Transformers

Andrej Karpathy

Karpathy's 90-Second Tour Through 33 Years of Neural Nets

Andrej Karpathy compresses 33 years of neural network evolution into a 90-second retrospective, tracing the architectural lineage from early MLPs to today's transformer-based generative models powering synthetic media.

LLM Infrastructure

Breaking the KV Wall: Scaling LLM Inference at Scale

As LLMs handle longer contexts and more concurrent users, the KV cache has become the dominant bottleneck in inference. New architectural approaches aim to break through this memory wall for next-generation serving.

LLM Inference

The Silent Speedup: How KV Cache Makes AI Feel Instant

KV caching is the unsung optimization that makes modern LLMs feel real-time. Here's how it transforms transformer inference from quadratic drudgery into a fast, token-by-token stream.

AI Infrastructure

SubQ: Miami Startup's Attention Trick Runs 52x Faster

A four-person Miami startup called SubQ claims a new attention mechanism that runs 52x faster than standard transformers at one-fifth the cost of Claude Opus, hinting at cheaper long-context inference for video and multimodal AI.

Multimodal AI

The Evolution of Encoders: From Basics to Multimodal AI

Encoders have evolved from simple feature extractors into the backbone of multimodal AI, powering today's video, image, and audio generation systems. Here's how they got here and why it matters for synthetic media.

super-resolution

Rank-Factorized Neural Bias Enables Scalable Super-Resolution

New research combines rank-factorized implicit neural bias with FlashAttention to scale super-resolution transformers efficiently, advancing high-quality image synthesis for AI-generated content.

LLM

KV Caching: How This Optimization Makes LLM Inference Viable

Key-value caching is the hidden optimization that makes large language models practical. Learn how this technique eliminates redundant computation during inference.

Hugging Face

Hugging Face Transformers v5: Simplified APIs for AI Development

Hugging Face releases Transformers v5 with cleaner APIs, unified model loading, and breaking changes that simplify building AI applications across text, image, and video domains.

Transformers

Positional Encoding Methods: Why Token Order Matters in AI

Transformers process tokens in parallel, losing sequence information. Four positional encoding methods—sinusoidal, learned, RoPE, and ALiBi—solve this fundamental challenge differently.

LLM

KV Cache Explained: The Hidden Engine Powering Fast LLM Inference

Understanding Key-Value caching in transformer architectures reveals how modern LLMs achieve fast token generation. This core optimization technique is essential for efficient AI inference.

NLP

Research Reveals How AI Transformers Distort Business Sentiment

New research exposes systematic sentiment bias in NLP transformers, showing how AI language models struggle to maintain neutral tone in business communications, raising concerns for automated content generation.

LLM Research

New Research Maps LLM Embeddings Using Hamiltonian Physics

Researchers propose a physics-inspired framework treating LLM token embeddings as discrete semantic states governed by Hamiltonian dynamics, offering new insights into transformer interpretability.