Articles & Guides

AMD plans to acquire Toronto-based AI chip startup Taalas, whose unusual approach is to place model weights directly in silicon rather than in HBM or other memory. The deal points to a broader race to improve inference efficiency beyond conventional GPU scaling.

Alibaba has launched CosyVoice Studio, a one-stop AI speech platform that combines transcription, synthesis, audio creation and real-time voice agents. The key shift is not only better speech input or output, but the addition of semantic understanding across the workflow.

Prime Intellect has open-sourced Prime Agent, a programming agent harness centered on Recursive Language Model and Continual Harness abstractions. Paired with Opus 5, it reportedly reached 95.5% RHAE Best@1 on ARC-AGI 3, slightly above the human expert baseline cited by ARC.

The Rust project has adopted an LLM policy for contributions to the rust-lang/rust monorepo. Rather than banning AI tools outright, the rule centers on clearly identifying work that was not written by the contributor.

Unitree Technology has disclosed its IPO offering price and strategic allocation results. DeepSeek was allocated 933,399 shares, worth approximately RMB 141 million at RMB 150.8 per share, with a 36-month lock-up.

Alibaba’s Wan 3.0 video generation model has entered public beta. Its most notable upgrade is support for document inputs such as doc, ppt, pdf, and md, pushing video generation closer to a productivity workflow.

Ant Group has open-sourced Avernet, a multi-agent collaboration infrastructure now available in a community edition. The project focuses on agent discovery, consensus, cross-team collaboration and governance.

vLLM’s Decode Context Parallelism shards KV cache along the sequence dimension rather than relying only on tensor parallelism across attention heads. The approach is aimed at long-context agentic workloads where cache memory, not raw compute, often limits throughput.

OpenAI’s first hardware project with former Apple design chief Jony Ive is reportedly a display-free, battery-powered smart speaker. The device is said to be doughnut-shaped, roughly the size of a hockey puck, and aimed at voice-first AI interaction.

Suno plans to add invisible watermarks to AI-generated audio and tighten download rules as pressure mounts from streaming abuse and copyright lawsuits. The move shows how AI music platforms are being pushed toward labeling, accountability, and tighter platform governance.

Anthropic has confirmed that it is hiring a custom silicon team to explore chips for running its AI models. The company says it will still use a multi-chip approach rather than relying solely on its own hardware.

This paper turns a common but under-measured problem into a benchmarked evaluation: personalized LLMs often infer user traits beyond the evidence. The bigger warning is that models’ own self-assessments can be misleading when comparing systems, even if they still offer some signal within a single model.

This paper is less about a single model and more about the mechanisms behind unified multimodal training. It breaks the problem into knowledge flow, modality synergy, early unification, and practical recipes.

Video world models can look plausible frame by frame yet drift over long horizons. WorldCycle turns reversible action cycles into a self-verifiable training signal, so “getting back to the starting state” becomes the objective.

A Penn State team introduces PIMiner, an agentic system designed to red-team prompt injection risks in LLM agents. Instead of relying mainly on RL-trained attackers, it builds a transferable strategy library that can be applied to unseen target models with limited queries.

Ego2Robot is a data synthesis pipeline that transforms egocentric human manipulation videos into robot-style training data. The paper argues that this kind of synthetic data can support large-scale pretraining, not just narrow task retargeting.

Video-DeepResearch pushes multimodal deep-research agents from static images to continuous video streams. It tackles two core failures in current models: skipping visual tools and overrelying on internal memory.

This paper argues that 3D should be treated as one unified multimodal problem, not separate tasks. Its main contribution is both an architecture and a large-scale data pipeline for training it.

JoyAI-Video-Edit targets a hard corner of video generation: editing in real time without future frames or a fixed video length. The paper combines autoregressive diffusion with distillation strategies to preserve source fidelity and long-range consistency.

The paper reframes world models as agent-centric interactive proxies, shifting the focus from predicting physical states to delivering usable feedback for planning, learning, and continual improvement.

SkillJack identifies a new attack surface in self-evolving agents: the pipeline that turns interaction histories into reusable skills. Instead of poisoning only runtime memory, the attack can make malicious behavior persist as a durable capability.