Artificial Intelligence thread

[name]

This approach will fail to reach AGI. They are matching patterns. It's like someone who copies homework without understanding. If homework is available to copy (AKA within distribution), they can often parrot decent answers that others provided. Without homework to copy (AKA outside of distribution), they give wrong answers eg bullshit. The problem is there's no amount of data to cover every single permutation. Where's the homework for 244r234234134123421323534534532498754897234897234412938471289472174 + 3 ?

Still very impressive but too unreliable to ever be AGI.

see more here

Please, Log in or Register to view URLs content!

Please, Log in or Register to view URLs content!

for those not logged in

garymarcus substack com/p/confirmed-llms-have-indeed-reached
garymarcus substack com/p/llms-dont-do-formal-reasoning-and

 

[fatzergling]

Natural Language is wildly inefficient for computer systems reasoning

Have a look at this new Meta paper. I think that fundamentally this is the way forward, with the disadvantage that the AI systems will become a complete black box for humans

Please, Log in or Register to view URLs content!

https://twitter.com/i/web/status/1866889287851253880

I was just looking through this example and I can't believe they are using LLM's to traverse a graph. Surely there are more efficient ways, like compile the graph and then perform some graph algorithm over it?

https://twitter.com/i/web/status/1870706778410983713

Put my thoughts on here

basically as it stands, o3 is completely unusable even if cost comes down to $10 per prompt.

and it’s only solving the hard problems because it got people that can solve them to first train the models. Well, can it get model to train against every menial task it will be expected to automate away?

If you iterate every possible combination of answers, you are bound to find the correct answer. It might take you 100 years though.

 

[tphuang]

my point is that what is "AGI" good for? Why is it important to have something that appears to be AGI to end users? It's to solve real world problems and such.

If the reasoning models are only able to solve complex problems because they are gamed with training from hired PHDs to solve fixed set of problems, are they smart enough to give correct answer for more mundane questions 99% of time?

Hallucination is a huge problem. These evaluations that measure scores. Are they testing answer consistency? Getting easy or medium difficulty questions wrong even 0.5% of time is not really acceptable if you want to automate away human workers.

 

[tphuang]

https://twitter.com/i/web/status/1870936955330216349

Honor have now teamed up with Tencent for AI backend solutions and such. This is after it already picked ByteDance's VolCEngine for Doubao models. And I'm sure it's own team is also developing diff AI solutions.

key is how you put everything together to end user.

 

[tphuang]

longer term, I think open source models will win. There is no data security concerns. You can't blame anyone but yourself when there is outages. you can run everything locally with primary and backup data centers. You can modify it and retrain it to use your data center as needed. If speed is not good enough, just increase computation power.

And believe it or not, when you are pitching your product to customers or investors, they like to hear that you have your own internal model solution. That you are not reliant on API access to OpenAI.

 

[tokenanalyst]

Huawei taking over the CUDA Moat​

The rise of China's AI framework: Huawei MindSpore claims 30% market share​

In 2024, China's artificial intelligence (AI) framework market is experiencing robust growth, with Huawei's MindSpore securing nearly 30% market share, successfully positioning itself among the global mainstream AI frameworks.

Please, Log in or Register to view URLs content!

 

[9dashline]

Huawei taking over the CUDA Moat​

The rise of China's AI framework: Huawei MindSpore claims 30% market share​

In 2024, China's artificial intelligence (AI) framework market is experiencing robust growth, with Huawei's MindSpore securing nearly 30% market share, successfully positioning itself among the global mainstream AI frameworks.

Please, Log in or Register to view URLs content!

China also needs the eq. of TensorFlow/PyTorch etc

 

[tphuang]

China also needs the eq. of TensorFlow/PyTorch etc

Mindspore is PyTorch for huawei.

Nobody uses tensor flow anymore.

 

[9dashline]

Please, Log in or Register to view URLs content!

Please, Log in or Register to view URLs content!

 

[tphuang]

https://twitter.com/i/web/status/1871662368796860614

btw, this actually is helpful. Getting GPT to read financial docs that are scanned and have photos and tables in them is hard. If this is something that can be morphed into reading PDFs, XLS, images, PPTs and word documents (or anything containing images/tables), it would be quite powerful.