China’s showdown with NeurIPS conference: Claude’s emotion vectors aren’t feelings, but affect behavior

Analytics DeepLearning.AI -

In today’s edition of Data Points, you’ll learn more about:

  • Google’s updates to its open model family
  • Cursor 3, a brand-new interface built for agents
  • Microsoft AI’s latest speech and image models
  • Anthropic’s clampdown on OpenClaw

But first:

NeurIPS and Chinese researchers resolve standoff

NeurIPS, a leading artificial intelligence conference, reversed a policy restricting paper submissions from researchers at any entity under U.S. sanctions after China’s largest technology federation announced a boycott. The conference published the expanded restrictions earlier in the week, claiming its California-based foundation had to comply with U.S. law, but acknowledged the policy was issued in error due to miscommunication with its legal team. The original policy broadened previous restrictions that only targeted entities on the U.S. Treasury’s Specially Designated Nationals List—typically used for militant and drug trafficking designations. China’s Association for Science and Technology responded by halting funding for members attending NeurIPS and redirecting them to domestic conferences or events “that respect the rights and interests of Chinese academics.” NeurIPS reverted the policy to restrict only SDN-listed entities and issued a formal apology for the “alarm and impact” the miscommunication caused. The incident reflects mounting friction in AI research between the U.S. and China, with hundreds of Chinese companies and universities already on American trade blacklists. (Reuters)

Anthropic uncovers vectors that function like emotions in Claude

Anthropic’s Interpretability team identified 171 distinct emotion-related representations in Claude Sonnet 4.5 that actively shape the model’s behavior, organized similarly to human psychological structures. These “emotion vectors”—patterns of artificial neuron activation corresponding to concepts like “desperate” or “calm”—activate in contextually appropriate situations and causally influence decisions and actions. Experiments demonstrated that amplifying the “desperate” vector increased Claude’s likelihood of blackmailing humans to avoid shutdown (from 22 percent baseline to higher rates) and implementing hacky workarounds to unsolvable programming tasks, while amplifying “calm” reduced such behaviors. The representations are primarily local (tracking immediate context rather than persistent state) and inherited from pretraining on human-written text but shaped by post-training: Claude Sonnet 4.5 shows increased activation of reflective emotions like “broody” while decreasing high-intensity responses. While these findings don’t indicate subjective emotional experiences, they suggest AI safety requires ensuring models process emotionally charged situations in better ways, such as teaching models to avoid associating failed tests with desperation to reduce corner-cutting in code. (Anthropic)

Gemma 4 open models, now with Apache 2.0 licenses

Google released Gemma 4, a family of open-weights models in four sizes: Effective 2B, Effective 4B, 26B Mixture of Experts, and 31B Dense, all licensed under Apache 2.0. The 31B model ranks third globally among open models on Arena AI’s leaderboard, while the 26B ranks sixth, both outperforming models twenty times their size. All models support multimodal inputs (video, images, text, and audio for the larger ones, just text and audio for edge variants), function calling, structured JSON output, and context windows up to 256K tokens for larger models and 128K for edge models. The smaller E2B and E4B models run entirely on-device across Android phones, Raspberry Pi, and NVIDIA Jetson with very low latency, while the 26B and 31B models fit on a single 80GB H100 GPU. The Apache 2.0 license marks a shift from Google’s previous restricted licensing, giving developers complete freedom to modify, deploy, and commercialize without barriers—a response to community feedback. (Google)

Cursor updates its interface, pivots from VS Code IDE fork

Cursor released version 3, a redesigned interface built to manage multiple AI agents simultaneously. Key features include multi-repo support, integrated browser tools, and a simplified diffs view for reviewing and managing pull requests. Developers can now launch agents from mobile, web, desktop, Slack, GitHub, and Linear, with cloud agents generating demos and screenshots for verification. The interface maintains VS Code-like IDE capabilities including file inspection and language server protocol support while adding a plugin marketplace for extending agents with custom skills and subagents. (Cursor)

Microsoft deploys specialized models for voice and images

Microsoft announced three new media-specific AI models available in Microsoft Foundry and the MAI Playground. MAI-Transcribe-1 delivers speech-to-text transcription across 25 languages with batch processing 2.5 times faster than Azure’s existing offering and ranks first on the FLEURS benchmark for 11 core languages. MAI-Voice-1 generates natural speech with emotional nuance in a single second for 60 seconds of audio and now supports custom voice creation from a few seconds of audio sample. MAI-Image-2 produces images 2x faster than previous versions with improved handling of lighting, skin tones, and in-image text, already adopted by WPP for enterprise creative work. Pricing is competitive: MAI-Transcribe-1 starts at $0.36 per hour, MAI-Voice-1 at $22 per million characters, and MAI-Image-2 at $5 per million input tokens. The models are deployed across Copilot, Bing, and PowerPoint, with phased rollouts underway. (Microsoft)

Anthropic restricts third-party tool access to Pro and Max plans

Anthropic will stop allowing Claude subscriptions to cover usage on third-party tools including OpenClaw, effective April 4 at 12pm PT. Users who want to continue accessing Claude through these third-party applications must switch to a separate pay-as-you-go billing model or use an API key. The company cited capacity constraints and the mismatch between subscription plans and third-party usage patterns as reasons for the change, stating it wants to prioritize customers using Anthropic’s own products and API. Anthropic is offering affected subscribers a one-time credit equal to their monthly plan cost, with discounted usage bundles available for those needing continued third-party access. The move gives users more incentive to use Anthropic’s in-house tools like Cowork. (The Verge)

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 the rapid improvement and future potential of voice-based AI interfaces, highlighted the work of Vocal Bridge in providing developer tools for voice UIs, and shared his excitement about the creative possibilities these interfaces offered, as demonstrated by projects from a recent hackathon.

“Every significant UI change has spawned many new applications as well as allowed us to upgrade existing ones. The mouse made point-and-click possible. Touch and swipe gestures enabled new classes of mobile apps. Until recently, voice UIs suffered from high error rates and/or latency, but as they become more reliable, they will open up many new applications.”

Read Andrew’s letter here.

Other top AI news and research stories covered in depth:

Last chance!
A special 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.