novice ai researcher & engineer
Joined August 2025
- Tweets 24
- Following 110
- Followers 1
- Likes 1
Photos and videos
17 Aug 2025
Today I learned about git rebase, here is a mini-thread about it:
1
7
17 Aug 2025
Why rebase?
- Cleaner history
- Easier to read logs
- Looks like your work was always based on the newest main
1
6
17 Aug 2025
But be careful!!! rebase rewrites history.
Never rebase shared branches.
Safe rule: only rebase your own local branches.
6
12 Aug 2025
Ok, so I get there are two main approaches for solving RL problem - value-based method and policy-based method. But is there a rule of thumb for choosing between them?
1
2
12 Aug 2025
Let me explain the attention mechanism in Transformers really simply.
Let's say there is this sentence:
"The animal didn't cross the street because it was too tired."
We want the model to focus on "the animal" when processing "it" - like humans do.
1
2
12 Aug 2025
Finally, we use these weights to combine the Value vectors from all words. This gives "it" a context vector - a smart blend of relevant info.
In our example, the weight for "animal" will be the largest, so most of the context vector for "it" comes from "animal’s" Value.
1
12 Aug 2025
And that’s attention:
1. Q asks the question
2. K says “this is what I am”
3. dot product of Q and K gives a score
4. V provides the answer
5. Softmax decides who to listen to most
11 Aug 2025
huggingface deep rl course is awesome and I am going to go through them for the next few days :)
huggingface.co/learn/deep-rl…
6
11 Aug 2025
Why do we need value function when we can just try to maximize the reward?
tl;dr
- unfairness: someone who worked hard and improved a lot must not be penalized because of low absolute reward
- instability: we should aim for stable 90 , not one time 100 with extreme strategy
1
3