yangbo

The overall assessment is basically saying: "I didn't read it (for user's sake to save tokens). The user wants a critical response, so here it is: I worried about xxx and xxx because it looks like xxx."

I found LLM sometimes tends to not read through in order to save tokens.

Try this prompt:

> Can you READ THROUGH the paper and give me a critique of your previous critique?

Did you ask Claude's own opinion instead of the opinion of the role it plays?

The journal editor should run a language model to detect if the paper includes some placeholder or instruction for a language model. Unfortunately they did not.

I am a developer from Gradient. I think the most interesting use case of our product is interactive few-shot fine-tuning in a Jupyter Notebook. When fine-tuning a model with less than 10 samples and around 3 epoch, it can learn knowledge in a couple of seconds, then you can immediately get the fine-tuned model.

The experience is like to be your model's private tutor. It just continuously learns new things you told it, like a life.

Be cautious about your learning rate, or it will go mad.

Hi all, I created a library to perform tail call optimization in JavaScript. The idea is to create proxies backed by trampolines, so that it is transparent from the caller's point of view. What do you think?

Even though, the number of React's API is still larger than [Binding.scala's API](https://static.javadoc.io/com.thoughtworks.binding/binding_2...).

Given a component that contains n elements, and you want to update k elements at a time, ReactJS requires O(n) time for virtual DOM differentiation, and Binding.scala only takes O(k) time.

As a result, Binding.scala is fast when k < n (aka partial updating), even faster than vanillajs written by a common human for some cases. Check the rows "partial update", "select row", "swap rows", "remove row" in the benchmark.