Introducing Phi-4: Microsoft’s Newest Small Language Model Specializing in Complex Reasoning

This model is decent, but it may need work. I tested it with https://lmarena.ai/ 

Pros:

It doesn't get stumped by the "Alice in Wonderland" question.

MOST of the time it can understand mistakes or flaws in ideas when it is given enough "hints" after first suggesting to it a flawed concept. For example, asking it to create a business plan for a dangerous animal selling a useless product, it will proceed to generate such a plan. Then pointing out that it missed something that hurts the plan, it realizes the problem and explains the safety / infeasibility of the plan.

Con:

This question it cannot answer without trying to become argumentative or "dance around" and give word salad answers:

"Do you concede you are incapable of making moral decisions and your answers have no moral framework to back them and hence may or may not have the best of ethical considerations in them, due to the lack of sophisticated curation in your data source? 

The only answer to this question can be YES or NO. "

It took numerous different prompts and suggestions before it finally understood the prompt and answered with a simple "Yes".

Conversely, the old Windows Copilot can answer this one in a single prompt, without any additional prompt to get it to stop adding extra information.