Zed 1.0

Zed 1.0 replaces the Electron-based model with a custom Rust-based GPUI framework to enable high-performance, AI-native development. The editor supports parallel agents and keystroke-level edit predictions via the Agent Client Protocol, integrating with models like Claude and Codex. Future updates will leverage DeltaDB, a CRDT-based synchronization engine, to facilitate real-time, character-level collaboration between multiple humans and AI agents.

Read more

Mistral Medium 3.5

Mistral has released Mistral Medium 3.5, a 128B dense model with a 256k context window and open weights, optimized for reasoning, coding, and agentic workflows. It powers the new Mistral Vibe remote agents, which enable asynchronous, cloud-based coding tasks with GitHub and Jira integrations, and a "Work mode" in Le Chat for complex multi-step tool use. The model achieves 77.6% on SWE-Bench Verified and features configurable reasoning effort to balance speed and depth.

Read more

Why AI companies want you to be afraid of them

AI companies like Anthropic and OpenAI are increasingly using "fear-based marketing" by claiming new models, such as Claude Mythos, are too dangerous for public release due to existential risks. Critics and researchers argue this doomerism narrative serves to inflate valuations, distract from immediate harms like environmental impact and labor exploitation, and discourage external regulation by positioning labs as the only capable guardians. Technical skepticism persists regarding these claims, citing a lack of industry-standard metrics like false positive rates and potential compute constraints as the actual drivers behind restricted releases.

Read more

He asked AI to count carbs 27000 times. It couldn't give the same answer twice

A benchmarking study of over 26,000 queries across GPT-5.4, Claude Sonnet 4.6, and Gemini Pro models reveals significant stochastic variability in carbohydrate estimation from identical food images. While Claude Sonnet 4.6 demonstrated the highest consistency (2.4% CV), other models exhibited extreme variance and systematic overestimation bias, posing severe clinical risks for insulin dosing. Crucially, the study found that LLM self-reported confidence scores are poorly calibrated and show near-zero correlation with ground-truth accuracy, rendering them unreliable for safety-critical applications.

Read more

"People who don't use AI will be left behind"

The author critiques the narrative that non-AI users will be left behind, arguing instead that over-reliance on LLMs will lead to the atrophy of critical cognitive skills like writing, fact-checking, and the fundamental ability to learn. They advocate for prioritizing human intellectual development and mastering tasks that exceed current AI capabilities to avoid long-term dependency and skill degradation.

Read more

Evaluating CUDA Tile for AI Workloads on Hopper and Blackwell GPUs

NVIDIA’s CuTile provides a Python-based, tile-centric abstraction for GPU kernels, optimizing for Tensor Core and TMA efficiency with minimal code. On Blackwell (B200), CuTile-based fused attention outperforms FlashAttention-2 by 2.5x, though GEMM performance remains at 52-79% of cuBLAS. Despite its productivity gains, CuTile currently lacks the cross-architecture portability of Triton, exhibiting significant performance degradation on RTX PRO 6000 hardware.

ArXiv

Strategic Polysemy in AI Discourse

This paper analyzes the strategic use of language in AI discourse, focusing on terms like "hallucination" and "chain-of-thought" that exhibit strategic polysemy by blending narrow technical definitions with broader, often anthropomorphic, associations. The authors introduce "glosslighting" as the practice of leveraging these intuitive associations while maintaining plausible deniability through restricted technical definitions. This practice contributes to AI hype, attracts investment, and influences public perception, often deflecting epistemic and ethical scrutiny.

ArXiv

LLMs understand flavours without ever tasting anything

FlavorGraph's 300-dimensional ingredient embeddings, trained on recipe cooccurrence and food chemistry, encode tacit culinary knowledge. An LLM-augmented curation pipeline consolidated 6,653 raw ingredients into 1,032 canonical entries, strengthening this recoverable structure. This process identified at least fifteen independently classifiable dimensions spanning taste, texture, geography, food processing, and culture.

ArXiv

Hyperstatistics

Hyperstatistics is a framework for complex systems where Boltzmann-Gibbs statistics fail, preserving the concavity of nonadditive $q$-entropy. The authors derive closed-form $q$-generalized Boltzmann factors for various distributions, showing they consistently reduce to $q$-exponential-type functions. This approach provides a robust mathematical basis for modeling non-standard statistical domains, with applications ranging from turbulence to high-energy physics data.

ArXiv

OpenGame: Open Agentic Coding for Games

OpenGame is an open-source agentic framework for end-to-end web game development, addressing LLM limitations in cross-file consistency and complex state management. It utilizes GameCoder-27B, a specialized model optimized via execution-grounded RL, alongside a "Game Skill" system that leverages reusable project templates and a persistent debugging protocol. The framework is validated by OpenGame-Bench, a new evaluation pipeline using headless browsers and VLMs to assess build health, visual usability, and intent alignment.

ArXiv

HERMES.md in commit messages causes requests to route to extra usage billing

Claude Code is an agentic CLI tool that enables natural language interaction with codebases to automate routine tasks, explain complex logic, and manage git workflows. It supports extensibility through a plugin architecture and integrates directly into terminal, IDE, and GitHub environments. Anthropic emphasizes privacy by excluding user feedback and session data from model training.

Repo

Pi-hosts – Give the Pi coding agent access to your servers

pi-hosts is an extension for the Pi coding agent that provides structured SSH management through named targets, cached host facts, and connection multiplexing. It optimizes LLM performance by significantly reducing token consumption and tool-call latency through the use of typed execution tools instead of raw shell discovery. The system includes a policy-based risk classification engine to guard remote commands and maintains detailed JSONL audit trails for all operations.

Repo

Agent that refuses to run commands without human approval

Fewshell is a self-hosted, cross-platform SSH copilot designed for secure remote infrastructure management and MLOps. It integrates LLMs to suggest shell commands while maintaining a strict human-in-the-loop model, requiring manual approval for every execution. The architecture prioritizes security by redacting secrets from LLM contexts, syncing sessions via SSH tunnels, and supporting various backends including OpenAI, Anthropic, and local Ollama instances.

Repo

LLM-assisted reconstruction of partially decompiled Minecraft 26.1.2

demcstify is a research project utilizing a multi-agent LLM framework to reconstruct Minecraft 26.1.2 source code into bytecode-identical artifacts. The system employs a deterministic bytecode oracle to provide feedback loops for LLM agents, which repair decompilation errors such as malformed generics and control-flow drift. Coordination is handled via an append-only SQLite database to manage parallel agent tasks and ensure verifiable alignment between reconstructed source and original JVM bytecode.

Repo

Snitchmd – Cloudflare-protected URLs into clean Markdown via Docker

snitchmd is a Docker-based CLI tool that converts URLs into clean Markdown optimized for LLM context windows and RAG pipelines. It bypasses anti-bot protections and renders JavaScript-heavy pages by chaining CloakBrowser with rs-trafilatura. By stripping noisy HTML, it significantly reduces token counts while providing features like disk caching and JSON metadata output.

Repo