LLM Post-Training: GRPO, DPO, RLHF & Fine-Tuning
How LLMs are trained and improved: RL methods like GRPO and DPO, distillation, fine-tuning, quantization, and retrieval — explained for practitioners.
AI 101
+1
11 min read
Jul 5, 2026
GRPO Explained: Group Relative Policy Optimization
GRPO (Group Relative Policy Optimization) is a critic-free RL algorithm for LLM training. Learn how it works, how it differs from PPO, and its role in DeepSeek-R1
AI 101
+1
10 min read
Mar 25, 2026
AI 101: Transformers Depth Is an Addressable Dimension
MoDA and Attention Residuals make Transformer depth queryable — not just a fixed pipeline. Learn how both approaches work and why it matters for deep LLMs.
AI 101
+1
14 min read
Mar 11, 2026
AI 101: Beyond RL: The New Fine-Tuning Stack for LLMs
How Doc-to-LoRA, Text-to-LoRA, LoRA-Squeeze, Kron-LoRA, MoA & Evolution Strategies replace expensive RL loops with a modular, cheaper post-training stack for LLMs.
AI 101
+1
14 min read
Feb 11, 2026
AI 101: "On-Policy Distillation Zeitgeist"
Three papers on on-policy self-distillation — OPSD, SDFT, SDPO — replacing sparse rewards with dense token-level feedback. Benchmarks and limitations covered.
AI 101
+1
10 min read
Feb 4, 2026
AI 101: Conditional Memory and the Rise of Selective Intelligence
Engram by DeepSeek gives LLMs conditional memory: models learn when to access knowledge. Covers architecture, U-shaped allocation law, and reasoning gains.
AI 101
+1
10 min read
Nov 5, 2025
What matters for RL? Precision! Switching BF16 → FP16
BF16 vs FP16: how switching precision during RL fine-tuning fixes training-inference mismatch, stabilizes GRPO, and why Karpathy applied it to nanochat.
AI 101
+1
10 min read
Oct 15, 2025
AI 101: What are Modular Manifolds?
Modular manifolds treat neural network layers as geometric modules for stable, scalable optimization. A deep dive into Thinking Machines Lab's approach.
AI 101
+1
9 min read
Sep 3, 2025
What is XQuant?
XQuant and XQuant-CL cut LLM KV cache memory up to 12x by storing input activations instead of keys and values. How the method works and when to use it.
AI 101
+1
13 min read
Aug 13, 2025
What's New in Test-Time Scaling?
How Chain-of-Layers, MindJourney, and Google's TTD-DR push test-time scaling further — and where inverse scaling shows its limits.
AI 101
+1
12 min read
Apr 23, 2025
What is MoE 2.0? Update Your Knowledge about Mixture-of-experts
The fresh angle on current Mixture-of-Expert. We discuss what new MoE techniques like S'MoRE, Symbolic-MoE, and others mean to the next generation AI
AI 101
+1
12 min read
Mar 26, 2025
Slim Attention, XAttention & KArAt: Three New Attention Mechanisms Explained
We explore three advanced attention mechanisms which improve how models handle long sequences, cut memory use and make attention learnable
AI 101
+1
10 min read
Mar 12, 2025
How to Reduce Memory Use in Reasoning Models
How LightThinker and Multi-Head Latent Attention (MLA) reduce memory use and speed up inference in reasoning models like DeepSeek-R1.
AI 101
+1
12 min read
Mar 5, 2025
Everything You Need to Know about Knowledge Distillation
This is one of the hottest topics thanks to DeepSeek. Learn with us: the core idea, its types, scaling laws, real-world cases and useful resources to dive deeper
AI 101
+1
10 min read
Feb 12, 2025
What are Chain-of-Agents and Chain-of-RAG?
CoRAG and Chain-of-Agents are two upgrades to standard RAG: one masters multi-hop reasoning, the other handles extremely long contexts. Compare both
AI 101
+1
5 min read
Jan 29, 2025
The Keys to Prompt Optimization
Prompt optimization improves LLM outputs by refining query structure. Covers expansion, decomposition, disambiguation, abstraction, and combining strategies effectively.
AI 101
+2
9 min read
Jan 8, 2025
What is HtmlRAG, Multimodal RAG and Agentic RAG?
Three RAG upgrades explained: HtmlRAG preserves HTML structure, Multimodal RAG retrieves images, and Agentic RAG reformulates queries for better results
