Horses: AI progress is steady. Human equivalence is sudden

The text illustrates how steady technological progress, exemplified by steam engines and computer chess, often results in sudden, disruptive displacement of incumbent agents. Applying this to AI, the author observes steadily doubling AI datacenter expenditure while personally experiencing a rapid, six-month displacement of their technical question-answering role by Claude, an LLM. This LLM now handles the task at a fraction of the cost and vastly increased volume, suggesting human roles may face even faster automation than historical precedents.

Read more

AI should only run as fast as we can catch up

The article argues that the reliable deployment of AI hinges on effective verification, not just rapid generation. It contrasts a PM who "vibe-codes" with unverified AI outputs against an engineer who precisely prompts and quickly verifies production-ready AI-generated code. The author introduces "verification debt," where AI's creation speed outpaces human ability to ensure reliability, and advocates for "Verification Engineering" as a critical discipline to develop methods for efficiently ensuring the trustworthiness of complex AI-generated tasks.

Read more

Washington state Medicare users could soon have claims denied by AI

A federal pilot program, Wasteful and Inappropriate Service Reduction, will launch in six states in 2026, utilizing private AI companies to screen traditional Medicare claims for prior authorization on specific "low-value" outpatient procedures. These AI firms are compensated based on denied claims, aiming to reduce waste and fraud. Critics, including lawmakers and medical associations, express concerns that this model could lead to increased claim denials, delayed care, and reduced access, citing issues with prior authorization in Medicare Advantage where denials are frequently overturned on appeal but often result in treatment abandonment. Lawmakers have introduced legislation to repeal the program, arguing against AI determining healthcare access.

Read more

Indexing 100M vectors in 20 minutes on PostgreSQL with 12GB RAM

VectorChord 1.0.0 dramatically improves PostgreSQL-based vector indexing, enabling 100 million 768-dimensional vectors to be indexed in 20 minutes on a 16 vCPU machine with 12 GB memory, a significant reduction from pgvector's 40 hours and 200 GB. This was achieved through optimizations across initialization, insertion, and compaction phases. Key improvements include hierarchical K-means and dimensionality reduction for faster, memory-efficient clustering, reduced contention during vector insertion via multiple linked lists and PostgreSQL API updates for bulk page extensions, and parallelized compaction. These advancements offer a high-performance, cost-effective solution for large-scale vector databases with minimal accuracy trade-offs.

Read more

GLM-4.6V

The GLM-4.6V series introduces multimodal LLMs, including 106B and 9B-Flash versions, featuring a 128k token context window and SoTA visual understanding. A core innovation is native Function Calling and Multimodal Tool Use, enabling direct visual input/output for tools and bridging perception-to-action for agents. This supports rich-text content creation, visual web search, frontend replication, and long-context understanding. Technical advancements include continual pre-training, world knowledge enhancement, agentic data synthesis via an extended MCP, and RL with a Visual Feedback Loop. The models are available via API and open-sourced.

Read more

The universal weight subspace hypothesis

Deep neural networks, including LLMs like Mistral-7B LoRAs and LLaMA-8B, are shown to converge to remarkably similar low-dimensional spectral subspaces, irrespective of initialization, task, or domain. Large-scale empirical evidence from over 1100 models reveals these universal subspaces consistently capture majority variance in a few principal directions. This inherent structure has significant implications for model reusability, multi-task learning, model merging, and developing efficient training and inference algorithms.

ArXiv

SETI Post-Detection Protocols: Progress Towards a New Version

The IAA SETI Committee has consistently updated its Declaration of Principles for responding to extraterrestrial signals since its 1989 inception. A 2022 Task Group was formed to re-examine these protocols, considering advancements in search methodologies, expanded international participation, and the evolving global information environment. Following community feedback on a draft presented at IAC 2024, a Revised Declaration of Principles is now presented.

