Anthropic launch Claude 3.7 Sonnet and Claude Code that collaborates with Humans

Anthropic has announced Claude 3.7 Sonnet, the first hybrid reasoning model that can switch between near-instant responses and extended, step-by-step thinking. API users can control the model’s reasoning time, balancing speed and accuracy.

Claude 3.7 Sonnet excels in coding and front-end development and maintains the same pricing as its predecessor ($3M per input token, $15M per output token). It is available across Claude Free, Pro, Team, Enterprise, and platforms like Amazon Bedrock and Google Cloud’s Vertex AI.

A major addition is Claude Code, a command-line agentic coding tool that helps with code editing, debugging, test-driven development, and GitHub integration. Early tests show it significantly reduces development time.

Claude 3.7 Sonnet is designed for real-world reasoning tasks beyond academic math and coding competitions. It has outperformed other models in code planning, full-stack updates, and complex workflows, with major endorsements from Cursor, Cognition, Vercel, and Replit.

For safety, Claude 3.7 Sonnet reduces unnecessary refusals by 45% and includes improvements in security, reliability, and resistance to prompt injection attacks.

This release signals a shift toward AI that collaborates more effectively with humans, helping developers and businesses augment their capabilities and streamline complex tasks.

Link to article

AI Agent as a Service(AaaS)

Software as a Service (SaaS) has driven digital transformation, but it still requires human intervention for key decisions. A new paradigm, Agent as a Service (AaaS), is emerging, where AI agents autonomously manage business operations, decision-making, and execution without human input.

Key Benefits of AaaS:

✔️ Real-time AI decision-making
✔️ Automated issue detection, prediction, and resolution
✔️ Dynamic workflow management, reducing inefficiencies

AI-Managed Layers in AaaS:

• Fully Autonomous: AI agents handle runtime, middleware, OS, virtualization, servers, storage, and networking.

• Semi-Autonomous: AI assists with data security, optimization, and application updates but requires human oversight.

Link to post

Google introduces AI co-scientist

The AI co-scientist is a multi-agent AI system built on Gemini 2.0 to serve as a virtual scientific collaborator. It helps researchers generate novel hypotheses, research proposals, and experimental protocols, aiming to accelerate scientific and biomedical discoveries.

The system addresses challenges in scientific research, such as managing the rapid expansion of literature and integrating interdisciplinary insights. Inspired by the scientific method, it employs specialized agents—Generation, Reflection, Ranking, Evolution, Proximity, and Meta-review—to iteratively generate, evaluate, and refine research hypotheses.

Key features include:

• Collaboration: Scientists can interact with the AI by providing seed ideas and feedback.

• Scaling compute: Uses test-time compute scaling to enhance reasoning and output quality.

• Self-improvement: Employs Elo-based auto-evaluation to refine hypotheses over time.

• Expert evaluation: In trials, domain experts found AI co-scientist’s results more novel and impactful compared to other models.

While promising, limitations include the need for improved literature reviews, factuality checks, and broader expert validation. To advance responsible deployment, a Trusted Tester Program will grant research organizations early access for evaluation and feedback.

Link to blog

Microsoft Muse - the first World and Human Action Model

Microsoft Research has introduced "Muse," the first World and Human Action Model (WHAM), which can generate game visuals and controller actions. Developed in collaboration with Xbox Game Studios' Ninja Theory, Muse is open-sourced, with its weights, sample data, and an interactive WHAM Demonstrator available on Azure AI Foundry.

The research was inspired by the rise of transformer-based models like ChatGPT and built using human gameplay data from Ninja Theory’s game Bleeding Edge. The team explored how generative AI could enhance game development, emphasizing ethical data collection and collaboration with creatives.

A key milestone was the WHAM Demonstrator, an interface allowing users to interact with the model. It helped identify three essential capabilities: consistency (realistic gameplay sequences), diversity (varied outputs), and persistency (retaining user modifications). Scaling up training was a challenge, but advances in GPU usage enabled the model's development. The research aims to support game creators by integrating AI into their workflows from the start.