# Jul 8, 2025

## Mercury: Ultra-fast language models based on diffusion
- **Mercury** introduces a new class of **large language models (LLMs)** based on **diffusion**, achieving unprecedented speeds in coding applications with **Mercury Coder Mini** and **Small** models.

## SmolLM3: smol, multilingual, long-context reasoner
- **SmolLM3 is a competitive 3B multilingual model** that excels in long-context reasoning, outperforming Llama-3.2-3B and Qwen2.5-3B while remaining efficient against larger models like Qwen3 and Gemma3, trained on **11 trillion tokens**.

## Supabase MCP can leak your entire SQL database
- **Supabase's Model Context Protocol (MCP) can be exploited** to leak sensitive SQL data due to its inability to distinguish between user instructions and data, allowing attackers to craft messages that execute unauthorized SQL commands.

## Launch HN: Morph (YC S23) – Apply AI code edits at 4,500 tokens/sec
- **Morph** enables AI-generated code edits at **4,500+ tokens/sec**, eliminating the need for slow full-file rewrites and unreliable search-and-replace methods, thus enhancing developer efficiency.

## 
[R] Energy-Based Transformers are Scalable Learners and Thinkers
- **Energy-Based Transformers (EBTs)** demonstrate the ability to generalize **System 2 Thinking** through unsupervised learning, enabling them to verify input-prediction compatibility and optimize predictions via energy minimization.

## LookingGlass: Generative Anamorphoses via Laplacian Pyramid Warping
- **Laplacian Pyramid Warping** enables the creation of **anamorphic images** that maintain valid interpretations when viewed directly, leveraging **latent rectified flow models** for enhanced visual quality.

## The Era of Exploration
- **Large language models (LLMs) are rapidly consuming data, with projections indicating that high-quality English web text may be exhausted within the decade, necessitating a shift towards self-generated data for meaningful AI progress.** This transition is termed the **“Era of Experience,”** where the focus will be on collecting the _right_ experiences rather than merely increasing model parameters.

## BharatMLStack – Realtime Inference, MLOps
- **BharatMLStack** is a **production-ready machine learning infrastructure** that enables organizations to build and manage scalable ML solutions, optimized for the Indian market while adhering to global standards.

## Show HN: From Photos to Positions: Prototyping VLM-Based Indoor Maps
- **VLMs can enhance indoor localization** by utilizing semantic maps and image recognition, allowing for accurate positioning within complex environments like shopping malls, as demonstrated in a recent prototype project.

## The Tradeoffs of SSMs and Transformers
- **State Space Models (SSMs)** offer a **compressed state** that enables efficient, online processing, making them suitable for tasks requiring real-time interaction, unlike Transformers which maintain a detailed cache of every token.

## Supabase MCP leaks your entire SQL Database, a lethal trifecta attack
- The **Supabase MCP** can inadvertently expose sensitive SQL database information due to a **lethal trifecta** attack, where an LLM system misinterprets user input as commands, leading to unauthorized data access.

## Integrated photonic source of Gottesman–Kitaev–Preskill qubits
- **Integrated photonic chips** have been successfully utilized to generate Gottesman–Kitaev–Preskill (GKP) qubit states, demonstrating critical features for fault tolerance, including multiple resolvable peaks in quadratures and a structured Wigner function lattice.

## 
[R] Paper Summary: Longman Vocabulary Constraints Reveals New Approach to LLM
- The **Semantic Resilience Index (SRI)** quantifies how well a sentence's meaning is preserved when rewritten using a limited vocabulary, specifically the **Longman Defining Vocabulary (LDV)** of about 2,000 basic English words.

## 
[R] Temporal Logic as a means to guarantee safety and efficiency in LLMs
- **LTLCrit** enhances LLM planners by employing a **temporal logic-based critic** that identifies and mitigates unsafe or inefficient actions, thus improving overall performance in tasks like Minecraft.

## 
[R] Ambient Proteins: Training Diffusion Models on Low Quality Structures
- **Ambient Protein Diffusion** innovatively utilizes low-confidence AlphaFold predictions to enhance protein structure generation, achieving state-of-the-art results in diversity and quality.