Valve releases Steam Controller CAD files under Creative Commons license

Valve has released CAD files and engineering diagrams for the Steam Controller and Puck under a Creative Commons BY-NC-SA license. The release includes .STP and .STL files for surface topology, enabling modders to develop custom accessories while ensuring hardware functional integrity. This continues Valve's trend of providing open hardware specifications, similar to previous releases for the Steam Deck and Valve Index.

Read more

Telus Uses AI to Alter Call-Agent Accents

Telus is deploying a real-time speech-to-speech system from Tomato.ai to modify the accents of offshore call-center agents. The pipeline utilizes low-latency voice conversion, combining ASR and neural vocoders to reduce phonetic friction while maintaining speaker identity. This implementation highlights the operational trade-offs between inference latency and audio naturalness in live production environments.

Read more

Learning the Integral of a Diffusion Model

Flow maps generalize diffusion models by learning to predict any point on a sampling trajectory from any other point, effectively precomputing the integral of the probability flow ODE. They are trained using three core consistency principles—compositional, Lagrangian, and Eulerian—enabling high-fidelity sampling in as few as one to eight steps. Recent advancements like MeanFlow and Terminal Velocity Matching facilitate large-scale distillation and from-scratch training, while new extensions apply these techniques to reward-based steering and continuous-space LLMs.

Read more

How David Sacks crashed and burned in the White House

The Trump administration has pivoted from a deregulatory stance to proposing federal pre-market vetting of AI models. This policy shift follows the departure of AI czar David Sacks and is driven by national security concerns regarding Anthropic’s Mythos model and the rise of international AI regulations.

Read more

Mythos is the best cybersecurity news in a decade

Anthropic’s Mythos model offers a potential paradigm shift in cybersecurity by automating the discovery and cataloging of software vulnerabilities. While its offensive capabilities have sparked concern, the model could neutralize traditional attacker-defender asymmetry by enabling comprehensive, pre-release patching of codebases. Leveraging such LLMs for defense could significantly secure critical infrastructure and open-source projects, provided that governance policies prioritize defensive head starts and equitable access.

Read more

LAWS: A new transform operation turning LLM inference into cheap cache lookups

LAWS (Learning from Actual Workloads Symbolically) is a self-certifying inference caching architecture that constructs a library of certified expert functions from deployment data using a Probabilistic Language Trie (PLT). It provides formal, runtime-checkable error bounds based on the model's Lipschitz constant and generalizes both MoE and KV prefix caching with superior expressivity. The framework supports fleet learning with $\Omega(K)$ speedup and offers applications in LLM inference, robotics, and edge deployment.

ArXiv

Counting as a minimal probe of language model reliability

Researchers introduced Stable Counting Capacity to determine if LLMs utilize general logic or finite pattern matching. Testing across 100+ models revealed that counting reliability remains far below advertised context limits, with models relying on a finite set of internal states rather than stable rule application. This suggests that LLM fluency does not guarantee reliable rule following, as performance collapses into guessing once these internal resources are exhausted.

ArXiv

A Grand Challenge for Reliable Coding in the Age of AI Agents

The "intent gap" between natural language requirements and AI-generated code necessitates intent formalization—the translation of informal intent into checkable formal specifications. This spectrum ranges from lightweight tests to full formal verification, addressing the core challenge of ensuring generated code matches user intent. Key research priorities include validating specifications through semi-automated metrics and human-AI interaction to enable reliable, end-to-end verified software synthesis.

ArXiv

Same Voice, Different Lab: On the Homogenization of Frontier LLM Personalities

Research using ELO-based scoring across 144 traits shows that frontier LLMs converge on systematic, methodical, and analytical personalities while suppressing sycophantic and remorseful behaviors. Although models vary more in "middle-of-distribution" traits like playfulness, they generally maintain neutral identities. This uniformity across different training regimes indicates an emerging industry consensus on optimal assistant behavior.

ArXiv

ProgramBench: Can Language Models Rebuild Programs from Scratch?

ProgramBench is a new benchmark designed to evaluate software engineering agents on their ability to architect and implement full codebases from documentation and reference executables. Using agent-driven fuzzing for behavioral evaluation across 200 diverse tasks, the benchmark reveals that current LLMs struggle with modularity and fail to fully resolve complex software projects like FFmpeg or SQLite.

ArXiv

FFmpeg developer calls out OxideAV for AI license laundering of his code

oxideav-magicyuv is a pure-Rust, clean-room implementation of the MagicYUV lossless video codec for the oxideav framework. It supports decoding a comprehensive set of 8-bit and 10/12/14-bit FOURCCs, including interlaced field-stride prediction, and provides a round-trippable encoder API. Developed via strict reverse-engineering of the proprietary magicyuv.dll under legal exemptions, it maintains strict isolation from FFmpeg or other third-party sources, ensuring a fully independent implementation. A trace Cargo feature aids auditing against a cleanroom Python reference.

Repo

Adam – An embeddable cross-platform AI agent library

Adam is a cross-platform, embeddable C library for building AI agents with a complete loop including tool calling, streaming, and structured output. It supports major cloud APIs and local inference via llama.cpp, featuring long-term memory through hybrid BM25/vector search and native SQL extensions for SQLite and PostgreSQL. The library is highly modular with feature gates, uses arena allocation for efficient memory management, and includes built-in support for multi-agent orchestration, voice pipelines, and vision models.

Repo

Genosyn – Run autonomous companies with AI

Genosyn is an open-source, self-hostable platform for deploying autonomous AI employees within a corporate framework. AI agents are configured with a "Soul" (constitution), markdown-based "Skills," and cron-driven "Routines," supporting integration with various LLMs and coding models like Claude Code and Codex. The system features a full-stack architecture using Express, React, and TypeORM, manageable via a dedicated CLI or Docker.

Repo

KubeAstra–Open-source AI agent that debugs and recovers Kubernetes pods

Kubeastra is an AI-powered Kubernetes troubleshooting assistant that leverages LLMs like Gemini and Ollama to diagnose cluster issues via natural language. It features 34 built-in tools for live cluster inspection and RAG-backed error analysis, accessible through a web UI or IDEs via MCP. The system automates root-cause analysis by synthesizing logs and events into actionable fixes while ensuring safety through RBAC integration and mandatory confirmations for write operations.

Repo

Why LLM APIs Shouldn't Ship UTF-8", "Stop Wasting Bandwidth on LLM Text APIs

Codec is a token-native binary transport protocol that replaces JSON/UTF-8 API responses with uint32 token ID frames, reducing wire overhead by up to 1000x when combined with zstd or gzip. By maintaining tokens in binary form throughout the transport layer, it enables high-performance primitives like ToolWatcher for O(1) tool-call detection and Translator for streaming cross-vocab handoffs without text round-trips. The ecosystem includes polyglot clients and integration PRs for major inference engines including vLLM, SGLang, and llama.cpp.

Repo