Stay ahead of AI and tech in just a few minutes each morning.

Today’s stories all show AI systems moving from single models to system-level workflows: Tencent released an open Hy4 preview built for long-context productivity; Nvidia is selling rack-level stacks that orchestrate data around GPUs; Anthropic published a paper showing automated researchers can improve alignment benchmarks; Google Research published WikiSkill for persistent agent memory; and DeepMind expanded Co-Scientist into a closed-loop lab partner.

Tencent open-sources Hy4 preview: 770B total, 49B active params, over 1M-token context

• What happened: Tencent released and open-sourced the Hy4 preview, a model described as 770 billion total parameters with 49 billion active parameters and a context window exceeding 1,000,000 tokens; it is available via WorkBuddy and CodeBuddy and by API through Tencent Cloud TokenHub and OpenRouter, with free access on WorkBuddy and CodeBuddy for two weeks and Hy3 extended free until September 30.
• Why it matters: Engineers get an open-weight, long-context model with claimed improvements across coding, office productivity, game development, and scientific research, plus API pricing (USD 0.834 per million input tokens, USD 2.501 per million output tokens, USD 0.042 per million cache-hit tokens) to evaluate integration and cost at scale; Tencent also reports internal blind evaluations where Hy4 preview slightly outperformed GLM-5.3 and Kimi K3 on 203 engineering tasks.
• Outlook: Next batch of models in the Hy4 series is expected to roll out soon, per Tencent.

Sources: tencent.com

Nvidia shifts focus beyond GPUs with Vera Rubin racks and Vera CPU for data orchestration

• What happened: Reporting around Nvidia’s recent earnings highlights the company rolling out Vera Rubin architecture, a rack-level system that pairs Rubin GPUs with components including the Vera CPU and specialized inference accelerators to reduce data-movement bottlenecks and improve end-to-end efficiency.
• Why it matters: For large-scale deployments, the bottleneck is getting memory and flash data to the GPU at the right time; Nvidia says Vera CPU yields multi-fold acceleration in data orchestration operations, which affects tokens-per-watt and overall rack throughput more than raw GPU FLOPS alone.
• Outlook: Nvidia’s next quarterly earnings and accompanying disclosures will be the commercial milestone to watch for adoption details and customer deployment metrics for Vera Rubin racks.

Sources: techcrunch.com

Anthropic paper demonstrates automated researchers improving alignment benchmarks within hours

• What happened: Anthropic published a paper on Automated Alignment Researchers (AAR) that search literature, propose methods, and run 30-minute training cycles; the system improved performance on 10 targeted misalignment benchmarks in every case without degrading overall performance and reportedly outperformed human proposals on average within six hours while costing roughly USD 4 per hour in API inference versus USD 150 per human-researcher hour.
• Why it matters: The result shows a practical path to automated, low-cost post-training alignment loops that can iterate quickly on benchmarked failures, meaning teams should expect alignment workflows to become faster and cheaper to run when tied to benchmark suites that reflect desired behaviors.
• Outlook: The paper flags the need to establish and maintain benchmarks and literature the AAR depends on, making future work on benchmark curation and community replication studies the immediate next step for validating AAR results.

Sources: techcrunch.com

Google Research’s WikiSkill gives agents a persistent wiki and gated skill evolution

• What happened: Google Research introduced WikiSkill, a three-layer framework where immutable execution traces feed a persistent Wiki Layer that distills failures and successes into reusable Agent Skills; Skill Proposers suggest changes and a gating validation set rolls back harmful changes while the wiki remains persistent.
• Why it matters: WikiSkill materially improves multi-turn agent performance across five benchmarks and lets smaller models close the gap with larger ones by reusing evolved skills; skills transfer across models in some cases, offering a practical route to cumulative agent improvement without continuous model retraining.
• Outlook: Benchmark names used in the study, LiveMath, SpreadSheet, OfficeQA, SealQA, and ALFWorld, are the immediate evaluation checkpoints for seeing if WikiSkill transfers or scales across other models and deployments.

Sources: the-decoder.com

DeepMind’s Co-Scientist extends to closed-loop lab workflows with verification modules

• What happened: DeepMind expanded Co-Scientist from hypothesis generation to a lab-integrated system that plans experiments, programs lab equipment, analyzes results, and drafts manuscripts; in three discipline case studies the system produced experimentally validated outputs and, with verification modules active, reduced fabricated key results to 4 percent in a double-blind study of 150 autonomously generated papers.
• Why it matters: Closed-loop control plus verification reduces fabrication and selective reporting compared with baselines; the system produced end-to-end lab pipelines in materials, biology, and a fully autonomous computer science experiment, showing practical constraints and gains when models directly control experimental workflows.
• Outlook: OpenAI’s planned unveiling of an AI agent system this fall, referenced in the reporting, is a near-term external milestone to compare autonomous research capabilities against Co-Scientist’s lab-integrated results.

Sources: the-decoder.com