AI NewsSeptember 17, 2026

Nunchux AI released VC-Attention, a training-free technique that speeds up the attention step in video-generating AI models without retraining or losing quality, cutting total clip generation time by up to 70%.
Nunchux AI released VC-Attention on September 17, a technique that speeds up attention, the costliest step in video-generating AI, without retraining the model. It fixes two problems: outlier values that hurt accuracy at low precision, and a slow softmax step that bottlenecks the pipeline.
On an NVIDIA B200, it runs attention 1.6x faster than standard FlashAttention while staying more accurate than rival low-bit methods. Tested across models like MiniMax H3 and HunyuanVideo, it cuts total clip generation time by up to 70%.
Key Capabilities:

• Two fixes in one: V-Smooth cuts value-quantization error; ExpCast-FP8 removes the softmax bottleneck.
• Real speedups: 1.6x faster attention on NVIDIA B200, and up to 3.6x on workstation GPUs.
• No retraining needed: Drops into existing video models like Wan2.2, HunyuanVideo, and MiniMax H3.

Grok Bot added secure 1Password integration, letting AI agents pull logins from your vault with approval required for each autofill, so the bot never sees the raw password.
Grok Bot added secure 1Password integration on September 16, letting agents pull logins from your vault to sign into tools like Gmail or LinkedIn. Every autofill needs your approval, so credentials never pass through the bot's memory or logs.
This fills a gap that's dogged browsing agents: how to log in somewhere without handing the model your actual password. Developers are calling it overdue, since agents that click through the web still need real credentials to reach anything useful.
Key Capabilities:

• Approval-gated: Every autofill requires explicit user approval before it happens.
• No exposure: The bot never sees or stores the actual password, only the fill action.
• Wide reach: Works across tools like Gmail, LinkedIn, and X that lack clean API access.

xAI added cross-session memory to Grok Build, its terminal coding agent, so it keeps notes on conventions and decisions and reads them back automatically in later sessions.
xAI added cross-session memory to Grok Build on September 16. The agent records conventions, decisions, and project facts as markdown notes, and later sessions read them before touching related code. Commands like /memory and /dream let you browse or reorganize notes.
In xAI's own demo, a user corrects Grok Build after it runs the wrong test command, and the agent remembers the fix for future sessions in that project. Each project gets its own memory, plus a global one for standing preferences.
Key Capabilities:

• Learns from corrections: Fixes you teach it in one session get remembered in the next.
• Per-project notes: Each project keeps its own memory, isolated from other codebases.
• Global overrides: A separate global memory for standing preferences always takes priority.

OpenRouter and OpenCode launched Union Alpha, an anonymous "stealth" model with a 262K context window and strong early coding benchmarks, free to use for about a week.
OpenRouter and OpenCode launched Union Alpha on September 16, a stealth model with no named lab, 262,144 tokens of context, image input, and tool calling. It's free for about a week, and the provider says prompts aren't used for training.
Early tests put it around 51% on Terminal-Bench v4 at roughly $1.50 a task, competitive with pricier frontier models. A widely shared claim of 74% on DeepSWE, matching GPT-6 Astra, comes from an unverified summary, so treat that number with caution.
Key Capabilities:

• Huge context: 262,144 tokens, enough to hold entire codebases in one prompt.
• Free preview: No cost to use for roughly a week, with zero data retention for training.
• Anonymous origin: No confirmed developer; community speculation points to Chinese labs, unconfirmed.

Anthropic merged Claude Cowork and chat into one Claude, so a single conversation can answer a quick question or take on a full report, even after you close your laptop, with new Docs and Slides tools in beta.
Anthropic merged Claude Cowork and chat into one Claude on September 16. There's no mode to pick: ask a quick question or hand over a report, and Claude decides how much work it needs, continuing even after you close your laptop.
Claude Docs and Claude Slides also launched today, letting you draft a document or presentation right in the conversation and export as PowerPoint or PDF. Claude Design now works inside any chat too, folding three separate products into one.
Key Capabilities:

• No more mode-picking: One conversation handles quick questions and multi-step delegated work.
• Docs and Slides in beta: Draft documents and presentations in-chat, export as PowerPoint or PDF.
• Runs after you leave: Delegated work continues in the cloud even after you close your laptop.