Prompt Injection and LLM Security
A deep technical report on the attack surface of LLM applications, direct and indirect injections, real-world vulnerabilities, and architectural defenses.
Sep 2026
Artifacts
An artifact is what remains after the thinking stops — a question worth sitting with, reduced to the few numbers that made it unavoidable.
RLHF, DPO, and GRPO: The Alignment Method Landscape
A deep technical comparison of the three dominant post-training alignment methods for LLMs, their trade-offs, and the shift toward verifiable rewards.
Aug 2026
The KV Cache Bottleneck
How PagedAttention eliminated 80% memory waste and unlocked continuous batching to multiply LLM inference throughput.
Aug 2026
Model Collapse and the Synthetic Data Frontier
How recursive training on AI-generated data leads to model collapse, and the metrics, mechanisms, and ratios required to prevent homogenization.
Aug 2026
Constitutional AI: The Self-Critique Alignment Paradigm
How AI systems use explicit principles and self-correction to bypass the human feedback bottleneck.
Aug 2026
Speculative Decoding
How drafting architectures overcome memory bandwidth bottlenecks to accelerate LLM inference, shifting from autoregression to parallel tree generation.
Aug 2026
KV Cache and the Memory Bottleneck
How PagedAttention, continuous batching, and prefix caching solved the 60 percent memory fragmentation problem in LLM inference.
Aug 2026
Watermarking LLMs: The Mathematics of AI Detection
How green-list token biasing, entropy thresholds, and z-score verification work to detect AI text, and why robust watermarking remains elusive.
Aug 2026
Inference-Time Scaling
How scaling test-time compute through search and reward models trades latency for frontier reasoning, ending the pre-training monopoly.
Aug 2026
Inference-Time Scaling: The Test-Time Compute Revolution
How PRMs, MCTS, and adaptive compute are shifting the focus from pre-training parameters to test-time search, yielding breakthrough LLM reasoning.
Aug 2026
Breaking the Autoregressive Bottleneck
How speculative decoding, Medusa, EAGLE, and Multi-Token Prediction (MTP) bypass the memory bandwidth wall to accelerate LLM inference.
Aug 2026
State Space Models and the Mamba Architecture
The math, memory economics, and hardware tricks behind State Space Models, Mamba-2's Structured State Space Duality, and AI21's Jamba.
Aug 2026
Solving the Memory Wall of LLM Inference
How continuous batching and PagedAttention solved the memory fragmentation wall and unlocked 24x higher throughput for LLM serving.
Aug 2026
The Distributed Training Matrix: 3D Parallelism
How frontier models shard parameters, gradients, and optimizer states across 100,000 GPUs to beat the memory bandwidth wall.
Aug 2026
Speculative Decoding and the Multi-Token Frontier
How draft models, Medusa heads, and n-gram lookup trade surplus compute for memory bandwidth to break the autoregressive bottleneck.
Aug 2026
Token Sampling: The Mathematics of LLM Generation
How temperature, top-p, and min-p algorithms manipulate probability distributions to balance creativity and coherence in large language models.
Aug 2026
Inference-Time Scaling and the Test-Time Compute Frontier
How LLMs use MCTS, verification, and repeated sampling to trade inference compute for reasoning performance, bridging the gap between small and frontier models.
Aug 2026
The KV Cache Bottleneck
How PagedAttention eliminated the 80% memory waste of LLM serving, allowing engines like vLLM to achieve 4x higher throughput via continuous batching.
Aug 2026
Speculative Decoding: Breaking the Memory Wall
How draft models, multiple heads, and multi-token prediction bypass the memory bandwidth bottleneck of autoregressive inference.
Aug 2026
Model Pruning: The Math of Sparsity
Structured vs unstructured sparsity, Wanda, SparseGPT, and the economics of zeroing out LLM weights to reclaim inference compute.
Aug 2026
Inference-Time Scaling
How test-time compute is replacing pretraining parameters to unlock the next frontier of LLM reasoning.
Aug 2026
Adaptive Computation and Mixture-of-Depths
How LLMs learn to route tokens dynamically, bypassing layers to save compute and breaking the paradigm of static inference FLOPs.
Aug 2026
Model Merging: The Math of Blending Intelligence
How SLERP, TIES, and DARE combine fine-tuned LLMs without backpropagation, resolving parameter interference to create emergent capabilities.
Aug 2026
Beyond Transformers: Mamba and State Space Models
How selective State Space Models solve the Transformer's quadratic bottleneck, offering constant memory and up to 5x throughput at extreme sequence lengths.
Aug 2026
The Parallelism Matrix
How ZeRO, Megatron-LM, GPipe, and Sequence Parallelism combined to break the memory wall and enable trillion-parameter distributed training.
Aug 2026
State Space Models and Mamba
How selective state spaces and hardware-aware parallel scans break the quadratic attention bottleneck, enabling infinite context windows.
Aug 2026
Speculative Decoding and Multi-Token Prediction
How feature-level drafting, Medusa heads, and EAGLE-2 break the memory-bandwidth bottleneck of autoregressive LLM inference.
Aug 2026
PagedAttention and KV Cache Economics
How static allocation wasted 80% of GPU memory during LLM inference, and how PagedAttention solved it by managing the KV cache like an operating system.
Aug 2026
The Edge Inference Stack
An analysis of MLX, llama.cpp, and Ollama, and the renaissance of local LLM inference on consumer hardware.
