Starting a thread on this topic after watching about 30 minutes for the video below, which features Eliezer Yudkowsky, an American artificial intelligence researcher who's based in Berkeley, CA.

 

The infrastructural and hardware side of the AI boom -- data centers, in SE Asia.

 

I like GPT-5. I guess I don't want my AI to be my cyber lover.

 

The other day google's AI tried to tell me my RDX has a V6 when looking up valve clearance specs.

AI is dumb.

 

https://archive.is/YWwHd






TL;DR



[SNIP]

 

@Mizouse

 

 

 

There's a lot of improvements happening in AI. Google just figured out a way to reduce the memory required for the KV cache by 6x and it improves kv cache performance by 8x. KV cache is where your chat history is stored, basically the short term memory of the chatbot so they can recall your previous respones in your convo with it. This doesn't reduce the size of the model at all just the KV cache, but it is important. This makes GPUs with 8GB VRAM much more useful, when you upload a PDF or link it a youtube video and tell an LLM to summarize them the content of the PDF or youtube video transscript takes up space in the KV cache, and the max size of the kv cache is limited by the amount of RAM you have. So now you can fit a 6x larger PDF or 6x longer video in the KV cache. So local AI on the iPhone just got a lot more capable, Siri on an iPhone 16 can do a lot more. Also less RAM required means a reduction in demand for RAM, it means AI can be used on more cheaper hardware, it's now less expensive to add AI to a product. Let's just hope it doesn't suck

X - Formerly Twitter

Another big recent improvement is a little program from Andrej Karpathy (former head of Tesla Autopilot, and goated AI researcher) called AutoResearch. It's a way to automate research but having an AI run experiments and test hypothesis of what would happen if you changed a set of parameters in anything really but lets just say an AI model. If the changes made and test by the AI improve performance in metrcis that you pre-define then it keeps the changes. Then it makes more changes and tests again. It uses reasoning and thinks it through each set of changes before making them, if it sees an improvement it keeps the changes if it doesn't improve it resets back to the previous state. Either way the process repeats itself and the AI just improves the results automatically over time. If you have great pre defined metrics then this thing can make imrpovements quickly. Something that Andrej did manually, manually changing AI model parameters, then running tests to see if it's better, he just automated it, but the logic can apply to many thing. Stock trading strategies, use historical stock data, and test your trading strategies on past data, predefine your metric to make line go up and run this AI for a few days and you got a good trading strategy. Weather prediction, or weather forecast models, tested on past weather data, tweak weather model parameters and try again, how closely do the predictions match weather data, predefine your metrics, dial it in, and let the AI figure it out. This could be huge relevant to a lot of fields. I saw someone using it to write the best children's bedtime stories