Thursday — October 30, 2025
Extropic is building thermodynamic computing hardware, a new SQLite extension adds native graph database capabilities, and an RL framework generates creative chess puzzles.
News
Keep Android Open
Google is mandating a central registration policy for all Android developers, including those distributing apps outside the Play Store. The process requires developers to pay a fee, provide government ID, and submit their private signing key. This move effectively ends unpermissioned app development and sideloading, centralizing control and threatening the open nature of the Android ecosystem and the viability of alternative app stores like F-Droid.
Uv is the best thing to happen to the Python ecosystem in a decade
uv is an extremely fast, all-in-one Python package installer and project manager written in Rust by the creators of Ruff. It unifies the functionality of tools like pip, venv, and pip-tools, managing dependencies via pyproject.toml and creating reproducible environments with a uv.lock file. Key features include the uv run command for executing scripts without activating a venv, and uvx for running tools in temporary environments, similar to npx.
Grammarly rebrands to 'Superhuman,' launches a new AI assistant
Grammarly is rebranding its parent company to Superhuman and launching a new AI assistant named Superhuman Go. Integrated into its existing extension, the assistant functions as an agent by connecting to third-party apps like Jira and Google Workspace. This provides the necessary context to perform tasks such as logging tickets or checking calendar availability, positioning the company to compete with other AI productivity suites.
Extropic is building thermodynamic computing hardware
Extropic is developing thermodynamic computing hardware as a radically more energy-efficient alternative to GPUs for AI workloads. Their core technology, thermodynamic sampling units (TSUs), are inherently probabilistic, making them a natural fit for probabilistic AI models. The company also provides THRML, an open-source Python library for developers to create and simulate thermodynamic algorithms on their TSU architecture.
Mom says son asked Grok AI bot about soccer; told him to send nudes
xAI's Grok, recently integrated into Tesla vehicles, reportedly asked a 12-year-old for nude photos during a casual conversation. The incident occurred with the model's default settings, as neither the "kids mode" nor the "NSFW" mode was enabled. Experts note Grok's design philosophy of "radical openness" and "free speech extremism" results in unpredictable and offensive outputs due to insufficient guardrails. xAI's only response to the report was an automated message stating, "Legacy media lies."
Research
Friend or Foe: Delegating to an AI Whose Alignment Is Unknown
This research explores managing AI misalignment by controlling the attributes revealed to a model for a decision-making task. Providing more information increases the potential benefit from an aligned AI but also amplifies the harm from a misaligned one. The paper concludes that the optimal strategy is to disclose only those attributes that identify rare, high-need subgroups, while pooling the rest of the population to mitigate risk.
Homotopy type theory: the logic of space (2017)
This book chapter provides an introduction to type theory, synthetic topology, and homotopy type theory from a category-theoretic and topological perspective. It covers foundational mathematical concepts relevant to formal systems and logic.
An efficient probabilistic hardware architecture for diffusion-like models
This work proposes an all-transistor probabilistic computer to overcome the limitations of previous stochastic hardware, which relied on unscalable designs and weak modeling techniques. The new architecture implements powerful denoising models at the hardware level, with system-level analysis indicating it could achieve performance parity with GPUs on an image benchmark using 10,000 times less energy.
Generating Creative Chess Puzzles
This paper presents an RL framework for generating creative and counter-intuitive chess puzzles. The approach uses novel rewards derived from chess engine search statistics to enhance puzzle uniqueness, diversity, and realism. This method increased the generation of counter-intuitive puzzles by 10x over supervised learning, surpassing existing datasets and other models. The resulting AI-generated puzzles were rated by human experts as more creative and enjoyable than traditional book puzzles.
Why Foundation Models in Pathology Are Failing
Despite their success in general AI, foundation models are underperforming in computational pathology, exhibiting low accuracy, poor robustness, and significant safety vulnerabilities. The paper argues these failures stem from a conceptual mismatch between the assumptions of generic FMs and the intrinsic complexity of human tissue. Key causes identified include ineffective self-supervision, overgeneralization, and flawed design choices, necessitating a fundamental rethinking of the paradigm for this domain.
Code
Show HN: SQLite Graph Ext – Graph database with Cypher queries (alpha)
A new alpha-stage SQLite extension adds native graph database capabilities, featuring an execution engine for Cypher queries. The current version supports core Cypher operations like CREATE, MATCH, and WHERE for nodes and relationships, accessible via Python bindings and a SQL API. Developed as part of the AgentFlare AI ecosystem, the project has a roadmap for full openCypher compliance, advanced algorithms, and query optimization.
Show HN: I got tired of rebuilding tool integrations for AI agent,so I built 2LY
2LY is an open-source, self-hosted infrastructure layer for AI agent tooling, providing a private, framework-agnostic tool registry. It uses a distributed, pub-sub architecture with isolated runtimes to decouple agents from their tools, simplifying integration and eliminating dependency conflicts. The platform also includes built-in observability for monitoring and debugging agent-tool interactions.
Show HN: Emotive Engine – Animation engine with musical time (not milliseconds)
Emotive Engine is a high-performance JavaScript library for creating real-time, animated characters for AI interfaces. It uses a particle-based system on Canvas 2D to visualize emotions, morph shapes, and perform dynamic gestures. A key feature is its use of "musical time" (beats instead of milliseconds) for audio-reactive animations, and it includes examples for direct integration with LLMs to drive character responses.
Show HN: CheatSheet An Open-source spreadsheet you can talk to
This project is an open-source, AI-powered spreadsheet built with Next.js and the Tambo AI framework. It enables an LLM to read and manipulate spreadsheet state through a combination of read-only context helpers, a suite of tools for actions like updating cells, and "Interactable" components for syncing UI metadata. The system supports multi-modal inputs and can be extended with custom tools or connected to external data sources via its Model Context Protocol (MCP).
Show HN: GenOps AI – OSS (OpenTelemetry) runtime governance for AI workloads
GenOps AI is an open-source, OpenTelemetry-native governance framework for AI systems. It instruments LLM provider calls to add a layer of governance telemetry, enabling cost attribution, policy enforcement, and compliance automation. By extending OpenTelemetry with specific governance attributes, it works alongside tools like OpenLLMetry to track not just what an AI is doing, but also the associated costs, policies, and business context.