# May 13, 2026

Daily

### Show HN: Needle: We Distilled Gemini Tool Calling into a 26M Model
- **Needle** distills **Gemini 3.1** into a **26m parameter Simple Attention Network**, enabling local finetuning on personal devices, achieving **6000 toks/sec prefill** and **1200 decode speed** in production.

### Interaction Models from Thinking Machines Lab 
- **Interaction models** represent a transformative shift in AI, enabling real-time collaboration across **audio, video, and text**, thus overcoming the limitations of traditional turn-based systems.

### A History of IDEs at Google
- **Google's IDE landscape evolved from fragmentation to a unified platform** with the introduction of Cider V, which now supports **80% of development** in the main codebase, enhancing productivity through better integrations and AI features.

### Beyond Semantic Similarity
- **Direct Corpus Interaction (DCI)** allows agents to search raw data using terminal tools, bypassing traditional retrieval systems that limit access and hinder multi-step reasoning.

### The US is winning the AI race where it matters most: commercialization
- The **US leads in AI commercialization**, leveraging cloud infrastructure, data, and strategic capital, significantly outpacing competitors like China in revenue and adoption since the launch of DeepSeek R1 in January 2025.

### An idiot's guide to lead optimisation for proteins
- **Lead optimisation** is a critical phase in drug design, where existing molecules are refined to enhance their efficacy, leveraging **machine learning** to propose and test modifications efficiently, as exemplified by the **Cradle-1 pipeline**.

### Show HN: Statewright – Visual state machines that make AI agents reliable
- **Statewright** enhances AI agent performance by implementing **state machines** that limit tool access, allowing models to focus on specific tasks and improve efficiency across various platforms like Claude Code and Codex.

### Fc, a lossless compressor for floating-point streams
- **`fc` is a lossless compressor for IEEE-754 64-bit doubles, achieving a compression ratio of 3.07 and a decoding speed of 1277 MB/s, outperforming competitors on structured data.**

### $δ$-mem: Efficient Online Memory for Large Language Models
- **$δ$-mem** introduces a **lightweight memory mechanism** that enhances large language models by utilizing a compact online state, allowing for efficient historical information management without the need for extensive context window expansion.

### Human-level performance via ML was *not* proven impossible with complexity theory 
- The **Ingenia Theorem** proposed by Van Rooij et al. claims that achieving AGI via ML is impossible, but this assertion is fundamentally flawed due to the lack of a precise definition for "human-level classifier."

### TabPFN-3 just released: a pre-trained tabular foundation model for up to 1M rows 
- **TabPFN-3** is a **pre-trained tabular foundation model** capable of handling **up to 1M rows** in a single forward pass, significantly enhancing efficiency with **10x-1000x faster inference** than its predecessors.

### Elastic Attention Cores for Scalable Vision Transformers 
- **Elastic Attention Cores** introduce a **core-periphery block-sparse attention** structure for Vision Transformers, reducing computational costs from **N²** to **2NC + C²** for **C** core tokens, enhancing scalability at higher resolutions.

### I Found a Hidden Ratio in Transformers That Predicts Geometric Stability 
- A **hidden ratio** in transformer models, derived from Lyapunov spectral analysis, predicts **geometric stability**, indicating whether a model will collapse to rank-1 based on the **MLP and attention spectral norms**.

### Learning, Fast and Slow: Towards LLMs That Adapt Continually 
- **Fast-Slow Training (FST)** enables large language models (LLMs) to adapt continually by utilizing **fast weights** for task-specific learning while maintaining **slow weights** for general reasoning, enhancing efficiency and performance.

### Solve the Loop: Attractor Models for Language and Reasoning
- **Attractor Models** enhance language modeling and reasoning by iteratively refining embeddings, achieving a **46.6% improvement in perplexity** and **19.7% in accuracy** while reducing training costs compared to standard Transformers.