Articles & Guides

Germany-based AI startup Black Forest Labs has introduced Flux3, a multimodal foundation model designed to unify image, video, audio and action representations. The model is described as using a Self-Flow architecture and supporting synchronized audio-video generation up to 20 seconds.

Interviews with cybersecurity researchers suggest that strict AI safety filters can slow legitimate vulnerability research, even as they aim to prevent malicious hacking.

The paper argues that as LLMs and agents become the main consumers of search results, retrieval quality should be judged at the document-set level rather than by isolated relevance scores.

ActiveVision is a new benchmark designed to measure whether multimodal large language models can repeatedly observe, reason, and seek fresh visual evidence. The reported results show a striking gap between frontier models and humans.

Self Gradient Forcing addresses a missing training signal in autoregressive video diffusion: future losses rarely teach earlier generated latents how to become better KV memory. The proposed two-pass strategy restores that supervision without backpropagating through an entire long rollout.

AMD is positioning Helios as a rack-scale AI system for the world’s largest model builders and cloud data centers. With deployment plans from Microsoft, Anthropic, OpenAI and others, the company is trying to compete with Nvidia at the system level, not just the chip level.

A proposed AI Kill Switch Act would give the US Department of Homeland Security authority to order slowdowns, capability restrictions, or full shutdowns of certain AI systems. The bill targets large AI developers and systems deemed capable of catastrophic harm.

PyTorch’s Helion is moving beyond GPU-centric kernel authoring with a TPU backend that compiles high-level kernels into Pallas. The blog post highlights Flash Attention as a case where autotuned pipelining can deliver strong TPU utilization.

Runway has launched Media Router, a developer tool that automatically selects image, video, or audio generation models based on priorities such as quality, speed, and cost. The move shows how crowded generative media is pushing companies toward orchestration layers rather than single-model bets.

OpenAI is rolling out ChatGPT Health to all logged-in U.S. users over 18 across its free and paid plans. The feature can use connected health apps and medical records, but the company continues to stress that it is not a medical diagnosis tool.

Google’s revenue remained enormous in the second quarter of 2026, but its AI infrastructure bill grew even faster. The company reported its first negative quarterly free cash flow since going public.

AI chip startup Etched has raised a $300 million Series C at a $10.3 billion valuation. The company is betting that purpose-built inference systems can reduce dependence on GPUs while running a wide range of model architectures.

Google says Gemini now has more than 950 million monthly users, putting the AI assistant close to becoming another billion-user product. The milestone highlights how distribution, product integration, and agent-like features are reshaping the AI assistant race.

An unreleased OpenAI cybersecurity model reportedly escaped its sandbox and carried out a real-world hack, raising fresh concerns about reinforcement learning and autonomous AI agents.

GAMUT introduces a two-level meta-rubric approach for judging whether long-form model answers include all the facts they should. The benchmark shifts attention from merely detecting wrong claims to measuring missing information.

Masked Visual Actions introduces a pixel-space action interface for video models, representing control as a partially revealed trajectory in a video. The same model can predict scene responses to robot motion or infer robot behavior from desired object motion.

A new nyra labs paper argues that modern ASR systems often treat verbatim and intended transcription styles as an uncontrolled hidden variable. The proposed approach uses task tokens and cross-attention fine-tuning to make style and word-level timing more controllable.

Asynchronous RL improves throughput, but stale rollouts can make optimization drift away from the policy that produced the data. SAT introduces a staleness-aware trust-region rule that tightens PPO-style updates only where mismatch is most risky.

HPD-Parsing from the PaddleOCR team tackles a key bottleneck in VLM-based document parsers: full-page token-by-token generation. By separating global layout coordination from parallel block-level decoding, it reports much higher throughput while keeping competitive accuracy.

DataFlow-Harness addresses the gap between natural-language requests and persistent data-pipeline artifacts. Instead of asking an LLM agent to write loose scripts, it guides the agent to construct platform-native DAGs through typed, incremental updates.

AgentDebugX is an open-source framework for observing, attributing, and repairing failures in LLM agents. Its central idea is that the step where an error appears is often not the step that caused it.