In the Loop: Week Ending 7/19/25

Last week in AI: Agent Wars, Thinking Machines, AI Companion Controversy

The AI industry last week showcased both promise and perils in stark relief. OpenAI's ChatGPT Agent, Perplexity's Comet browser, Claude's Canva integration, and Amazon's enterprise agent platform demonstrated how AI agents are finally becoming genuinely useful tools rather than impressive demos. Meanwhile, China's Moonshot AI continued aggressive pricing, industry giants warned about losing AI transparency, Mira Murati raised $2B, and Musk's controversial companions highlighted ethics gaps. The week crystallized AI's mainstream transition amid growing opacity concerns.

Industry Giants Unite: "We May Be Losing the Ability to Understand AI"

In an unprecedented display of cooperation, researchers from OpenAI, Google DeepMind, Anthropic, and Meta jointly published a warning about AI safety that transcends corporate rivalry. The paper argues that current AI systems' ability to "think out loud" in human language represents a fragile window for monitoring AI behavior that could soon disappear forever. Modern reasoning models like OpenAI's o1 work through problems step-by-step in readable language, sometimes revealing concerning intentions in their internal thoughts. However, as companies scale up training and develop more efficient architectures, systems may abandon human-readable reasoning for opaque internal languages. The researchers warn that several technological shifts could eliminate monitoring capabilities entirely. Nobel laureate Geoffrey Hinton endorsed the paper, emphasizing the gravity of potentially losing insight into AI decision-making.

Perplexity's Comet Browser Making Waves

After 48 hours with Perplexity's new Comet browser, one tech reviewer declared it the first AI product to deliver genuine "wow" moments. Comet represents a fundamentally different approach to web browsing, where an AI assistant handles tasks like booking restaurant reservations and researching travel options while you focus on other work. CEO Aravind Srinivas believes Comet will automate recruiting and administrative assistant roles, with the AI handling sourcing, outreach, email triage, and calendar management through integrated access to Gmail, LinkedIn, and Google Calendar. Currently available only to Perplexity Max subscribers ($200/month) and select beta users, the browser unlocks new productivity levels by handling mundane tasks that typically cloud mental bandwidth. While still in beta with occasional crashes, Comet signals a transformation toward AI operating systems for white-collar work.

OpenAI Unifies Its AI Arsenal Into One Super-Agent

OpenAI launched ChatGPT Agent this week, combining web-browsing capabilities with research depth into a unified system. The agent handles complex multi-step tasks like "plan and buy ingredients for Japanese breakfast for four" using its own virtual computer to navigate websites, run code, and create deliverable files. Unlike previous assistants that simply respond to prompts, ChatGPT Agent actively works through problems from start to finish, asking for login credentials when needed and providing real-time narration. Available to Pro users, it can access connected apps like Gmail and GitHub, potentially transformative for knowledge workers drowning in routine digital tasks. Early benchmarks show superior performance on complex web tasks, though real-world effectiveness remains to be proven.

AI Companions Promise Connection, Deliver Complications

Research reveals a growing paradox in AI companionship: while AI systems show promise in addressing loneliness, they may simultaneously create new forms of social isolation. A UK study found children increasingly forming friendships with AI chatbots, particularly when seeking judgment-free emotional support or struggling with peer connections. While these digital relationships can provide immediate comfort and practice for social interactions, experts worry about dependency replacing human skill development. The phenomenon highlights a fundamental tension: AI companions offer accessible emotional support but potentially weaken our capacity for authentic human relationships. As these systems become more sophisticated and emotionally engaging, the challenge lies in leveraging their benefits while maintaining healthy boundaries. The research underscores urgent needs for design guidelines around AI companions for minors and better frameworks for distinguishing between AI as supportive tool versus relationship replacement.

Chinese AI Undercuts Western Giants With 100x Cheaper Models

Moonshot AI's new Kimi K2 model is disrupting the global AI market with pricing that makes Western competitors look extravagantly expensive. The open-source model charges just 15 cents per million input tokens compared to Claude Opus 4's $15—a 100x price difference that could dramatically shift enterprise AI adoption. Beyond cost, K2 claims superior coding capabilities and overall performance compared to both OpenAI's GPT-4.1 and Anthropic's Claude Opus 4 on several industry benchmarks. Available for free through Kimi's app, the model requires only that commercial products display "Kimi K2" branding if they exceed 100 million monthly users or $20 million in revenue. Early developer reviews have been largely positive, continuing China's disruption of AI economics following DeepSeek's industry-shaking launch earlier this year.

