Nvidia's AVO Agent Achieves Perfect ARC-AGI-3 Score
In what may be the most significant AI benchmark result of the year, Nvidia's Agentic Variation Operators (AVO) system has cleared all 183 levels of the ARC-AGI-3 benchmark — the first architecture to achieve a perfect score on the test designed to measure general reasoning ability beyond pattern matching.
What makes the result even more remarkable is efficiency: AVO used approximately 12% fewer actions than competing systems that only managed partial completion. The architecture combines persistent memory with a supervision loop that detects and redirects stagnant strategies, essentially allowing the system to recognize when it's stuck and try fundamentally different approaches.
The ARC-AGI benchmark series, created by François Chollet, has long been considered one of the more meaningful tests of general intelligence because it requires genuine abstraction rather than memorization. A perfect score doesn't mean AGI has arrived — the benchmark tests a specific type of reasoning — but it signals that agentic architectures with self-monitoring capabilities are pushing past barriers that stumped previous generations of models.
DeepSeek Drops V4-Flash-Vision: Multimodal AI Gets Cheaper and Better
DeepSeek released V4-Flash-Vision-Exp, adding image understanding to its popular mixture-of-experts model family. The results are turning heads: the system surpassed Claude Opus 4.8 on visual reasoning benchmarks while reducing processing costs by 73%.
This continues DeepSeek's pattern of delivering frontier-level performance at a fraction of the cost charged by Western AI labs. The multimodal addition means DeepSeek's model can now process and reason about images alongside text, making it suitable for applications from document analysis to visual question answering.
The release intensifies price pressure across the industry, where the cost per intelligence unit has already dropped roughly 50% across multiple model tiers in August 2026 alone.
Nevada Greenlights Thousands of Robotaxis for Las Vegas
Nevada's transportation regulator unanimously approved permits for Tesla, Waymo, and Uber to operate paid robotaxi services across Clark County, home to Las Vegas. The approvals authorize up to 7,000 fully autonomous vehicles in total — a dramatic escalation from Tesla's earlier interim permit that capped its fleet at just 10 vehicles.
Tesla received the largest allocation at up to 5,000 robotaxis, with Waymo and Uber each authorized for 1,000. Uber plans to operate through partnerships with Hyundai subsidiary Motional and Amazon's Zoox unit.
The decision makes Las Vegas one of the largest autonomous vehicle deployments in the world and comes just weeks after Amazon's Zoox began paid commercial service in the resort corridor, following the first-ever NHTSA commercial exemption for a purpose-built autonomous robotaxi.
Local taxi companies and gig-economy drivers have voiced concern about long-term job displacement, particularly around the busy Strip-downtown-airport corridor.
Nvidia Acquires Poolside in $7 Billion Coding AI Deal
Nvidia is committing $7 billion to acquire Poolside's Model Factory platform — $6 billion in licensing fees plus a $1 billion equity investment — while offering positions to all 109 employees. The deal represents Nvidia's most aggressive move yet into the AI software layer, building out its coding model development capabilities beyond its dominant position in hardware.
Poolside's technology specializes in generating and fine-tuning code-focused AI models. For Nvidia, the acquisition creates a vertically integrated stack: from the GPUs that train models, to the software platform that builds them, to the developer tools that deploy them.
The deal comes at a time when Nvidia is also raising prices on its next-generation Vera Rubin and Grace Blackwell systems by over 15% starting in early 2027, driven by rising memory chip costs — marking the first major hyperscaler sticker shock of the new hardware generation.
Google's A2A Protocol Joins the Agentic AI Foundation
Google's Agent-to-Agent (A2A) protocol has formally joined the Linux Foundation-directed Agentic AI Foundation (AAIF), consolidating the emerging standards for how AI agents communicate with each other. The foundation now counts more than 250 members, including major cloud providers and AI labs.
This is a significant step toward reducing integration friction in multi-vendor agent architectures. As enterprises deploy agents from different providers — coding agents, research agents, workflow agents — they need a common language for these systems to collaborate. A2A under the Linux Foundation's governance gives it the neutrality and governance structure that proprietary protocols lack.
The move comes during a week when agentic AI is clearly accelerating: Salesforce introduced Slack Code for agent-assisted collaborative coding, Cloudflare's Kitesurf browser runtime for agents hit beta, and Binance launched Agent OS for AI-powered trading — all signs that the infrastructure layer for autonomous AI systems is rapidly maturing.