Latest posts

AI productivity pressure floods teams with unreviewed merge requests. How to push back and spend the time on deeper, better work instead.

Should you read the code your AI agents write? Yes. Why the hard part of software was never only the code, and why agents still need review.

Complex-valued versus real-valued neural networks on modulation classification and RF fingerprinting, with results and when complex layers help.

Complex numbers for deep learning, with interactive visuals: polar form, rotation, holomorphic functions, and Wirtinger calculus for backprop.

IEEE 754 floats explained, then TF32, BF16, FP16, FP8, and FP4 for deep learning: bit layouts, ranges, GPU support, and when to use each.

AI agents do not make you a 10x engineer. Why vibe-coded software breaks, why the productivity gains are overstated, and why skills still matter.

Good code versus bad code is the wrong question. Code is a vehicle for solving problems, shown through API, debugging, and memory layout examples.

Training an LLM to understand images with an MLP projection, Q-Former, and Perceiver Resampler, with attention maps and inference comparisons.

How multi-modal models understand images and audio: projection layers, cross attention, and VQ-GAN tokens, with their tradeoffs.

What a good PyTorch checkpoint saves: model, optimizer, scheduler, GradScaler, and RNG states, plus tracking checkpoints with MLflow.

AI slop, layoffs, and broken interviews in the tech industry, and practical advice for developers and students on how to get through it.

Write and launch CUDA kernels from Python with NVIDIA cuda-python, using both cuda.bindings and cuda.core, with a Gaussian blur and timing results.

Run Python on the GPU with CuPy and Numba. Benchmarks of CuPy against NumPy and SciPy, and writing custom CUDA kernels with Numba.

What batch normalization does, why internal covariate shift is not the real reason it works, and how smoother loss landscapes explain its benefits.

Models break in production when pre- and post-processing drift. How to save processing steps with checkpoints, TorchScript, and baked-in pipelines.

Binding C and C++ to Python with the C API, ctypes, CFFI, SWIG, Boost.Python, and pybind11, with timing results and when each is worth it.

Debug neural network training by probing gradients: detect vanishing and exploding gradients, measure gradient signal-to-noise ratio and variance.

How to visualize the loss landscape of a neural network in PyTorch, compare initialized and trained weights, and track the loss path during training.

The 4 pillars of OOP are not unique to OOP. Encapsulation, inheritance, polymorphism, and abstraction compared with C, composition, and ECS.

Handling imbalanced datasets with undersampling, oversampling, SMOTE, and combined methods, using a credit card fraud dataset and imbalanced-learn.

Deep abstraction layers make code hard to follow. How to keep program control flow readable, testable, and easy for teammates to change.

Why this blog exists: the gap between ML research and real software, why ML work is hard to estimate, and the topics Gradient Rage will cover.