In today’s edition of Data Points, you’ll learn more about:
- Dreamer, a new agent platform that acts like an OS
- A proposal to control access to infectious disease data
- Tiny Aya, a small model family for India, Africa, West Asia, etc.
- Code to Canvas, where Figma teams up with Claude Code
But first:
Claude Sonnet 4.6 matches Opus 4.5 at a much lower price
Anthropic launched Claude Sonnet 4.6, which delivers substantial improvements in coding, computer use, instruction following, and long-context reasoning. Early testing shows developers preferred Sonnet 4.6 over Opus 4.5 59 percent of the time, citing better instruction following, less overengineering, and more consistent execution on multi-step tasks. On OSWorld, the standard benchmark for AI computer use, Sonnet 4.6 achieved human-level performance navigating complex spreadsheets and filling multi-step web forms across real software like Chrome, LibreOffice, and VS Code. A one million token context window available in beta can hold entire codebases. Safety evaluations showed major improvements in resistance to prompt injection attacks. The model is now the default for Free and Pro users on claude.ai and Claude Cowork, and launches immediately via API, major cloud platforms, and Claude Code at $3/$15 per million tokens of input/output. (Anthropic)
Sony builds fractional copyright detector for AI music
Sony Group has created technology that identifies which songs were used to train AI music generators, quantifying each original work’s contribution to the final output. The system works in two modes: when AI developers cooperate, Sony connects directly to their base models; when they don’t, the technology estimates original sources by comparing AI-generated music against existing catalogs. The technology enables a potential royalty distribution system where songwriters, composers, performers, and record labels could collect compensation based on how much their music contributed to AI training and generation. Sony, which owns major music labels and publishing catalogs including half of Michael Jackson’s estate, sees this as both a revenue opportunity and a tool to prevent unauthorized use. The company has yet to announce when it will deploy the system commercially, but envisions AI developers embedding it into their models and content companies using it for license negotiations. (Nikkei)
Former Stripe CTO launches personal agents platform
Dreamer, rebranding from /dev/agents, opened beta access to a platform for building AI agents through natural language without code. The platform positions itself as an operating system for intelligent software, handling infrastructure, data storage, cross-device compatibility, and privacy through isolated VM execution. Its three-layer architecture consists of Tools (data sources and actions, including MCP servers for integrations with services like Google Workspace and GitHub), Agents (user-created applications with triggers, prompts, custom UIs, and databases), and Sidekick (a system-level agent that mediates between components and generates new agents on request). Agents expose Functions that enable composition — for example, one agent can invoke another’s capabilities through Sidekick, creating Unix-style pipelines for AI workflows. The platform monetizes through tool distribution, with partners like Parallel Web Systems offering premium capabilities to developers building on top. (David Singleton)
AI biosecurity framework seeks guardrails on viral datasets
Over 100 researchers from leading universities are proposing controlled access to infectious disease datasets that could enable AI systems to design dangerous pathogens. The framework restricts only high-risk biological data linking viral genetics to traits like transmissibility and immune evasion, while keeping most biological research data publicly available. The concern stems from AI models trained on DNA sequences that can learn to design pathogens if exposed to sensitive datasets, and the reality that once data enters the open web, it cannot be recalled or controlled. The researchers argue that legitimate scientists should retain access while preventing anonymous distribution, and call for basic safety assessments before releasing biological AI models—currently absent in frontier research. The proposal comes as the Trump administration’s Genesis Mission pushes rapid AI development on massive scientific datasets, creating an urgent window to establish protective measures before dangerous capabilities spread. (Axios)
Cohere debuts small LLM family for offline translation and other uses
Cohere launched Tiny Aya, an open-weight model family with 3.35 billion parameters that supports over 70 languages including Bengali, Hindi, Tamil, and other South Asian languages. The models run on consumer devices without internet connectivity, enabling offline translation and local deployment. Cohere trained the base model on a single cluster of 64 H100 GPUs and optimized the architecture for on-device usage, requiring less compute than comparable multilingual models. The company released four variants: TinyAya-Global for broad language support, TinyAya-Fire for South Asian languages, TinyAya-Earth for African languages, and TinyAya-Water for Asia Pacific, West Asia, and Europe. Models are available on HuggingFace, Kaggle, and Ollama for immediate download. For developers in linguistically diverse regions like India, offline-capable models unlock translation and localization applications without dependency on continuous cloud infrastructure. (TechCrunch)
Figma partners with Anthropic on AI-to-collaborative-design pipeline
Figma announced a partnership with Anthropic on Tuesday to launch Code to Canvas, a feature that converts code generated by Claude Code directly into editable designs within Figma’s canvas. The integration allows teams to take working interfaces built through AI code generation, refine them collaboratively, compare design options, and align on decisions without rebuilding from scratch. The partnership reflects confidence that agentic coding tools complement rather than replace design work. However, the feature carries strategic risk: if AI-generated code quality continues improving, teams may eventually skip the design refinement stage entirely, reducing Figma’s role in the development pipeline. The announcement comes as Figma’s stock has fallen 85 percent from its 52-week high of $142.92 reached in August, caught in the broader “SaaSpocalypse” selloff that has hit software-as-a-service companies including Salesforce, ServiceNow, and Intuit. (CNBC)
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 his experience at the Sundance Film Festival, where he engaged with Hollywood professionals to address their concerns about AI, highlighting the cultural differences and the industry’s apprehension towards AI’s impact on jobs and intellectual property.
“This wave of technological change feels forced on them more than previous waves, where they felt more free to adopt or reject the technology. For example, celebrities felt like it was up to them whether to use social media.”
Read Andrew’s letter here.
Other top AI news and research stories covered in depth:
- SpaceX acquired xAI and announced plans for data centers in space, marking a new frontier for AI infrastructure.
- Anthropic updated its flagship model, Claude Opus 4.6, which placed first on the Intelligence Index by tackling more complex problems.
- OpenAI alumni founded Averi to set standards for AI model audits, aiming for more consistent and reliable evaluation practices.
- Researchers introduced Dr. Cabot, an agent trained to diagnose complex conditions, promising more robust medical diagnoses.
A special offer for our community
DeepLearning.AI recently launched the first-ever subscription plan for our entire course catalog! As a Pro Member, you’ll immediately enjoy access to:
- Over 150 AI courses and specializations from Andrew Ng and industry experts
- Labs and quizzes to test your knowledge
- Projects to share with employers
- Certificates to testify to your new skills
- A community to help you advance at the speed of AI
Enroll now to lock in a year of full access for $25 per month paid upfront, or opt for month-to-month payments at just $30 per month. Both payment options begin with a one week free trial. Explore Pro’s benefits and start building today!