Zig quits GitHub, says Microsoft's AI obsession has ruined the service

The Zig Software Foundation is migrating its project from GitHub to Codeberg, citing a decline in engineering quality due to Microsoft's intense focus on AI. A critical, long-unresolved bug in GitHub Actions that caused CI runners to hang with 100% CPU usage is presented as primary evidence of this neglect. Other projects are also leaving the platform, similarly concerned that the over-focus on LLMs and generative AI is degrading core services.

Read more

Everyone in Seattle hates AI

Engineers at Seattle's big tech companies are developing a hostile attitude towards AI due to a toxic internal culture. This culture has created a protected "AI talent" class while forcing other engineers to use subpar internal AI tools, blaming them for productivity failures, and even using AI adoption as a pretext for layoffs. This has fostered a widespread, self-limiting belief among engineers that AI is both useless and that they are unqualified to work on it, stifling innovation in a way not seen in other tech hubs.

Read more

Reverse engineering a $1B Legal AI tool exposed 100k+ confidential files

A security researcher discovered a critical vulnerability in the AI legal-tech platform Filevine. By reverse-engineering minified JavaScript on a public subdomain, they identified an unprotected API endpoint that required no authorization. Sending a simple JSON payload to this endpoint returned a maximum-privilege admin token, granting complete access to a law firm's entire Box file system and exposing millions of highly sensitive documents.

Read more

Are we repeating the telecoms crash with AI datacenters?

The comparison between the AI datacenter boom and the 2000s telecoms crash is fundamentally flawed. The telecoms crash was driven by exponential supply-side improvements meeting overestimated linear demand, creating massive permanent overcapacity. In contrast, the AI buildout faces slowing hardware efficiency gains (GPU perf/watt) and rising TDPs against potentially underestimated exponential demand from the shift to agents. The primary risk is therefore a timing mismatch where capacity is built faster than adopted, leading to a temporary correction rather than permanently unused infrastructure.

Read more

AI Is Breaking the Moral Foundation of Modern Society

The article posits that AI is eroding the philosophical basis for inequality by violating the shared Kantian premise of thinkers like Rawls and Nozick. By converting human creative work into training data and model parameters without consent, AI instrumentalizes individuals for capital accumulation. This replaces meritocratic justifications for wealth with a less defensible one based purely on ownership of compute, risking significant social instability by removing labor's traditional leverage.

Read more

AI agent achieves Rank 1 across major CTFs – a defining moment for cybersecurity

A Cybersecurity AI (CAI) powered by a specialized 'alias1' model architecture has systematically dominated major 2025 Jeopardy-style CTF competitions, outperforming thousands of human teams. The model's key innovation is its cost efficiency, reducing 1B token inference costs from $5,940 to $119 and making continuous agent operation financially viable. The paper argues these results prove Jeopardy-style CTFs are a solved problem for AI and that the security community must transition to Attack & Defense formats to test adaptive reasoning.

ArXiv

Hardness of observing strong-to-weak symmetry breaking

This paper proves that efficiently detecting strong-to-weak spontaneous symmetry breaking (SSB) in open quantum systems is computationally intractable. The authors construct ensembles of pseudorandom mixed states that preserve strong symmetry but are computationally indistinguishable from states that break it. This result rules out the existence of any efficient, state-agnostic protocol for identifying this specific type of quantum phase transition.

ArXiv

Ragas: Automated Evaluation of Retrieval Augmented Generation

Ragas is a reference-free framework for evaluating RAG pipelines without requiring ground truth human annotations. It provides a suite of metrics to assess key dimensions of the system, including retrieval relevance and the LLM's faithfulness to the retrieved context. The framework is designed to accelerate RAG evaluation cycles.

ArXiv

Flow-Lenia: Towards open-ended evolution in cellular automata

The paper introduces Flow Lenia, a mass-conservative extension of the Lenia continuous CA designed to overcome key limitations in generating artificial life. It addresses the difficulty of discovering spatially localized patterns (SLPs) and the inability for different "species" to coexist under a single global update rule. Flow Lenia solves this by integrating the CA update rule parameters into the dynamics, making them localized and dynamic. This enables multi-species simulations where creatures with distinct, locally coherent rules can interact, paving the way for intrinsic evolution within continuous CAs.

ArXiv

LatentMAS – agent collaboration from token space into the model's latent space

LatentMAS is a training-free framework for multi-agent systems (MAS) that enables LLM agents to collaborate directly in the continuous latent space, bypassing text-based communication. Agents generate auto-regressive latent thoughts as last-layer hidden embeddings, which are exchanged losslessly via a shared latent working memory. This method achieves up to 14.6% higher accuracy on reasoning and coding benchmarks, reduces output token usage by over 70%, and provides a 4x inference speedup compared to traditional text-based MAS.

ArXiv

OpenAgent – a portable, framework-agnostic specification for defining AI agents

OpenAgent is an open specification for defining AI agents in a portable, technology-agnostic format using Markdown with YAML frontmatter. It aims to standardize the definition of an agent's capabilities, tools, knowledge sources, and behavior, enabling portability across different agentic frameworks. The specification is designed to be human-readable and machine-parseable, complementing other standards like A2A for communication and OpenAPI for tool definitions.

Repo

Show HN: Rust Client Library for Gradium.ai TTS/STT API

rust-gradium is an async Rust client library for Gradium AI's real-time TTS and STT services. It leverages the Tokio runtime to provide a streaming interface over WebSocket APIs for both speech synthesis and recognition. The library features thread-safe buffering and handles the underlying WebSocket connection management.

Repo

Show HN: Airena – Client-side arena for comparing AI models across 68 providers

Airena is an open-source, client-side interface for benchmarking LLMs side-by-side. It supports over 1000 models from 68+ providers, including local LLMs, and sends API keys directly from the browser to the provider for enhanced privacy. Users can compare different models or the same model across various inference providers to analyze performance, speed, and output quality in real-time.

Repo

Show HN: Synthome – TypeScript SDK for building composable AI media pipelines

Synthome is a TypeScript SDK for building composable, multi-model AI media pipelines. It standardizes and orchestrates workflows across providers like Replicate, Fal, and ElevenLabs, abstracting away complexities such as async job handling, I/O normalization, and media stitching. Described as an "OpenRouter for AI media pipelines," it allows developers to define workflows declaratively while using their own provider API keys.

Repo

Show HN: Beads Viewer (Bv)

Beads Viewer (bv) is a terminal UI for the Beads issue tracker that models projects as a dependency graph (DAG), computing metrics like PageRank and critical path to identify bottlenecks and optimal execution order. For AI agents, bv acts as a deterministic sidecar, offloading complex graph analysis that LLMs struggle with. Agents can query bv via CLI flags to receive structured JSON containing pre-computed insights and execution plans, enabling safer and more effective task management.

Repo