With AGI it should be able to do inference on outside of distribution data. I would argue that the current performance of our best model has seen the test data somewhere in their training set - Data leakage.
I would love to hear from you on my statement! (4/4)

We optimized the model with a gradient descent methodology minimizing a selected loss function that should give us optimal weights models for that specific dataset, and that generalize well on the validation set - Similar data distribution. (3/4)

We are using generative AI algorithms trained on as many good quality data as possible. So in theory the model has as much information of the world as it has been trained on - Which is a lot! (2/4)

Here is the official Artificial Intelligence for the Earth Systems journal link!

journals.ametsoc.org/view/journal...

We should create a new research direction - Maximizing researcher contributions to humanity.

I’m sure lot of great researchers have awesome ideas for that!

I think Canada needs a clearer definition of what “top talent” means before prioritizing recruitment from abroad.
Many of us already here are struggling to find opportunities - not because of a lack of talent.
Policies like this only push young researchers to leave and do great research elsewhere.

As a beginners myself, I highly recommend this French video from one of the most exceptional professors in the domain to get started with QC: youtu.be/XkGaitu3EbI?...

I take a lot of pride in this paper since it was done as an undergraduate student-Nothing is impossible with hard work.

I love research, and I truly do it out of pure passion. I feel alive when collaborating with highly talented people and being around those who understand and challenge my ideas.

3.We showed that a model trained only with the predictand achieved similar performance to a model trained with multiple predictors, challenging prevailing beliefs in our field.

2.We created a model that incorporates learned interpolation directly within the network instead of performing it beforehand. By reducing the model inputs by a factor of 16 (4× downscaling), we showed that learned interpolation achieves performance equal to that of traditional interpolation.

We decided to challenge these ideas by:
1.Comparing models trained with multiple sparse and nearby domains. We showed that using sparse domains maintains performance equal-better than that of nearby domains.
We also demonstrated that adding more domains keeps performance unchanged or improves it.

I will never retire from research.
I truly do it for fun, and I would even pay just to do it with talented people.

The US lost a lot of credibility in my eyes.