# Jan 12, 2025

## AI Founders Will Learn The Bitter Lesson

- **AI founders must recognize that general methods leveraging computation consistently outperform specialized approaches**, as evidenced by historical patterns in AI research, where reliance on human domain knowledge has led to stagnation and missed opportunities for innovation.

## Phi 4 Available on Ollama

- **Phi 4** is a **14B parameter** open model from Microsoft, designed to enhance generative AI applications through a combination of synthetic datasets and academic resources, ensuring robust performance in **memory-constrained** and **latency-sensitive** environments.

## Have Transformers Won in Computer Vision?

- **Transformers** have gained traction in **Computer Vision** since 2020, with the introduction of **Vision Transformers** challenging traditional architectures like **ResNets** for tasks such as image classification.

## A Hard Algorithmic Benchmark for Future Reasoning Models

- **A new benchmark** for AI models challenges them to deduce hidden logic from transformed data, revealing significant shortcomings in current models like **GTP, o1, Sonet, and Gemini**.

## Show HN: SemHash – Fast Semantic Text Deduplication for Cleaner Datasets

- **SemHash** is a **lightweight** tool for **semantic text deduplication**, utilizing **Model2Vec** for fast embedding generation and **Vicinity** for efficient similarity search, making it suitable for both single and multi-dataset deduplication tasks.

## FuseGPT: Learnable Layers Fusion of Generative Pre-trained Transformers

- **FuseGPT** introduces a novel approach to **layer fusion** in Generative Pre-trained Transformers, enhancing model efficiency and performance through learnable parameters that adaptively combine multiple transformer blocks.