# Apr 22, 2025

## Daily

### Show HN: Dia, an open-weights TTS model for generating realistic dialogue
- **Dia is a 1.6B parameter text-to-speech model** that generates realistic dialogue from transcripts, allowing for emotion and tone control, as well as nonverbal cues like laughter and coughing.

### AI assisted search-based research works now
- **AI-assisted search-based research has evolved significantly**, with tools like OpenAI's o3 and o4-mini and Google's Gemini 2.5 Pro now delivering reliable, real-time answers without hallucinations, marking a pivotal shift in their utility for users.

### ClickHouse gets lazier (and faster): Introducing lazy materialization
- **ClickHouse's new lazy materialization** optimizes query performance by deferring column reads until necessary, achieving a **1,576× speedup** in specific scenarios without altering SQL code.

### The Future of Compute: Nvidia's Crown Is Slipping
- **NVIDIA's dominance in AI hardware is threatened** as hyperscalers like Google and Microsoft consolidate demand and develop custom silicon, undermining NVIDIA's market position and profitability.

### Fujitsu and RIKEN develop world-leading 256-qubit sup quantum computer
- **Fujitsu and RIKEN have developed a groundbreaking 256-qubit superconducting quantum computer**, enhancing computational capabilities fourfold from its 64-qubit predecessor, which was launched in October 2023 with support from Japan's MEXT.

### CaMeL: Defeating Prompt Injections by Design
- **CaMeL** is a novel defense mechanism designed to protect **Large Language Models (LLMs)** from prompt injection attacks by creating a secure layer that isolates untrusted data from affecting program flow.

### π0.5: A VLA with open-world generalization
- The **π0.5 model** represents a significant advancement in **open-world generalization** for robots, enabling them to perform tasks in entirely new environments, such as cleaning kitchens and bedrooms not seen during training.

### Does RL Incentivize Reasoning in LLMs Beyond the Base Model?
- **Yang Yue** is pioneering **new paradigms** for incentivizing **LLM/MLLM reasoning** and exploring **generalized world models**, aiming to enhance the **generalization of VLA**.

### [R] One Embedding to Rule Them All
- **OmniSearchSage** revolutionizes traditional two-tower architectures by employing a unified query embedding that enhances retrieval of pins, products, and queries through **multi-task learning**.

### Forecaster reacts: METR's bombshell paper about AI acceleration
- **METR's recent findings indicate an exponential acceleration in AI capabilities**, suggesting that models like Claude 3.7 could complete tasks in hours that currently take humans days, with projections extending to year-long tasks by 2031.

### [R] [DeepMind] Welcome to the Era of Experience
- **AI is entering a transformative phase** where agents will achieve superhuman capabilities by learning predominantly from _experience_, rather than relying solely on human-generated data.

### Quantum-assured magnetic navigation with higher positioning accuracy than GPS
- **Quantum-assured magnetic navigation (MagNav)** significantly enhances positioning accuracy, achieving up to **46x better error** than traditional velocity-aided inertial navigation systems (INS) in airborne trials, with a final accuracy of **22m** or **0.006%** of flight distance.

### Can a single AI model advance any field of science?
- **AI's transformative potential** is being harnessed at Los Alamos National Laboratory (LANL) to tackle complex scientific challenges, with projects like ArtIMis aiming to develop foundation models that can generalize across various scientific domains.

### [P] How to measure similarity between sentences in LLMs
- **Measuring similarity** between sentences in LLMs reveals that traditional methods like **cosine similarity** yield misleadingly high scores, indicating a need for more nuanced approaches to understand internal representations.

### Making Brain Waves: AI Startup Speeds Disease Research With Lab in the Loop
- **BrainStorm Therapeutics** leverages **AI-powered computational drug discovery** and **organoids** to accelerate the development of treatments for neurological disorders, aiming to reduce the **93% clinical trial failure rate** associated with traditional methods.