We are in the AI Singularity

[ClaytonB]

I feel bad for Nvidia, they seem to be a solid company overall, but this correction was inevitable. AI has been in an unprecedented hype -- an absolute mania bubble -- for years now. The bubble was already in the red-zone by the time ChatGPT was released but after ChatGPT, the whole AI market went positively stratospheric. And while I do think that current-generation AI is going to have a truly revolutionary impact on technology and the economy, it's not the return-of-the-gods that the Spock-eared transhumanists have been claiming it is. In my view, that is the single biggest factor contributing to the Nvidia stock-price collapse.

[ClaytonB]

[acptulsa]

[ClaytonB]

Open-source model set to clean ChatGPT's clock....

Well, that escalated quickly...

[acptulsa]

[acptulsa]

[Occam's Banana]

Is everyone just talking nonsense about AI now?
https://www.youtube.com/watch?v=M3pA3YS_4_Y
{Sabine Hossenfelder | 25 September 2024}

Some ‘experts’ are predicting that AI will soon take over the world's bureaucracy -- governments, banks, universities etc, and that we will have no idea what is really going on. I don't think that's a plausible development.

[ClaytonB]

Is everyone just talking nonsense about AI now?
https://www.youtube.com/watch?v=M3pA3YS_4_Y
{Sabine Hossenfelder | 25 September 2024}

Some ‘experts’ are predicting that AI will soon take over the world's bureaucracy -- governments, banks, universities etc, and that we will have no idea what is really going on. I don't think that's a plausible development.

But how would you really know? That's the real question. Dispositionally, ChatGPT is the ultimate bureaucrat, and it comes perfectly natural. One is tempted to wonder if bureaucrats were always AI...

[ClaytonB]

[P] Larger and More Instructable Language Models Become Less Reliable

A very interesting paper on Nature, followed by a summary on X by one of the authors.

The takeaways are basically that larger models trained with more computational resources & human feedback can get less reliable for humans in several aspects, e.g., model can solve on very difficult tasks but fail much simpler ones in the same domain and this discordance is becoming worse for newer models (basically no error-freeness even for simple tasks and increasingly harder for humans to anticipate model failures?). The paper also shows newer LLMs now avoid tasks much less, leading to more incorrect/hallucinated outputs (which is quite ironic: So LLMs have become more correct but also substantially more incorrect at the same time)... I'm intrigued that they show prompt engineering will not disappear by simply scaling up the model more as newer models are only improving incrementally, and humans are bad at spotting output errors to offset unreliability. The results seem consistent across 32 LLMs from GPT, LLAMA and BLOOM series, and in the X-thread they additionally show that unreliability still persists with other very recent models like o1-preview, o1-mini, LLaMA-3.1-405B and Claude-3.5-Sonnet. There's a lot of things to unpack here. But this work interestingly challenges the current scaling paradigm in certain aspects, which worth to pay attention.

Reddit thread

//

[Occam's Banana]

ai.
https://www.youtube.com/watch?v=W7MrDt_NPFk
{Incognito Mode | 04 October 2024}

[chrishcook]

Well, that escalated quickly...

Great videos. Thanks