A research team at Berkeley said they

A research team at Berkeley said they reproduced the basic version of DeepSeek R1-Zero for $30 so it was noisy. Using the DeepSeek R1-Zero algorithm, we presented the purpose and method of problem solving, and asked AI to find a way to solve it.

The problem they were trying to solve is a math problem presented by the British game show Countdown. It's a quiz game that randomly selects six numbers, adds, subtracts, multiplies, and divides them to become the target number, as shown in the figure below.

The bottom line is: It just works. The basic model used Qwen-2.5-Base model provided by Alibaba, which had different performance depending on the parameters used in the model (AI size), but the 500 million parameter model didn't work well, and the 1.5 billion parameter found the solution well on its own.

... The cost of this is about 30 dollars. Hahaha.

I wondered where to use this... but I thought there would be some work to find a solution in software. For example, at Ion Mall in Japan, there is a place where AI is entrusted with the sale price before the deadline?

This has been decided by senior employees, but since leaving it to AI, sales have increased and employees' mental burden has been reduced. Naturally, fewer products have been discarded.

Basic data should be prepared, but I think finding solutions using AI that has already been disclosed to these various problems can be widespread. Nowadays, more and more places are marked with electronic ink, so you can mark the discount price right away.

... And if you can connect this to the AI broadcasting system and automatically broadcast it like, "From now on, I'll sell beef at up to 30% off at the butcher's store"... Hmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmm.

Anyway, the AI that I see is more a function than a product. It shines when it is integrated into the work process (software) rather than doing everything on its own, just like program coding does the basics.

And I think deep shock can help when integrating AI functionally like this.

+ The problem is that preparing basic data is not as easy as you think...

+ I think of the countdown game as a counter strike game, and the trap of thinking, "What problem should I solve here?" I had a hard time saying the wrong name of this game during the crowd strike.

+ Reject 'Do something with AI' because AI is in vogue without thinking about a problem to solve. Try it while doing this and that for fun with AI. I don't know if it's 'Do something with AI'.