focusing on industry and manufacturing seems more productive than western ones used to kill off artists and designers. Maybe its more boring and not flashy but plans should not be designed about how eye appealing they are.

 

[tphuang]

Please, Log in or Register to view URLs content!

Pangu 5.0 has been launched with a whole host of different sized models and applications.

Huawei really focused on industrial usage