← Engineermaxxing

Nuranors

Deep technical series and architecture overviews. From first principles to production-level detail, with interactive visualizations throughout.

⌕

No matches found.

Deep-Dive Series

10-Part Series

LLM Internals

Demystifying how large language models actually work — from the mechanics of tokenization to the algorithms behind alignment and the systems that serve them at scale.

Tokenization Attention Training LoRA Alignment Quantization KV Cache FlashAttention Semantic Spine Two-Tower Retrieval

10 articles ~395 min 35+ demos

8-Part Series

Diffusion & Flow Matching

Demystifying diffusion models and flow matching — from probability foundations and DDPMs to score-based models, SDEs, modern architectures, and state-of-the-art generation.

Generative Models DDPMs Score Matching SDEs Flow Matching DiT Samplers Applications

8 articles ~290 min 30+ demos

8-Part Series

Vision-Language Models

How images and text merge into a single intelligence — from vision transformers and CLIP to multimodal fusion, visual instruction tuning, and spatial grounding.

Vision Foundations ViT CLIP Multimodal Fusion LLaVA Training Grounding Applications

8 articles ~280 min 30+ demos

8-Part Series

Vision-Language-Action Models

Where perception meets physical action — from imitation learning and behavioral cloning to RT-2, OpenVLA, and the foundation models teaching robots to act.

Embodied AI Imitation Learning Robot Vision Language Policies RT-2 / OpenVLA Sim-to-Real Planning Deployment

8 articles ~280 min 30+ demos

Deep-Dive · Field Manual

The Robot Learning Stack

A teardown of the architectures, losses, and training recipes that move modern manipulators — from behavior cloning’s first sin to flow-matched VLAs and the pixel-RL renaissance.

Behavior Cloning Diffusion Policy ACT VLAs PPO / SAC Sim-to-Real World Models Deployment

25 sections ~90 min 15+ demos

Deep-Dive · Research Teardown

Claude Plays Robotics

Anthropic wired frontier models to pendulums, robot dogs, and humanoids — why the control interface matters as much as the model, with three runnable labs.

Control Interfaces Latency Wall LIBERO VLA Supervision Navigation Project Fetch 3 Live Labs

10 chapters ~90 min 3 runnable labs 9 demos

Deep-Dive · Architecture Teardown

From GPT-2 to Kimi K3: One Memory, Seven Eviction Policies

The 22,580× headline, audited — and the real story: KV cache → linear attention → delta rule → gates → KDA → K3, told as one 4×4 memory you watch succeed, blur, overwrite, fade, and survive. Every listing runs; every figure is draggable.

KV Cache Linear Attention Delta Rule Gated DeltaNet KDA / MLA Latent MoE AttnRes

9 chapters ~90 min 10 interactive widgets

Deep-Dive · Deployment

Scaling, Optimizing & Deploying Foundation Models

VLMs, VLAs, and World Models — from architecture internals to quantization, inference acceleration, and production serving. Includes autonomous driving perception/planning and five portfolio projects.

VLM / VLA World Models Scaling Laws Quantization KV Cache vLLM / TensorRT Autonomous Driving Portfolio Projects

14 sections ~55 min 11+ demos

Deep-Dive · Research Teardown

Sesame CSM: Crossing the Uncanny Valley

The research note that made text-to-speech conversational, read closely — one transformer over two kinds of audio token, the compute-amortization trick that pays for it, and why voice presence defeats every benchmark we have.

Voice Presence Semantic vs Acoustic Tokens RVQ Delay Compute Amortization Contextual TTS CMOS Saturation Full Duplex

10 chapters ~150 min 10 interactive widgets

Deep-Dive · Field Manual

The Voice Agent Stack

Every millisecond between a mouth and an ear, itemized — the latency ledger line by line, cascade versus speech-to-speech, endpointing, barge-in, transport, and the evals nobody puts in a demo.

Latency Ledger Cascade vs S2S Endpointing Barge-In Paralinguistics WebRTC / Telephony Voice Evals

11 chapters ~175 min 11 interactive widgets

Thinking Machines Blog Deep Dives

Thinking Machines · July 2026

Inkling, Dissected

A block-by-block teardown of the 975B open-weights MoE that wakes 41B per token — sigmoid routing, Muon, the 5:1 attention weave, the thinking-effort dial, and the model that fine-tuned itself. With four live labs you run in the page.

MoE Routing Muon Sliding-Window Attention Thinking Effort Self-Fine-Tuning Calibration 4 Live Labs

11 chapters ~100 min 4 runnable labs 12 demos

Thinking Machines · May 2026

The Long-Running Agent Stack

The four-lane map: managed harnesses, frameworks, specialized agents, and durable execution. MCP + A2A protocols, 43 platforms compared, 12 concepts shown working.

MCP A2A Durable Execution Agent Harnesses HITL Sandboxing

6 chapters 18 simulations 43 platforms

Thinking Machines · May 2026

Interaction Models

Real-time human-AI collaboration via 200ms micro-turns, encoder-free fusion, and streaming sessions.

Micro-turns dMel Flow Matching Batch Invariance MoE

9 chapters 10 canvases 11 papers

Thinking Machines · Oct 2025

On-Policy Distillation

Combining on-policy sampling with dense teacher supervision for 9–30x cheaper training than RL.

Reverse KL RLHF Continual Learning Math Reasoning

8 chapters 7 canvases

Thinking Machines · Sep 2025

LoRA Without Regret

When LoRA matches full fine-tuning, why all layers matter, and the information theory of RL capacity.

LoRA PEFT eNTK Information Theory RL

8 chapters 7 canvases

Thinking Machines · Sep 2025

Modular Manifolds

Constraining weight matrices to Stiefel manifolds with spectral-norm budgeting across layers.

Manifold Optimization Muon SVD Lipschitz

8 chapters 6 canvases

Thinking Machines · Sep 2025

Defeating Nondeterminism

Why LLM inference is nondeterministic (batch variance, not atomics) and how to make it bitwise reproducible.

FP Arithmetic Batch Invariance FlashDecoding On-Policy RL

8 chapters 6 canvases

Architecture Atlas

12 Interactive Overviews

Modern AI Architecture Atlas

A bird's-eye map of every major AI architecture — interactive diagrams, intuitive explanations, and the research context to navigate the field.

Transformer Mamba / SSM Diffusion Flow Matching VAE / VQ-VAE GAN CLIP / Contrastive VLM VLA World Models NeRF / 3DGS Reward / Alignment

12 overviews ~180 min 48+ diagrams