Key AI news

NVIDIA brings local agents to PC and team clusters

NVIDIA shows the local product line in Berlin: RTX Spark for personal agents on Windows PCs, DGX Spark for developers, and DGX Station for teams. Underdog, a desktop agent with fully local processing, is available for RTX Spark. DGX Spark 64GB has been rolling out at partners since Friday, 23.10. NVIDIA Sync will get the Model Launcher in October, plus a software update with Wi-Fi Assistant. DGX Station for Windows is announced for December. Some RTX Spark laptops ship with an unlocked bootloader. That helps you pilot a private agent with no data leakage. Vendor claims.¹

Build AI factories faster and with power in mind

In the keynote NVIDIA introduced DSX POD, a reference design. That means a pre-validated build template for “AI factories” in modular containers. Commissioning is supposed to take 14 days after site prep. According to NVIDIA, DSX can need up to 100 percent less water for cooling and up to 30 percent less power for the same AI performance. These are vendor numbers without independent verification. In parallel, partners announced concrete DSX uses: from Hitachi Energy to Schneider Electric and Siemens Energy. Phaidra reports about 10 percent more compute in the same power budget. If your data center is on the grid, talk early with energy, facilities, and IT together.²³

OpenShell: blueprint for safer agents

NVIDIA outlined OpenShell as an open reference architecture with a kernel sandbox and minimal rights per agent run. The goal: adjust permissions, policies, and monitoring dynamically. For you this means: strictly limit privileges and measure every action before you let agents into production networks. Announced, source details from NVIDIA.²

GPT-6.1 Terra: cheaper mid-range use

OpenAI introduces GPT-6.1 Terra, according to the announcement with stronger visual justification and faster computer control. Price: 1 US dollar per million input tokens and 5 US dollars per million output tokens. An OpenAI employee places Terra as faster and cheaper than Sol for communication, vision, and computer use. With trade-offs on other tasks. The source is a summary with pricing and placement. For you this means: pick Terra for screen automation, form work, and clear text when Sol is too expensive.⁴

Highlights for your workday

• Code security with GPT-6.1 Sol at CodeRabbit: A security agent inspects risks in large repos and suggests fixes. Use the idea as a second opinion before merge, especially on legacy code. Vendor example.⁵
• Greptile tests workflows with GPT-6 Astra: Agents check usage flows, harden security, and cut rework. Borrow the idea and automate your most common end-to-end tests. Vendor example.⁶
• Forum for Facebook Groups: Meta ships Admin AI and Admin Assist in the separate Forum app in the US. You can review posts faster, apply rules, and schedule posts. Useful if you run a community for brand or support.⁷
• MiniMax-M3.1-Flash local: A 2-bit quantization. That means heavily compressed model weights. According to the vendor it keeps 91.5 percent top-1 accuracy and runs from 150 GB RAM. This is for workstations. You can test sensitive prompts locally. Vendor claims.⁸

Tools and updates

• Transformers 5.21: MiniCPM-OCR reads text-dense images like receipts. Pocket TTS generates speech fast on CPU. Uses: capture receipts, create voice snippets. Requirement: Python and Transformers 5.21.⁹
• Seam Cutter: Small web tool that makes a horizontal or vertical cut and removes the middle. Use: black out prompt blocks in long screenshots without relayouting everything.¹⁰
• Governance for agents in n8n: Guide to identity, guardrails, and approvals in autonomous flows. Use: retrofit roles, limits, and approval steps systematically.¹¹
• QD-Rec from Sakana AI: Recommendation approach for diverse options in decisions. Use: explore outreach targets, target customers, or a reading list. Research with code, not a finished product.¹²

Try it in five minutes

Requirement: An existing AI chat of your choice.

1) Copy this short task into the chat: “Review before PR. Task: add safe price calculation and validate inputs. Context: web shop.” Then paste the code:

"""

bug: unsafe evaluation and rounding

import math

def total_price(qty, unit_price, tax_rate_str): tax = eval(tax_rate_str) # unsafe subtotal = qty * unit_price return round(subtotal * (1 + tax), 2) """

Ask for: a) findings triaged by risk, b) concrete fix suggestions, c) two reusable review standards.¹³

2) Check the result: Did the model flag eval as a security risk and name a safe alternative? Was rounding handled cleanly?

3) Outcome: A short list of prioritized risks, concrete patches, and a mini checklist for next reviews. Acceptance criterion: eval is marked critical, and there is a patch without eval.

Sources

1. Bringing the Spark to Europe: Local Intelligence Powers Next Wave of Agents (blogs.nvidia.com)
2. GTC Berlin 2026: Live Updates From NVIDIA CEO Jensen Huang’s Keynote (blogs.nvidia.com)
3. How Companies Are Using NVIDIA DSX to Break the Power Bottleneck (blogs.nvidia.com)
4. GPT-6.1 Terra (simonwillison.net)
5. GPT-6.1 Sol helps CodeRabbit strengthen security across complex codebases (openai.com)
6. Greptile takes bug detection beyond code review with GPT-6 Astra (openai.com)
7. New Tools for the Admins Who Build Communities on Forum (about.fb.com)
8. @unslothai: MiniMax-M3.1-Flash can now run locally!

We made 2-bit GGUF retain 91.5% top 1 accuracy while being 45% smaller. Run on 150GB RAM.

MiniMax-M3.1-Flash is the strongest open model f (x.com) 9. Transformers v5.21.0 (github.com) 10. Seam Cutter (simonwillison.net) 11. @n8n_io: Your agents are already in production. They authenticate as users, call internal APIs, and run workflows autonomous. 🛡 But autonomy needs governance. 👉 Our guide to embedding ide (x.com) 12. @sakanaailabs: 多様な提案を進化させ、意思決定を支える推薦手法「QD-Rec」を開発

ブログ：https://t.co/2AV6dap8Ve 論文：https://t.co/zg1gtm0OlK GitHub：https://t.co/EKNOoEIBEP

どの企業を買収候補として検討するか。どの顧客に新しい製品を提案するか。膨大な論文の中から、次に何を読むか。こうし (x.com) 13. 7 AI code review best practices #deeplearningai #qodo (youtube.com)