Silicon Valley products, a rival to DeepSeek’s AI model, are excellent in some complicated tasks. The model uses inference time calculation to divide the query into a smaller and managed task.
China’s AI Lab DeepSeek has recently released an AI model that matches or exceeds the top offering of Silicon Valley.
DeepSeek uses an approach called test time or inference time calculation. This slices the query into a small task and turns it into a new prompt for each model. In each step, you need to execute a new request known as the AI inference stage.
Business Insider has recently tested one of the DeepSeek models using Deepshink mode. This shows all steps in the thinking process.
We have given the multiple mathematics issues proposed by Charlie Snell, a UC Berkeley AI researcher.
Snell stated that DeepSeek works well on complex math issues. Researchers chose a problem from the US Invitation Mathematics Examination. This is a challenging test of the crying of high school mathematics.
“I asked the DeepSeek model from the question,” Snell told BI in an interview. “I read the chain of thinking. I understand that.”
DeepSeek demonstration
The mathematics issues proposed to BI’s DeepSeek demonstration are as follows: ” +, -, /, *find the sequence. Just once.”
BI put the prompt in the website’s Deepseek chat window. This model responded by first layout issues.
“Now, there is this problem here. It is necessary to use numbers 7, 3, 11, and 5 to combine each number of values with addition, subtraction, multiplication, and division operation. It will be 24.” I answered. “At first glance, this seems a little difficult, but I think that you can understand it with systematic thinking.”
After that, we took multiple steps beyond about 16 pages, including mathematics and equations. The model made it wrong, but found it and didn’t give up. Instead, I moved quickly to try another possible solution, then another solution.
At one point, “I got closer to 33 /7 *5≈23.57, but not 24. You may need to try another approach.”
Later, the DeepSeek model seemed to have repeated potential solutions.
“Wait, I have already done it,” he wrote. “Well, maybe I need to consider using the split in another way.”
A few minutes later, a correct answer was found.
“You can try various ideas and try backtracks,” said Snell. He emphasized that this part of the chain of Deepseek’s thinking was particularly worth noting.
“This really takes time. Maybe I need to consider another strategy,” says the AI model. “Instead of combining two figures at a time, you must probably find a different way to group them, or use the operation in nested methods.”
