Highlights
Top Insights
1. People begin to mirror the style, reasoning patterns, and even opinions of AI tools they use. Users exposed to AI-generated viewpoints tend to shift their own opinions toward those views.
2. In experiments on social issues (e.g., the death penalty), people who used AI to write about topics later expressed opinions closer to the AI’s position than those who didn’t use AI.
Source: AI can ‘same-ify’ human expression — can some brains resist its pull? (Nature)
Top News
1. Google released Gemini Embedding 2, its first natively multimodal embedding model.
2. NVIDIA published Nemotron 3 Super, an open model positioned for high-throughput agentic reasoning.
3. Google announced new Gemini AI capabilities across Docs, Sheets, Slides, and Drive that can generate documents.
4. Zoom launched an AI-powered office suite with Docs, Slides, and so on.
5. Microsoft introduced “Copilot Cowork” for long-running multi-step work.
Additional Insights
1. Hustlers are cashing in on China’s OpenClaw AI craze (MIT Technology Review)
China is experiencing a “gold rush” around OpenClaw, an open-source AI agent capable of autonomously performing tasks on a user’s computer, which has rapidly spread beyond developers to entrepreneurs and everyday users. Early adopters and technically skilled individuals are profiting by selling installation services, training, and customized tools to people eager to use the technology but lacking the expertise to set it up themselves. At the same time, a broader ecosystem of businesses—from cloud providers to API resellers and hardware vendors—is monetizing the surge in demand created by the hype. Many users are motivated by the belief that running AI agents could automate work or create new income streams, fueling a wave of experimentation and speculative activity. The phenomenon highlights how emerging AI tools can quickly spawn secondary markets where the biggest profits often go not to end users but to those selling the infrastructure, services, and expertise surrounding the technology.
2. Generative AI changes how employees spend their time (Ideas Made to Matter)
Generative AI is reshaping work by shifting employees toward higher-value core tasks and away from coordination-heavy or administrative work, as shown in research on developers using GitHub Copilot. In the study, developers spent more time coding and less time on project management, suggesting AI can change the composition of work rather than merely speed it up. The biggest gains appeared among less-experienced workers, reinforcing the idea that AI can accelerate skill development and make junior employees more productive rather than obsolete. At the same time, the findings raise concerns that AI may reduce collaboration and peer interaction, potentially weakening teamwork and the informal learning that comes from shared problem-solving. The broader insight is that companies should treat generative AI as a tool for augmenting capability and learning, while deliberately managing its side effects on collaboration, training, and foundational skill development.
3. Using AI Can Stifle Innovation. But It Doesn’t Have To (HBR)
This article introduces a formal model showing that as the “reuse” of knowledge becomes essentially free via AI, independent exploration falls and organizational innovation flattens. The underlying problem is the erosion of “absorptive capacity”—the ability to evaluate, adapt, and improve ideas rather than simply copy them. When AI makes research effortless, people stop investing in discovering better approaches. The researchers suggest that CEOs must build “calibrated friction” into workflows—requiring independent human attempts before AI is consulted and assessing how people use AI, not just what they produce. The mandate for 2026 is to ensure that AI does not make organizations faster overnight but shallower over time.
4. Your Data Agents Need Context (a16z)
This article explains why most “chat with your data” experiments fail: a lack of proper context. Agents are essentially useless if they cannot decipher vague business definitions or reason across disparate data sources. The solution is the construction of a “context layer”: a multi-dimensional corpus where code lives alongside natural language, capturing the “tribal knowledge” of an enterprise.
Innovation Radar
1. AI Model Releases and Advancements
Sarvam released Sarvam 30B and 105B as open-source reasoning models (Apache 2.0), trained fully in India and optimized across agentic and Indian-language benchmarks (Sarvam).
Microsoft introduced Phi‑4‑Reasoning‑Vision‑15B as a 15B open-weight multimodal model designed for reasoning over text and images (MarkTechPost).
Google released Gemini Embedding 2, its first natively multimodal embedding model (text, images, video, audio, documents) in public preview via Gemini API and Vertex AI (Google).
NVIDIA published Nemotron 3 Super, an open hybrid Mamba-Transformer MoE reasoning model positioned for high-throughput agentic reasoning (NVIDIA).
2. AI Tools and Features
Cursor launched Automations, enabling AI agents to run tasks in the background, including scheduled or triggered workflows (TechCrunch).
AWS launched Amazon Connect Health as an AI agent platform to help healthcare providers with scheduling, documentation, and patient verification (TechCrunch).
OpenAI introduced interactive learning experiences in ChatGPT to help users explore math and science concepts through dynamic visuals (OpenAI). OpenAI announced Codex Security, an agent that builds project context and an editable threat model, then prioritizes, validates, and proposes fixes for vulnerabilities with an emphasis on reducing false positives and triage noise (OpenAI). OpenAI described how prompt injection is evolving toward social-engineering-like attacks and outlined a defensive framing built around constraining dangerous “source-to-sink” pathways and requiring safeguards around sensitive actions or data exfiltration (OpenAI).
Google announced new Gemini AI capabilities across Docs, Sheets, Slides, and Drive that can generate documents, spreadsheets, and presentations from prompts, refine writing, unify tone, and search across Workspace files to produce summaries and answers (TechCrunch).
Zoom launched an AI-powered office suite with Docs, Slides, Sheets, agent-building tools, meeting voice translation, and said its photorealistic meeting avatars will arrive later in March 2026 (TechCrunch).
Anthropic updated Claude for Excel and PowerPoint so add-ins share full conversation context across open files and introduced Skills to make workflows repeatable (Claude). Anthropic rolled out the ability for Claude to generate visuals inline in chat responses, letting the assistant decide when charts or diagrams are useful and enabling iterative updates as the conversation evolves (The Verge).
Microsoft framed Wave 3 as moving Copilot beyond “assistance” toward embedded agentic capabilities across Word, Excel, PowerPoint, Outlook, and Copilot Chat, including “Copilot Cowork” for long-running multi-step work and a multi-model approach (including Anthropic and OpenAI models via Microsoft’s programs) (Microsoft). Microsoft’s security blog argued that agent adoption can create visibility gaps and “agent sprawl,” and positioned “Microsoft Agent 365” as a unified control plane to observe, govern, and secure agents across an organization, integrated with existing admin and security workflows (Microsoft).
Salesforce positioned Agentforce Contact Center as a natively unified contact center offering that combines voice and digital channels with CRM context and AI agents in a single system for self-service, handoffs, and real-time visibility (Salesforce).
3. Other
Hyperscale Power, a startup developing compact solid-state transformers for data centers, is aiming to replace the 140-year-old iron-core transformer technology by making power systems far smaller and more efficient (TechCrunch).
Researchers are exploring manufacturing AI chips on glass substrates instead of traditional silicon, enabling better optical connections, improved heat management, and potentially far more densely interconnected processors for AI workloads (MIT Technology Review).
Leave a Reply