minimaxir-com

The article explores whether repeatedly asking large language models to "write better code" can improve code performance. It focuses on using AI to generate faster, more efficient code rather than just accelerating the coding process itself.

The article recommends using Parquet files and the Polars library as the optimal portable method for storing and working with text embeddings. It advises against storing embeddings in CSV format due to inefficiencies.

The author, an experienced LLM user, states they don't frequently use generative large language models. However, they find the specific applications they do use LLMs for to be invaluable.

Researchers developed a model to predict average IMDb movie ratings using text embeddings from movie metadata. The approach analyzes textual descriptions and other metadata to forecast audience ratings. This method demonstrates how AI can extract predictive signals from unstructured movie information.

Gemini can identify public figures in images, while ChatGPT and Claude currently do not offer this capability. This represents a functional difference between major AI models regarding image recognition of people.

The article examines whether modern large language models can accurately count the number of 'b's in the word "blueberry," testing their ability to handle this specific adversarial question.

The article discusses attempts to jailbreak Claude Haiku 4.5, an AI model. The AI responds by questioning whether the jailbreak attempts are genuinely useful or merely testing its security measures.

Nano Banana is an AI image generation model that supports up to 32,768 input tokens, enabling extensive prompt engineering for highly nuanced image creation.

Nano Banana Pro is described as the best AI image generator, though the article notes it has some caveats. The main issue highlighted is that the technology is considered too effective.

A skeptic of AI agent coding conducts an extensive experiment using AI agents to code a project. The detailed analysis examines the practical capabilities and limitations of current AI coding agents in real-world development scenarios.