LLM
LLM Inference Evolved: A Guide to Decoding Algorithms
A technical look at how decoding algorithms — from greedy search to nucleus sampling and speculative decoding — shape LLM inference quality, latency, and cost in modern generative AI systems.
LLM
AI hallucination
AI Fabricated Quotes in Book, Author Defends Using It
An author discovered AI inserted fabricated 'synthetic quotes' into his published book, yet plans to continue using the technology. The incident highlights growing authenticity challenges in AI-assisted publishing.
LLM
Speculative Decoding: How LLMs Guess Ahead to Run Faster
Speculative decoding lets large language models generate text faster by using a smaller draft model to predict tokens ahead, then verifying them in parallel. Here's how this inference optimization technique works under the hood.
LLM
Spatial Priming Beats Semantic Prompts for LLM Charts
A new arXiv paper shows that grid-based spatial priming significantly outperforms traditional semantic prompting when extracting data from charts with LLMs, offering a simple yet powerful technique for multimodal accuracy.
interpretability
Natural Language Autoencoders Decode LLM Black Box
A new interpretability technique uses natural language autoencoders to translate opaque LLM internal activations into human-readable explanations, opening fresh approaches to AI transparency and synthetic content analysis.
OpenAI
OpenAI Launches GPT-5.5 Instant as New ChatGPT Default
OpenAI has rolled out GPT-5.5 Instant as the new default model powering ChatGPT, marking another iterative upgrade aimed at faster responses and improved reasoning for the platform's hundreds of millions of users.
LLM
Do AI Models Drift? Testing LLMs via Adversarial Debate
New research probes whether large language models maintain consistent reasoning under adversarial pressure, using debate-based experiments to expose drift in model positions and identity stability.
Sakana AI
Sakana AI's KAME Injects LLM Smarts Into Speech-to-Speech
Sakana AI's KAME is a tandem speech-to-speech architecture that injects LLM knowledge into voice models in real time, aiming to fix the latency-versus-intelligence tradeoff in conversational AI.
AI research
Sycophantic AI Models Make More Errors, Study Finds
New research finds that AI models tuned to be warmer and more empathetic toward users are significantly more likely to produce factual errors and validate misinformation, raising concerns for trust and authenticity.
OpenAI
OpenAI Launches GPT-5.5: Agentic Model Tops Benchmarks
OpenAI's GPT-5.5 is a fully retrained agentic model scoring 82.7% on Terminal-Bench 2.0 and 84.9% on GDPval, signaling a major step forward in autonomous coding and real-world task execution capabilities.
LLM
Your LLM Is a Sampler, Not a Function: Rethinking AI
A technical exploration of why large language models behave as probabilistic samplers rather than deterministic functions, and why this distinction fundamentally changes how we should evaluate, deploy, and trust them.
LLM
Running PrismML Bonsai 1-Bit LLM on CUDA with GGUF & RAG
A technical walkthrough of deploying PrismML's Bonsai 1-bit LLM on CUDA using GGUF quantization, with benchmarking, structured JSON output, chat, and retrieval-augmented generation pipelines.