d/d

AI

The AI category covers machine learning engineering and practical applications of artificial intelligence. It focuses on understanding how models work and deploying them in real-world contexts:

  • Model Architecture: How neural networks are designed, trained, and optimized.
    • Topics: Transformers, quantization, inference frameworks, scaling laws.
  • Edge AI: Running models on consumer hardware, browsers, and mobile devices.
    • Topics: bitnet.cpp, WebGPU inference, on-device fine-tuning, memory optimization.
See All Categories
AI

August 2026

What Happens to the Frontend When Agents Use the Web?
August 15, 2026

What Happens to the Frontend When Agents Use the Web?

Browser Run lets artificial intelligence agents operate the web we already have. WebMCP gives websites another option: expose their actions directly.

Read
What a Feed-Forward Network Does Inside a Transformer
August 13, 2026

What a Feed-Forward Network Does Inside a Transformer

Follow one contextualized token through expansion, activation, contraction, and a residual path, then see why Transformer FFNs hold so many parameters.

Read
How a 120B Mixture of Experts Activates 5.1B Parameters
August 12, 2026

How a 120B Mixture of Experts Activates 5.1B Parameters

A token-level look at expert routing, sparse computation, load balancing, and why fewer active parameters do not remove memory and communication costs.

Read

July 2026

Sparse Attention Guide: What Should an LLM Ignore?
July 4, 2026

Sparse Attention Guide: What Should an LLM Ignore?

Sparse attention cuts full attention's n^2 graph, but every missing edge changes what an LLM can remember, route, cache, and safely ignore in long contexts.

Read

May 2026

The Phase Shift Attack: How Fourier Math Strips an AI Watermark
May 9, 2026

The Phase Shift Attack: How Fourier Math Strips an AI Watermark

How a Fourier transform exposes an AI watermark's frequencies, and how a phase shift drops SynthID detection by 91% at 43 dB PSNR.

Read
Talkie-1930: An LLM That Has Never Heard of Computers Just Learned to Code
May 2, 2026

Talkie-1930: An LLM That Has Never Heard of Computers Just Learned to Code

Talkie-1930 is a 13B language model trained only on text published before 1931. It has never seen a computer. Given a few Python examples, it writes code anyway. Here is what that means for the reasoning vs memorization debate.

Read

April 2026

Run a 1-Bit LLM on Your Mac with bitnet.cpp
April 13, 2026

Run a 1-Bit LLM on Your Mac with bitnet.cpp

A first-person walkthrough of running BitNet b1.58 on an M1 Pro: real benchmarks (22.5 tok/s), five real gotchas, a 3B model in under 1 GB of RAM, and the GPU never touched.

Read
How SynthID Was Broken: Three Attacks That Defeated Google's AI Watermark
April 12, 2026

How SynthID Was Broken: Three Attacks That Defeated Google's AI Watermark

Researchers broke Google's SynthID watermark 3 ways. Spectral analysis drops detection by 91.4%, diffusion re-nosing overwrites it, and synonym swaps defeat text marking.

Read
BitNet b1.58: The 1-Bit LLM That Matches Full-Precision Models
April 9, 2026

BitNet b1.58: The 1-Bit LLM That Matches Full-Precision Models

BitNet b1.58 replaces every Transformer weight with -1, 0, or 1, cutting energy use by 71.4x while matching FP16 quality at 3B+ parameters. Here is how.

Read