few weeks back, I presented DSPy to a group of my research colleagues, with some cool demos from online. Today, I am writing my research with DSPy & co-coatching an undergrad on research with DSPy! Once you try DSPy, you will never go back! Thanks Omar Khattab

keep in mind that LLMs favor highly probable tokens, which will have higher density in a give distribution (i.e. the internet). LLMs are, by design, unable to produce originality!

Apparantly, with the current landscape of AI coding agents, it is becoming clear that code is a major market of LLMs vendors. Why? Code: - can build products & genertae money. - is well structreded, verbose, redundant -> has less entropy (easy to learn). - is simple by design &

When working on LLMs on text modality, I've found it helpful to think about four categories of tasks, each with unique characteristics and patterns: Text Operations (TextIO): These tasks transform input text into a different format while preserving core information. Examples

Another way to verify the LLM cutoff date: Ask it to pin a Python requirements file without searching the web.

Now I am further realizing that getting tired and lazy is a bless, not a blast! These feelings force to innovate shortcuts and respond only with what really matters. LLMs should learn this skill!

It is interesting that the launch of GPT-5 is giving Meta some indirect credit!

Yesterday, I received the notification of acceptance of the two works I co-authored for #EMNLP2025. Truly delighted, as this contribution marks my third contribution in a row to the prestigious #EMNLP conference. My EMNLP2023 contribution was this paper: Consonant is all you