# Oct 17, 2025

## Gemini 3.0 spotted in the wild through A/B testing
- **Gemini 3.0** has been observed in A/B testing via Google AI Studio, showcasing impressive SVG generation capabilities, particularly with an Xbox 360 controller image that outperformed existing models.

## Claude Skills are awesome, maybe a bigger deal than MCP
- **Claude Skills** enhance model performance by allowing Claude to load specific instructions and resources relevant to tasks, improving efficiency in specialized applications like Excel and brand compliance.

## State of AI Report 2025
- The **State of AI Report 2025** reveals that **OpenAI** maintains a slight edge in AI development, while **China's DeepSeek** and others are rapidly closing the gap in reasoning and coding tasks, marking a significant shift in global AI leadership.

## Bringing AI to the next generation of fusion energy
- **Google DeepMind** is collaborating with **Commonwealth Fusion Systems (CFS)** to harness AI for advancing fusion energy, aiming to achieve **net fusion energy** with the SPARC tokamak, a significant step towards sustainable energy.

## Nvidia DGX Spark and Apple Mac Studio = 4x Faster LLM Inference with EXO 1.0
- **EXO 1.0** achieves **4x faster LLM inference** by leveraging the **NVIDIA DGX Spark™** for prefill and **Apple Mac Studio** for decode, optimizing each phase's strengths.

## [R] Plain English outperforms JSON for LLM tool calling: +18pp accuracy, -70% variance
- **Natural Language Tools (NLT)** enhances LLM tool-call accuracy by +18 percentage points while reducing variance by 70% and token overhead by 31%, demonstrating a significant shift from structured JSON to natural language frameworks.

## Flight Simulator for the Brain Reveals How We Learn and Why Minds Go Off Course
- A new computer model, **CogLinks**, simulates brain decision-making and adaptability, revealing how misjudgments in context can lead to psychiatric disorders like **schizophrenia** and **OCD**.

## LaSeR: Reinforcement Learning with Last-Token Self-Rewarding
- **LaSeR** introduces a novel approach to **Reinforcement Learning** by simplifying self-verification into a last-token self-rewarding score, enhancing the efficiency of **Large Language Models (LLMs)** during reasoning tasks.

## Attention Is All You Need for KV Cache in Diffusion LLMs
- **Elastic-Cache** optimizes key-value (KV) cache management in diffusion large language models (DLMs) by adaptively refreshing caches based on token importance, significantly enhancing prediction accuracy and reducing decoding latency.

## Show HN: Searchable compression for JSON – ~99% page skip and sub-ms lookups
- **SEE (Semantic Entropy Encoding)** achieves a **combined size of ≈19.5%** of raw data while enabling **searchability** and **low I/O**, making it advantageous for workloads requiring quick access to compressed JSON data.

## [R] Tensor Logic: The Language of AI
- **Tensor Logic (TL)** aims to **unify Deep Learning and Symbolic AI**, offering a framework that integrates various AI models, including neural networks and graphical models, into a cohesive system.

## Asking AI to build scrapers should be easy right?
- **Skyvern's AI can now autonomously write and maintain code, achieving a remarkable 2.7x cost reduction and 2.3x speed increase in automation tasks.**

## How a Gemma model helped discover a new potential cancer therapy pathway
- **Google's C2S-Scale 27B model**, with 27 billion parameters, has successfully identified a novel cancer therapy pathway by predicting the synergistic effect of silmitasertib and low-dose interferon on antigen presentation.

## [R][D] A Quiet Bias in DL’s Building Blocks with Big Consequences
- **Deep learning's building blocks**, such as activation functions and optimisers, introduce a foundational bias that shapes network representation and reasoning, suggesting a new symmetry-based design axis for improved model performance.