LLM

I often test new LLM models with the following question — Given the following conditions, how many ways can Professor Y assign six different books to four different students? The most expensive book must be assigned to student X. Each student must receive at least one book. The first model that could solve this problem was OpenAI’s o1 model. The second one was Google’s Gemini 2.0 Flash Thinking model. Then more thinking models came out, like DeepSeek’s R1, Anthropic’s Claude 3.7 (with extended thinking mode), all of which can solve this problem without any mistakes. ...

I recently learned, from the ever-illuminating Simon Willison, that Google has released a new model, gemini-2-0-flash-thinking-exp. So, naturally, I posed it a question, a little puzzle if you will. Given the following conditions, how many ways can Professor Y assign six different books to four different students? The most expensive book must be assigned to student X. Each student must receive at least one book. It took the model some time, a moment of digital contemplation, to complete the task. But it did, correctly yielding the answer 390. So now, alongside gpt-o1, we have another model capable of solving this combinatorial problem, a problem that, it must be said, eludes most of my students. It seems that these “inference scaling” based models will be able to solve more such problems, each one a small victory for the silicon mind. Another reason, perhaps, why I no longer set homework and only give quizzes, a small act of surrender in the face of the inevitable. ...

I planned to give a talk on AI’s application at a local cafe. Unfortunately there was not a single one who signed up. In order to not to waste my entire effort in preparing for the talk, I translated my slides into a blog post. You can check out my original slides (in Chinese) here. 😨 Will AI Threaten Humanity? 🤖 What is AI? Artificial Intelligence (AI) is a broad term encompassing various technologies. Here are some key areas: ...

For some random reason, I decided to test whether ChatGPT and Claude could write poems in an ancient Chinese style. So, I sent Claude 18 poems by Tang Yin (唐寅, 6 March 1470 – 7 January 1524) and asked it to compose three new ones in Tang’s style. The initial results were impressive but sounded a bit modern, so I asked ChatGPT to revise them. Afterward, I gave them back to Claude, asking it to change any words that sounded modern. Below are four poems: one by Tang Yin and three by AI. If you can guess which is which without Googling, I would say you are an expert in ancient Chinese poetry. ...

I have been trying to use the Large Language Model (LLM) Claude to generate artifacts (apps written in JavaScript and HTML) to visualize algorithms for my COMPSCI 308 course. It works amazingly well. If I had to write them myself, it would have taken hours. It also has the added benefit that I can run them in a browser on my iPad, which I use for teaching. Previously the only way I found to run code on iPad is to use Wolfram Player, which is quite buggy. ...