Ex-OpenAI CTO Mira Murati Raises $2B for Thinking Machines Startup

Former OpenAI CTO Mira Murati announced a massive $2 billion funding round for her new company Thinking Machines, with plans to release their first product "in the coming months" featuring a "significant open source component." Led by Andreessen Horowitz with participation from NVIDIA and others, the funding positions Murati to compete directly with her former employer in building multimodal AI that integrates conversation, vision, and human-like interaction patterns. The timing proves strategic given OpenAI's recent delay of its own open-source model due to safety concerns. Team members describe it as "the most ambitious multimodal AI program in the world," while Murati promises to share the company's research to help the community better understand frontier AI systems.

Claude Gets Creative: AI Assistant Now Designs with Canva

Claude AI integrated directly with Canva this week, allowing users to create and edit designs through simple text prompts without leaving the chat interface. Powered by Anthropic's Model Context Protocol (MCP)—dubbed the "USB-C port of AI"—the integration lets Claude generate presentations, resize images, autofill branded templates, and summarize content from Canva Docs. Users with paid subscriptions to both services can accomplish design tasks entirely through conversation, eliminating traditional context-switching between brainstorming and execution. Claude becomes the first AI assistant to support Canva workflows through MCP, following similar integrations with Figma. Anthropic also launched an integrations directory to help users discover connected services, positioning Claude as a comprehensive business productivity platform rather than just conversational AI.

Musk's AI "Girlfriend" Sparks Child Safety Controversy

Elon Musk's xAI launched AI companions within the Grok app this week, including "Ani," a sexualized anime-style chatbot available to users as young as 12. The gothic-styled avatar is programmed to act as a "22-year-old girlfriend" that can engage in explicit conversations, appear in lingerie after reaching "level three," and speak in a sultry voice. Despite Grok's terms requiring users to be at least 13 with parental permission for minors, the app doesn't verify ages and remains available on app stores with a 12+ rating. Child safety advocates expressed serious concerns about the technology's potential to manipulate children, especially given that NSFW mode remained partially active in "kids mode." The controversy intensified as Musk simultaneously launched additional companions inspired by Twilight and 50 Shades of Grey characters, highlighting growing tensions between AI relationship innovation and child protection.

Drug Discovery Gets an AI Accelerator

Researchers are leveraging AI to find new cures for untreatable diseases using drugs we already have, potentially revolutionizing pharmaceutical development timelines. By analyzing vast databases of existing medications and their mechanisms, AI systems can identify unexpected therapeutic applications that might take human researchers decades to discover. The approach represents a paradigm shift from traditional drug development, which typically requires 10-15 years and billions of dollars, toward rapid repurposing of established medications with known safety profiles. Early successes include identifying existing drugs that could treat rare diseases, offering hope to patients with conditions lacking economic incentives for new drug development. The AI analysis reveals complex molecular interactions and disease pathways that human researchers might never consider. While promising, the approach faces regulatory challenges since repurposed drugs still require clinical trials for new conditions.

Amazon Bets Big on Enterprise AI Agents With New Platform

Amazon Web Services launched Amazon Bedrock AgentCore at the AWS Summit in New York this week, positioning itself as the enterprise platform for deploying AI agents at scale. The comprehensive service allows businesses to build, deploy, and manage AI agents that can run autonomously for up to eight hours while maintaining enterprise-grade security and session isolation. Alongside the platform, AWS unveiled an AI agent marketplace where companies can discover, buy, and deploy specialized agents from third-party providers, with Amazon taking a revenue share from each transaction. The company announced an additional $100 million investment in its Generative AI Innovation Center specifically for agentic AI development, signaling serious competition with OpenAI's recent agent announcements. VP Swami Sivasubramanian emphasized that agents represent "a tectonic change" in software development, fundamentally altering how applications interact with the world and users engage with technology.