Aug 2026
Tokenizer Wars
BPE, WordPiece, SentencePiece, Unigram and Tiktoken — how the subword layer shapes compression, multilingual reach and model cost. Aug 2026.
Aug 2026
The Flash Attention Revolution
How tiling attention onto GPU SRAM cut memory from quadratic to linear and made long context practical. FA1 to FA3, with real speedups.
Aug 2026
The LoRA Family
How low-rank adaptation shrank fine-tuning from retraining 175B parameters to a few million — and the variants that pushed it further.
Aug 2026
Speculative Decoding
How draft models guess tokens and target models verify them, cutting LLM decode latency by 2-3x without changing outputs.
Aug 2026
The Inference Cost Curve
How the price of a token fell 600x in three years — GPT-4 to GPT-5-nano, open-weights serving, and the techniques that bent the curve. Aug 2026.
Aug 2026
The Open-Weights vs Closed-Source Gap
How far open-weights models trail the closed frontier, and how fast that gap is closing — benchmarks, economics, and licenses. Aug 2026.
Aug 2026
Prompt Engineering in 2026
What still works in prompt engineering — and what the 1M-token context window and better models made obsolete.
Aug 2026
Sovereign AI
Why nations are building their own AI: national LLM programs, government compute, and the open-weights models that make sovereignty possible. Aug 2026.
Aug 2026
Hallucination Measurement
How we measure when an LLM fabricates: the benchmarks, the methods, the rates by model, and why the problem resists measurement.
Aug 2026
The Benchmark Contamination Problem
How training on test data inflates LLM benchmark scores, and the human-filtered, continuously-refreshed benchmarks built to resist it.
Aug 2026
Multimodal Fusion Architectures
How LLMs learn to see, hear and read at once — from CLIP's contrastive pairs to LLaVA's adapters to Gemini's native multimodal context.
Aug 2026
RAG vs Long Context
The cost and quality tradeoff between retrieval-augmented generation and stuffing everything into one long context window.
Aug 2026
The Alignment Tax
How RLHF, DPO and GRPO align models — and what that alignment quietly costs in capability.
Aug 2026
The Agent Framework Landscape
LangChain, CrewAI, AutoGen, LangGraph and the shift to direct model-plus-tool agents — which abstraction actually earns its complexity.
Aug 2026
The GPU Supply Chain
From A100 to Rubin: the NVIDIA roadmap, the memory-bandwidth bottleneck, and the supply chain that decides who gets to train frontier AI.
Aug 2026
Attention Variants
MHA, GQA, MQA and MLA compared on KV cache size, quality and throughput — how attention shrank from 3.8 MB to 67 KB per token.
Aug 2026
The Distillation Toolkit
Knowledge distillation for LLMs: from Hinton's logits to DeepSeek-R1's 800K-sample reasoning transfer, and the compression ratios that hold up.
Aug 2026
The MoE Routing Problem
Load balancing, expert collapse, and the routing strategies that decide which experts a token wakes up. From Switch to DeepSeek's aux-loss-free bias.
Aug 2026
The Scaling Laws Debate
Chinchilla said ~20 tokens per parameter. Modern models train on 100 to 800 times more. The debate over overtraining, and where it ends.
Aug 2026
MCP: The Universal Tool Protocol
Anthropic open-sourced a USB-C port for AI agents. Eight months later, the ecosystem has hundreds of servers and every major IDE. Aug 2026.
Aug 2026
SWE-bench and the Coding Arms Race
AI coding benchmarks went from 2% to 80% in under two years, but passing tests and writing production code are not the same thing.
Aug 2026
The Energy Economics of Inference
Measured joules per token, training megawatt-hours, and carbon grams per query — the energy cost of AI inference and training, mapped from silicon to grid.
Aug 2026
The Inference Engine Wars
Six open-source LLM inference engines compared on throughput, hardware coverage, and the innovations that reshaped serving.
Aug 2026
The Quantization Spectrum
From FP16 to 1.58-bit ternary weights: the format landscape, quality cliffs, and VRAM economics of every LLM quantization method in 2026.
Aug 2026
The Sub-8B Renaissance
Small LLMs that beat frontier models: distillation, textbook-quality data, and the end of bigger-is-always-better. Aug 2026.
Aug 2026
The Context Window War
From 4K tokens to 10M: the positional encoding tricks, KV cache economics, and the gap between advertised and usable context. Aug 2026.
Aug 2026
The Terabyte Frontier
Frontier open-weights MoE models — from 109B Scout to 2.8T Kimi K3 — and the exact GPU hardware each one needs to run locally. Aug 2026.
Aug 2026
Inside the J-Space
Anthropic's discovery of a hidden cognitive workspace inside Claude's latent layers — and the inference architecture it unlocked.
Aug 2026
The Immaturity of Diffusion Models
Benchmark data showing diffusion models fail at composition, typography, temporal stability, and compute efficiency versus autoregressive alternatives.
Aug 2026
The Local LLM Ceiling
Blackwell-class consumer hardware, the two phases of inference, and the exact VRAM point where one GPU stops being enough. Q2 2026.
Aug 2026
Software Is Becoming Disposable
A visual field note on how AI collapses the marginal cost of software — disposable apps, ephemeral interfaces, and the durable substrate underneath.
Aug 2026