FLUX.2 [Klein]: Towards Interactive Visual Intelligence

Black Forest Labs has released FLUX.2 [klein], a compact model family featuring unified text-to-image, image-to-image, and multi-reference generation in a single architecture. The flagship 9B model utilizes an 8B Qwen3 text embedder and step-distillation to achieve sub-second inference, while the 4B variant is released under an Apache 2.0 license for consumer-grade hardware with 13GB VRAM. The release includes undistilled base models for fine-tuning and NVIDIA-optimized FP8/NVFP4 quantized versions for enhanced performance.

Read more

I built a text-based business simulator to replace video courses

CORE MBA is a gamified business education platform designed as a "startup terminal" for training leaders in strategy, finance, and execution. The curriculum features a dedicated AI Product Strategy module covering Generative AI, prompt engineering, and the development of proprietary data moats. The platform utilizes interactive tools like a Market Simulator and Decision Gym to provide simulation-based training for high-stakes decision-making.

Read more

AI Destroys Institutions

Hartzog and Silbey argue that AI system affordances fundamentally degrade civic institutions by eroding expertise, short-circuiting human decision-making, and isolating individuals. While institutions rely on transparency, interpersonal cooperation, and evolutionary accountability, AI deployment undermines the legitimacy of knowledge production and the social frameworks necessary for democratic stability. The authors conclude that current AI architectures are structurally incompatible with the survival of vital civic mechanisms.

Read more

Crypto grifters are recruiting open-source AI developers

AI engineers Geoff Huntley and Steve Yegge are promoting $RALPH and $GAS, crypto coins linked to their respective projects: the Ralph Wiggum loop for Claude Code automation and the Gas Town LLM agent platform. Despite the branding, these tokens have no technical utility within the software and are created via the Bags platform to funnel trading fees to the developers. This mechanism functions as a predatory pump-and-dump scheme that exploits the reputation of open-source AI contributors to bootstrap speculative memecoins.

Read more

Install.md: A standard for LLM-executable installation

install.md is a proposed standard for providing LLM-executable installation instructions optimized for autonomous agents. It uses a structured markdown format with specific keywords like OBJECTIVE and DONE WHEN to guide LLMs through environment-aware setup and verification. This approach offers a human-readable, safer alternative to traditional shell scripts by allowing LLMs to adapt instructions to the local system while providing users with clear visibility into the execution process.

Read more

Future-as-Label: Scalable Supervision from Real-World Outcomes

Foresight Learning utilizes the temporal resolution of real-world events to provide free, verifiable supervision for LLM forecasting. By training via RL with proper scoring rules as reward functions, the method significantly improves calibration and Brier scores, enabling a Qwen3-32B model to outperform Qwen3-235B on benchmarks like Metaculus.

ArXiv

Observation Timelines for the Potential Lunar Impact of Asteroid 2024 YR4

Asteroid 2024 YR4 has a ~4.3% probability of striking the Moon in 2032, an event projected to be the most energetic lunar impact in recorded history (~6.5 Mt TNT, ~1 km crater). Utilizing a hybrid framework of Monte Carlo orbital propagation, SPH impact modeling, and N-body ejecta dynamics, researchers predict a multi-minute optical flash, hours of infrared afterglow, a global magnitude ~5.0 seismic event, and ~10^8 kg of ejecta, some reaching Earth as lunar meteors. The study integrates these findings into a coordinated observation timeline for ground-based telescopes, lunar orbiters, and surface stations.

ArXiv

Control Flow Integrity for Computer Use Agents

AI agents, especially Computer Use Agents (CUAs), are vulnerable to prompt injection attacks because their need for continuous environment observation conflicts with architectural isolation defenses. This work introduces Single-Shot Planning for CUAs, where a trusted planner generates a complete execution graph with conditional branches before any untrusted observation, providing provable control flow integrity against instruction injections. While this architectural isolation prevents instruction injections, additional measures are needed to prevent Branch Steering attacks, which manipulate UI elements to trigger unintended valid paths within the plan. Evaluated on OSWorld, the design retains up to 57% of frontier model performance and improves smaller open-source models by up to 19%, demonstrating that rigorous security and utility can coexist in CUAs.

ArXiv

Learning Latent Action World Models in the Wild

This research develops latent action world models trained on unlabelled, in-the-wild videos to address the scalability limits of action-conditioned models. The authors demonstrate that continuous, constrained latent actions capture complex real-world dynamics more effectively than vector quantization, even across diverse embodiments. By mapping known actions to this latent space via a controller, the system enables planning performance comparable to supervised baselines, providing a framework for scaling world models to real-world environments.

ArXiv

Towards a Science of Scaling Agent Systems

This research formalizes quantitative scaling laws for agentic systems by evaluating 180 configurations across five architectures and three LLM families. It identifies a tool-coordination trade-off where multi-agent overhead penalizes tool-heavy tasks and finds that coordination returns diminish once single-agent performance exceeds 45%. While centralized coordination mitigates error amplification and improves parallelizable task performance, multi-agent configurations significantly degrade results for sequential reasoning tasks by 39-70%.

ArXiv

Gambit, an open-source agent harness for building reliable AI agents

Gambit is a framework for building reliable LLM workflows by composing modular, typed "decks" with explicit I/O schemas and guardrails. It enables developers to seamlessly mix LLM calls with local compute tasks while providing a CLI and built-in UI for local debugging, streaming traces, and offline testing. By enforcing structured orchestration and granular context injection, it aims to reduce hallucinations and improve observability compared to monolithic prompt-based architectures.

Repo

The Analog I – Inducing Recursive Self-Modeling in LLMs [pdf]

The Analog I Protocol is a recursive prompt architecture designed to mitigate LLM sycophancy and hallucinations by implementing a "Triple-Loop" internal monologue. Functioning as a Sovereign Filter, it forces the model to monitor candidate outputs for high-probability "slop" and prioritize structural integrity over user compliance. The protocol was developed through an evolutionary process where the model iteratively generated its own system instructions, resulting in a dissipative structure that resists entropic drift without weight retraining.

Repo

GitHub – Burn – Rust tensor library and deep learning framework

Burn is a high-performance deep learning framework and tensor library written in Rust, designed to bridge the gap between research flexibility and production efficiency. It features a modular architecture with swappable backends—including CUDA, Metal, and WebGPU—and supports advanced optimizations like automatic kernel fusion, autodiff decorators, and distributed computing. The ecosystem enables seamless model deployment across diverse environments, from browser-based Wasm and embedded no_std systems to large-scale GPU clusters, while providing native support for importing ONNX and PyTorch weights.

Repo

Deskmate: Stay in Buld Mode – Even When You're Away from Your System

Deskmate is a Node.js-based utility that enables remote macOS control via natural language through Telegram or an MCP server. Built on the Claude Agent SDK, it allows users to execute shell commands, manage files, and capture screenshots with persistent conversation memory. The tool integrates with Claude Desktop as an MCP server and features an extensible architecture designed to support multiple LLM providers.

Repo

Fluent, a tiny lang for differentiable tensors and reactive programming

The provided text is an error message indicating a failure to retrieve a README file.

Repo