Joined January 2025
- Tweets 10
- Following 6
- Followers 2
- Likes 71
2 Photos and videos
After prolonged negotiations with softmax temperatures and causal semantics, Input-Adaptive Hard-Exit DARTS concluded that unnecessary FLOPs are a skill issue.
Reviewer #2, your move.
#AIResearch #DeepLearning #NAS
14
Jun 13
DARTS finally learned where the Exit button is đŞđ¤
CIFAR-10: cost_norm 0.991â0.772 (~23% less compute), test acc 0.862â0.854.
Hard-exit now actually exits (not just a layer-4 fan club đ).
Paper: sameerresearchai.github.io/aâŚ
Code: github.com/sameerresearchai/âŚ
30
Jun 13
Cloudflare hid the taxes better than I hide API keys.
Jun 12
@Cloudflare
Your domain checkout page needs some real transparency.
Just bought a new domain. It showed $26.00 throughout the entire process. Got charged $30.68. The extra $4.68 in taxes was never mentioned once during checkout. Only found out via the invoice email.
Please show the final all-in price (taxes included) upfront.
It's a small change that greatly improves customer trust. Fix this.
1
37
Jun 12
Me: "2000 GPU days on RL-based NAS."
DARTS: "Hold my gradient descent. Differentiable search in 4 days."
Bilevel optimization: weights on training loss, architecture on validation loss. More math, less crying.
sameerresearchai.github.io/aâŚ
40
May 31
My neural network performed a harmful architectural modification, detected the regression, rolled back to a previous stable state, and resumed optimization.
Meanwhile, I'm still overfitting on a conversation from 2017.
github.com/sameerresearchai/âŚ
#AI #MachineLearning
54
May 29
A PyTorch model exploring adaptive architectures:
⢠learns sparse connectivity
⢠adapts weights from activity signals
⢠prunes & grows neurons dynamically
Basically gradient descent with a mild obsession for reorganizing itself mid-life crisis.
46
May 24
Perceptron loss optimizes separability.
Hinge loss optimizes separability with margin maximization.
That tiny â 1â quietly converts a classifier from âbarely rightâ to statistically robust.
Inductive bias hidden in one constant.
#MachineLearning #AIResearch
50
May 24
Reproduced âAttention Is All You Needâ from scratch.
The Transformer achieved convergence.
I achieved gradient instability.
colab.research.google.com/drâŚ
#Transformer #AttentionIsAllYouNeed #NLP #PyTorch
57
May 23
What if intelligence isnât âunderstandingâ at all?
What if itâs just an extremely efficient latent-space compression and reconstruction of knowledge patterns?
And what if human intuition emerges the same way?
42
May 21
Maybe intelligence isnât hidden in larger models.
Maybe it begins the moment a neuron can change its own computation.
43