# Oct 30, 2025

## News Highlights

### Alphabet tops $100B quarterly revenue for first time, cloud grows 34%
- **Alphabet achieved a historic milestone** with quarterly revenue surpassing **$100 billion**, reporting **$102.35 billion**, driven by a **34% growth** in its cloud unit, which now has a **$155 billion backlog** of customer demand.

### Language Models Are Injective and Hence Invertible
- **Transformer language models are proven to be injective**, meaning they can uniquely map discrete input sequences to continuous representations without loss, a property maintained throughout training.

### NLP conferences look like a scam..
- **NLP conference papers often lack theoretical justification**, with 90% of reviewed papers failing to provide solid foundations, while the remaining 10% mislabel findings as _theorems_ under unrealistic assumptions.

### Extropic is building thermodynamic computing hardware
- **Extropic** introduces **thermodynamic computing**, a groundbreaking approach that leverages thermodynamic principles to enhance computational efficiency and performance, as showcased in their [launch video](https://www.youtube.com/watch?v=Y28JQzS6TlE).

### How the U.S. National Science Foundation Enabled Software-Defined Networking
- **NSF's investments in Software-Defined Networking (SDN) over two decades have fundamentally transformed network design**, enabling greater control and flexibility for network owners, which has led to innovations in both public and private sectors.

### Video‐rate tunable colour electronic paper with human resolution
- **Retina E-paper** technology achieves **ultra-high resolutions** exceeding **25,000 pixels per inch** (PPI) using electrochromic WO3 metapixels, enabling dynamic control over optical properties for immersive displays.

### FastJAM: a Fast Joint Alignment Model for Images (NeurIPS 2025)
- **FastJAM** introduces a **lightweight graph-based framework** for joint image alignment, achieving results in seconds compared to the hours required by previous methods, utilizing **sparse keypoints** and **graph neural networks (GNNs)** for enhanced efficiency.

### The Laplace Perceptron: A Complex-Valued Neural Architecture for Continuous Signal Learning and Robotic Motion
- The **Laplace Perceptron** introduces a novel approach to **temporal signal learning** and robotic control by utilizing **complex-valued damped harmonics**, enhancing both signal representation and optimization efficiency.

### Update on DynaMix: Revised paper & code (Julia & Python) now available
- The **revised paper** on DynaMix, focusing on **foundation models for dynamical systems**, is now available, along with the complete code base in **Julia** and **Python** for practical implementation ([arXiv](https://arxiv.org/abs/2505.13192), [GitHub](https://github.com/DurstewitzLab/DynaMix-julia)).

### Accelerating discovery with the AI for Math Initiative
- The **AI for Math Initiative** unites five prestigious research institutions to leverage AI in advancing mathematical research, aiming to identify and solve complex mathematical problems through innovative collaboration.

### Into the Omniverse: Open World Foundation Models Generate Synthetic Worlds for Physical AI Development
- **NVIDIA's updates to Cosmos enable the generation of **physically based synthetic data** at scale, crucial for training **physical AI** models in dynamic environments.

### BambooKG: A Neurobiologically-inspired Frequency-Weight Knowledge Graph
- **BambooKG** introduces a **frequency-weighted knowledge graph** that enhances multi-hop reasoning by incorporating non-triplet edges, addressing limitations in traditional knowledge graphs.

### ALDEN: Reinforcement Learning for Active Navigation and Evidence Gathering in Long Documents
- **ALDEN** enhances **vision-language models (VLMs)** by enabling them to actively navigate and gather evidence from long, complex documents, overcoming limitations of traditional fixed reasoning templates.