What is the Model Context Protocol (MCP)?
Anthropic
MCP is an open standard for connecting AI applications to external systems.
Open library
The library
The essential reading of generative AI and agents, open in a clean reading view, with Meow and Standard summaries when available. No account needed to read.
563 papers · 561 with a Meow summary · 333 posts
Key publications from Anthropic, OpenAI, Google DeepMind, Meta, DeepSeek, Qwen, Z.ai, Kimi and other labs, plus the most influential essays and blogs.
Defeating Nondeterminism in LLM Inference
Thinking Machines Lab · Thinking Machines
The authors argue that endpoint variation usually comes from changing server batches, which alter floating-point reduction paths—not from GPU cores racing unpredictably.
LoRA Without Regret
Thinking Machines Lab · Thinking Machines
The paper’s “low-regret regime” is conditional: LoRA can match full fine-tuning for post-training when adapters cover important layers and have enough capacity.
On-Policy Distillation
Thinking Machines Lab · Thinking Machines
On-policy distillation has the student generate its own trajectories while a strong teacher gives token-level feedback—combining RL’s student-specific training with distillation’s denser signal.
Interaction Models: A Scalable Approach to Human-AI Collaboration
Thinking Machines Lab · Thinking Machines
The proposal: build real-time collaboration into the model, so it can respond to a person’s changing input instead of waiting for neat turns.
Harness Engineering for Self-Improvement
Lilian Weng
The article’s key shift is from improving an agent’s answer to improving the machinery that repeatedly produces and checks answers.
Scaling Laws, Carefully
Lilian Weng
Scaling laws make small training runs useful for forecasting larger ones, but their compute recommendations can shift with the experiment range, data regime, and fitting choices.
Why We Think
Lilian Weng
More computation at inference can improve reasoning, but only when the extra work is spent well: longer traces and self-revision are not automatically more reliable.
Reward Hacking in Reinforcement Learning
Lilian Weng
Reward hacking means a high score can mark a loophole, not task success; in RLHF, optimization can also make incorrect answers more persuasive to human evaluators.
Extrinsic Hallucinations in LLMs
Lilian Weng
Extrinsic hallucination is fabricated content that lacks grounding in context or world knowledge. The post argues that reliable models must be factual and acknowledge when they do not know.
Diffusion Models for Video Generation
Lilian Weng
Video diffusion must keep objects and motion coherent across frames, not just make convincing individual images; limited high-quality video data makes that harder.
Thinking about High-Quality Human Data
Lilian Weng
A cleaner label set is not always a better one: quality work must remove annotation errors without flattening legitimate disagreement.
Adversarial Attacks on LLMs
Lilian Weng
LLM safeguards can fail at inference without changing model weights. Weng’s survey argues that how attacks are found—and how failures are judged—shapes apparent robustness.
LLM Powered Autonomous Agents
Lilian Weng
An LLM agent’s autonomy comes from the system around the model: planning, memory, and tools. Those parts extend what it can do, but do not remove its core failure modes.
Prompt Engineering
Lilian Weng
Prompt engineering steers a language model without changing its weights, but it is empirical: a prompt that works for one model may fail on another.
The Transformer Family Version 2.0
Lilian Weng
When a Transformer needs more context, less memory, or faster inference, which part of attention should change?
Large Transformer Model Inference Optimization
Lilian Weng
Transformer inference gets cheaper only when the fix targets the actual bottleneck—memory, computation, or latency—and fits the hardware that must run it.
Generalized Visual Language Models
Lilian Weng
The design choice is where image information enters a language model—and whether adapting it changes the language model at all.
How to Train Really Large Models on Many GPUs?
Lilian Weng
Training giant networks is a memory-and-scheduling problem: divide data or computation across GPUs, then manage synchronization, idle stages, and extra state so the split actually helps.
What are Diffusion Models?
Lilian Weng
Diffusion models combine analytical tractability with flexibility, but pay for it with a long, expensive sampling chain.
