In today’s edition of Data Points, you’ll learn more about:
- Anthropic’s multi-agent harness design
- A new voice agent benchmark
- Arm’s chip design for data center CPU
- U.S. White House’s AI proposal to Congress
But first:
OpenAI unexpectedly shuts down video generator Sora
OpenAI announced it will shut down Sora, its text-to-video generation platform, without providing a reason for the decision. The company said it would share timelines for discontinuing the app and API, along with details on preserving user work, but declined to comment further. The shutdown comes three months after Disney signed a major licensing agreement with OpenAI that included a planned $1 billion investment and rights to generate videos using over 200 Disney, Marvel, Pixar, and Star Wars characters. Disney has now terminated the partnership, with a company representative stating they respect OpenAI’s decision to exit video generation and shift priorities elsewhere. The shutdown eliminates Sora’s video capabilities from ChatGPT and occurs amid ongoing Hollywood tensions over AI copyright infringement—OpenAI’s Sora 2 had sparked industry backlash for its opt-out training model, where IP owners had to proactively request exclusion from the system rather than opt in. (Variety)
Tencent embraces OpenClaw, allows WeChat to control own version
Tencent launched ClawBot on Sunday, integrating open-source AI agent OpenClaw directly into WeChat as a contact, enabling the platform’s over one billion monthly active users to send and receive commands through chat. Users can use the agent to perform tasks such as file transfers and sending emails on their behalf. The move shows intensifying competition among Chinese tech giants to capture the growing AI agent market, which has seen rapid user adoption despite official warnings about security risks. Alibaba launched Wukong last week, an enterprise-focused platform coordinating multiple AI agents for tasks like document editing and meeting transcription, while Baidu released a series of OpenClaw-based agents across desktop, cloud, mobile, and smart-home devices. Tencent’s own agent suite (QClaw for consumers, Lighthouse for developers, and WorkBuddy for enterprises) launched earlier this month, positioning the company’s ecosystem to reach a massive existing user base through WeChat integration. (Reuters)
Anthropic details multi-agent harness for long-running applications
Anthropic developed a multi-agent system that lets Claude build complete full-stack applications over multi-hour sessions without human intervention. The architecture separates three distinct roles: a planner converts simple prompts into detailed specifications, a generator implements features in sprints using React, Vite, FastAPI, and SQLite, and an evaluator tests the running application with Playwright and provides feedback against defined criteria. The critical insight came from frontend design work: when agents evaluate their own output, they consistently overrate mediocre work. A separate evaluator agent tuned for skepticism delivers actionable feedback instead. The system runs continuous iteration loops where the generator refines or pivots entirely based on evaluator feedback, sometimes producing unexpected creative solutions like reimagining a museum website as navigable 3D space in CSS. (Anthropic)
EVA, a new two-dimensional benchmark for voice agents
ServiceNow AI Research released EVA, an evaluation framework that measures both task completion accuracy and conversational experience quality in voice agents through bot-to-bot audio simulations. The framework splits metrics into accuracy (task completion, factual faithfulness, speech fidelity) and experience (conciseness, conversation progression, turn-taking timing). Testing across 20 systems including cascade and audio-native architectures on 50 airline customer service scenarios revealed a consistent tradeoff: agents excelling at task completion typically deliver worse user experiences and vice versa. The open-source framework includes dataset, code, and evaluation judges on GitHub, marking the first benchmark to assess both dimensions through complete multi-turn spoken conversations rather than isolated components. Early results show a tradeoff between accuracy and experience: For example, popular model Whisper-Large has strong accuracy scores, but poor experience ones. (Hugging Face)
Arm partners with Meta for first self-designed chip
Arm released AGI CPU, its first production silicon product in 30 years, designed for agentic AI workloads in data centers. The chip features up to 136 Neoverse V3 cores per CPU, 6 gigabytes per second memory bandwidth per core, and 300-watt TDP, supporting configurations up to 45,000 cores per rack in liquid-cooled systems. Meta co-developed the chip and committed to deploy it in production, along with OpenAI, Cloudflare, Cerebras, F5, and SAP. Arm claims the AGI CPU delivers more than twice the performance per rack compared to x86 platforms and could save ten billion dollars in capital expenditures per gigawatt of data center capacity. Production systems from ASRock Rack, Lenovo, Quanta Computer, and Supermicro are available now, with broader availability expected in the second half of 2026, marking Arm’s strategic shift from licensing IP to competing directly against Intel and AMD. (Arm)
White House proposes federal AI policy preempting state laws
The Trump Administration released legislative recommendations calling for the U.S. Congress to establish a national AI policy framework that would preempt state AI laws deemed to impose “undue burdens” on the industry. The framework centers on seven areas: child protection, community safeguarding, intellectual property rights, free speech protections, innovation enablement, workforce development, and federal preemption of state regulations. The proposal takes no position on whether training AI models on copyrighted material constitutes fair use, stating the courts should resolve the issue. The Administration calls for regulatory sandboxes for AI applications, opposes creating new federal AI rulemaking bodies, and mandates that residential electricity customers not face rate increases from AI data center operations. The framework would preserve state authority over general laws like child protection and fraud prevention, but prevent states from regulating AI development itself, which the Administration characterizes as having inherent national security implications. (The White House)
Still want to know more about what matters in AI right now?
Read the latest issue of The Batch for in-depth analysis of news and research.
Last week, Andrew Ng talked about job insecurity due to AI advancements and geopolitical uncertainties, emphasizing the importance of community and skills as stable foundations for the future.
“Jeff Bezos famously said that knowing what’s not going to change in the next 10 years creates a stable foundation on which to build a business. Many many things in the world will still be the same in 10 years as now. But for individuals who are worried about job security, I would put forward two things that I think will be stable in that timeframe: Community and skills.”
Read Andrew’s letter here.
Other top AI news and research stories covered in depth:
- Iran struck AWS facilities in Bahrain and the UAE, threatening data centers throughout the Persian Gulf region.
- Alibaba’s Qwen3.5, an open-weights MoE model, outperformed bigger models and led vision benchmarks, with sizes starting from less than 1B parameters.
- DeepSeek made its upcoming 4.0 model available for performance testing to Chinese chipmakers, excluding U.S. ones, in a snub to Nvidia.
- Apple’s AToken introduced a single tokenizer for visual media, integrating a multimodal model with a unified encoder for images, videos, and 3D objects.
Attention!
An important event for our community
Andrew Ng and DeepLearning.AI are hosting AI Dev 26 × San Francisco, a two-day conference for AI developers taking place April 28–29 at Pier 48.
Join 3,000+ engineers, researchers, and builders working on modern AI systems.
The program includes top speakers, developer relations experts, and engineers from companies including Google, AMD, Oracle, Neo4j, and Snowflake (and of course DeepLearning.AI), all sharing their latest technologies and explaining how they’re building and deploying AI systems today.
At AI Dev 26, you’ll find:
- Technical talks from engineers building AI systems in production
- Hands-on workshops exploring new tools and techniques
- Live demos from startups and AI builders
- Opportunities to meet other developers and companies in the space
Get your ticket with a special discount!
Data Points is produced by human editors with AI assistance.