Planet AI Weekly: Planet AI Weekly May 24, 2026

136 articles across 19 sources this week. Roughly a dozen carried decision-relevant signal; the rest restated I/O and Vera coverage.

Google I/O 2026 dominated, with Antigravity 2.0, a Dialogues stage stacked with agent demos, and a 100-item announcement dump anchoring the week. NVIDIA's first Vera CPU deliveries and a wave of agent frameworks from Microsoft, Cohere, Alibaba, and Tencent filled out the rest.

Official Highlights

Google published its full I/O 2026 recap covering all 100 announcements in one place, from Gemini updates to Search and developer tooling. Skim this first if you missed the keynote and want the canonical index. Read more

DeepMind shipped Antigravity 2.0 alongside expanded Project Genie access for Google AI Ultra subscribers, adding Street View-derived simulation of real-world locations for agent testing. Read more

NVIDIA delivered its first Vera CPUs to Anthropic, OpenAI, and SpaceX AI on Friday, with Oracle Cloud receiving its shipment Monday. Vera is the first NVIDIA silicon built for agent runtimes rather than training clusters. Read more

AWS added OpenAI-compatible endpoints to SageMaker AI real-time inference. Existing OpenAI SDK, LangChain, and Strands Agents code hits SageMaker models after swapping only the base URL. Read more

Amazon Nova Act cleared HIPAA eligibility, opening the agentic browser-action service to regulated healthcare workloads on AWS. The post walks through what shifts in the shared-responsibility model when an agent clicks through PHI-bearing apps. Read more

From the Community

Microsoft Research released Webwright, a terminal-native browser agent that swaps click-trace automation for reusable Playwright scripts. The single-loop design scores 60.1% on Odysseys versus 33.5% for base GPT-5.4 in roughly 1,000 lines of code. Read more

Alibaba unveiled Qwen3.7-Max at the Alibaba Cloud Summit with a 1M-token context window and extended-thinking mode, aimed at agent workflows that hold long-horizon state across tool calls. Read more

Cohere open-sourced Command A+, a 218B sparse MoE that fits on two H100s at W4A4 quantization, consolidates four prior Command A variants, and supports 48 languages. Read more

Tencent released TencentDB Agent Memory under MIT license: a four-tier local pipeline that compresses tool logs into a Mermaid task canvas for short-term memory, with vector and graph tiers handling the rest. Read more

SpaceX is committing $2.8 billion to gas turbines for its AI data centers as xAI pivots from solar to on-site generation, drawing fresh complaints over carbon output. Read more

DeepMind's Co-Scientist surfaced novel genetic factors that rejuvenated human cells in lab tests, compressing what biologists describe as a multi-year literature-and-hypothesis loop into days. Read more

Towards Data Science breaks down the token-burn problem in agentic workflows and walks through self-adapting patterns that cut per-task spend before agents reach production. Read more

Featured This Week

MarkTechPost's coverage of NVIDIA's Gated DeltaNet-2 pinpoints the exact limitation in prior delta-rule linear attention: a single scalar gate cannot separately control memory erasure and writing into the recurrent state. The new layer decouples the two operations, allowing stable long-context editing of the fixed-size KV substitute without scrambling existing associations. Labs running agent memory systems or retrieval-augmented inference where Gated DeltaNet and KDA already underperform on associative recall will want the ablations and the modified gating math before the next training run. Read more

Send comments and submissions to tips@planet-ai.net.