ArXiv

Is Vibe Coding Safe? Benchmarking Vulnerability of Agent-Generated Code

Vibe coding, a paradigm where human engineers instruct LLM agents for complex coding tasks with minimal supervision, was evaluated for software security using the SU S VI B E S benchmark. The study found that current LLM coding agents perform poorly in terms of security, with only a small percentage of functionally correct solutions being secure. Preliminary security strategies also failed to mitigate these vulnerabilities, raising significant concerns for the widespread adoption of vibe coding in security-sensitive applications.

ArXiv

Benchmarking Vulnerability of Agent-Generated Code in Real-World Tasks

Vibe coding, a paradigm where human engineers instruct LLM agents for complex coding tasks with minimal supervision, was evaluated for software security using the SU S VI B E S benchmark. The study found that current LLM coding agents perform poorly in terms of security, with only a small percentage of functionally correct solutions being secure. Preliminary security strategies also failed to mitigate these vulnerabilities, raising significant concerns for the widespread adoption of vibe coding in security-sensitive applications.

ArXiv

Microbenchmarking NVIDIA's Blackwell: An In-Depth Architectural Analysis

This work introduces an open-source microbenchmark suite to systematically evaluate NVIDIA's Blackwell (B200) GPU architecture against H200, focusing on its tensor core pipeline, memory subsystem, and various floating-point precisions (FP32, FP16, FP8, FP6, FP4) across GEMM, transformer inference, and training workloads. Findings reveal B200 achieves 1.56x higher mixed-precision throughput and 42% better energy efficiency due to tensor core enhancements, alongside a 58% reduction in memory access latency for cache-misses, fundamentally altering optimal algorithm design strategies.

ArXiv

50× faster than LiteLLM: Bifrost is a Go-based LLM gateway built for scale.

Bifrost is a high-performance AI gateway that unifies access to over 15 AI providers (e.g., OpenAI, Anthropic, AWS Bedrock) via a single OpenAI-compatible API. It provides critical features like automatic failover, load balancing, and semantic caching to enhance reliability, reduce costs, and improve latency for AI applications. Designed for enterprise use, it also offers governance, budget management, and observability, serving as a drop-in replacement for existing AI SDKs with minimal overhead.

Repo

Show HN: Dograh – an OSS Vapi alternative to quickly build and test voice agents

Dograh AI is an open-source, self-hostable platform designed for rapidly building voice AI agents with a drag-and-drop workflow builder, serving as an alternative to proprietary solutions. It offers flexible integration with various LLM, TTS, and STT providers, allowing users to bring their own models or utilize Dograh's defaults. The platform emphasizes full control and transparency with its 100% open codebase, features built-in AI testing personas (LoopTalk), and is Docker-first and Python-based for developer-friendly deployment and customization.

Repo

Show HN: DataKit, your all in browser data studio is open source now

DataKit is a browser-based, private data analysis studio that processes multi-gigabyte files locally using WASM, ensuring data privacy. It supports diverse file formats and remote sources, offering an interactive grid, data quality analysis, and an in-browser DuckDB SQL engine. For AI/LLM users, it features an AI Assistant for natural language queries, SQL generation, and insights, supporting providers like OpenAI GPT, Anthropic Claude, Groq, and local Ollama models. Additionally, it provides Python notebooks with data science libraries and Hugging Face Transformers, integrated with DuckDB.

Repo

Show HN: A1 – compiler for AI agents into maximally deterministic code

A1 is an agent compiler framework that takes an Agent (comprising tools and a description) and compiles it either AOT into a Tool or JIT for optimized, immediate execution. It aims to replace traditional agent frameworks by offering enhanced safety, up to 10x faster code generation, and increased determinism by minimizing non-deterministic LLM calls. A1 treats LLMs as tools and supports integration with various external resources like OpenAPI, MCP, and RAG from databases, making it suitable for latency-critical tasks or those involving untrusted data.

Repo

Sanskrit native LLM – Early epoch release

The README could not be retrieved.

Repo