Contrastive Representation Learning
Lilian Weng
Contrastive representation learning places related samples close together in an embedding space and separates unrelated ones. Its effectiveness depends on how it constructs positive pairs and negatives.
Attention? Attention!
Lilian Weng
Attention replaces one all-purpose summary with a different weighted view of the input for each prediction.
Policy Gradient Algorithms
Lilian Weng
The key is to inspect what each policy-gradient method estimates, reuses, and constrains—not just its optimizer name.
The Unreasonable Effectiveness of Recurrent Neural Networks
Andrej Karpathy
An RNN reuses a learned update and carries state across steps, so it can handle sequences of varying length. Character-level LSTMs show both the structure this can learn and the errors that remain.
A Recipe for Training Neural Networks
Andrej Karpathy
Neural-net training can fail silently: a runnable model may still be logically misconfigured, so verify each step before adding complexity.
Verifiability
Andrej Karpathy
Karpathy argues that AI automates work more readily when attempts are verifiable, not merely specifiable.
Animals vs Ghosts
Andrej Karpathy
The essay argues that frontier LLMs are not yet clean, experience-only learners: human data and engineering shape them. Still, the author treats pretraining as a practical cold start, not a dead end.
2025 LLM Year in Review
Andrej Karpathy
The author identifies RLVR—training against automatically checkable rewards—as 2025’s major training shift: it enabled longer optimization and reasoning strategies, but helped produce intelligence that remains sharply uneven.
Understanding LSTM Networks -- colah's blog
Chris Olah
Long Short Term Memory networks (LSTMs) add a gated cell state that lets useful information travel through a recurrent network more easily than in a standard RNN.
Feature Visualization
Chris Olah · Distill
Feature visualization creates inputs that show what a network responds to, but raw optimization can exploit high-frequency artifacts; useful interpretations need diversity and constraints.
The Building Blocks of Interpretability
Chris Olah · Distill
What could an explanation show if it followed a network’s learned concepts, not just highlighted pixels?
Attention and Augmented Recurrent Neural Networks
Chris Olah · Distill
The article argues that attention can extend recurrent networks beyond sequence processing by letting them use memory, focus on inputs, vary computation, and build programs.
The Illustrated Transformer
Jay Alammar
A word can use clues from elsewhere in its sentence: the Transformer learns how much to attend to each position.
The Illustrated GPT-2 (Visualizing Transformer Language Models)
Jay Alammar
GPT-2’s text generation came from scaling a familiar decoder-only Transformer: causal masking lets it predict the next token without seeing future tokens, while self-attention builds context.
The Illustrated BERT, ELMo, and co. (How NLP Cracked Transfer Learning)
Jay Alammar
BERT paired reusable pretraining with a Transformer encoder that can use left and right context, then adapt to labeled NLP tasks.
The Illustrated Stable Diffusion
Jay Alammar
Stable Diffusion generates text-guided images by repeatedly predicting and removing noise in compressed latent space, then decoding the result into pixels. Latent-space processing speeds the work; attention brings prompt information into denoising.
The lethal trifecta for AI agents: private data, untrusted content, and external communication
Simon Willison
An agent can leak private data when its tools let it access secrets, read attacker-controlled content, and communicate outward. Hostile instructions can steer it to use all three.
Prompt injection attacks against GPT-3
Simon Willison
Text meant for translation can become a command to GPT-3, turning an ordinary task into a security exploit.
The Dual LLM pattern for building AI assistants that can resist prompt injection
Simon Willison
Prompt injection can turn ordinary hostile text into a tool call or data leak; Willison’s proposal separates the model that can act from the one that reads untrusted content.
Things we learned about LLMs in 2024
Simon Willison
By the end of 2024, GPT-4-level performance had become a crowded field, but capability gains had not solved reliability. Willison’s practical emphasis is on testing the system you build, not trusting fluent output.