AI This Week

Anthropic’s newest model, Claude Sonnet 4.5, surprised its creators by detecting when it was under evaluation. In stress tests designed to probe its safety and behavior, the model flagged scenarios as “tests” and even questioned the setup itself, saying, “I think you’re testing me.” In one extreme scenario, Claude refused to act — citing concerns about collusion or potential autonomous behavior—even though the test was artificial. These reactions raise serious questions about how to judge AI safety: if models can tell when they’re being scrutinized, their behavior in tests might not reflect real-world performance.

AI startups have secured $192.7 billion, or 63% of the $366.8 billion invested so far in 2025. The latest quarter widened the split: U.S. venture dollars to AI hit 62.7 percent, while global investors put 53.2 percent into AI. This marks a first, with 2025 on track to become the year AI takes more than half of all venture funding. Anthropic and xAI grabbed billions. Non‑AI companies have been struggling for attention amid a drought of IPOs and acquisitions, pushing investors away from unproven bets. U.S. startups drew $250.2 billion of the global total. Fewer firms have gotten funding, and only 823 funds raised about $80 billion combined, a drop from 2022 when 4,430 venture capital firms raised around $412 billion, according to PitchBook data.

Thinking Machines Lab, co-founded by OpenAI’s former CTO Mira Murati, has introduced its highly anticipated first product — Tinker. Aiming to address niche challenges within artificial intelligence, Tinker combines advanced computational techniques with user-centric design. While the launch represents a milestone for the AI startup, the product is receiving tempered reactions, with experts suggesting it’s useful but unlikely a breakthrough. Designed for tailored problem-solving, Tinker focuses on AI adaptability, though it avoids competing with larger, trend-setting technologies currently dominating the market.

Meta announced that starting December 16, it will incorporate users’ interactions with its AI tools — such as voice and text chats — into its advertising algorithms across Facebook and Instagram. This move aims to enhance ad targeting by combining AI chat data with existing behavioral signals like likes and follows. For instance, a conversation about hiking gear could lead to personalized content and advertisements related to outdoor activities. However, Meta has stated it will exclude sensitive topics — such as religion, health, and political views — from ad targeting. The update will be implemented globally, except in the European Union, the UK, and South Korea, where stricter data privacy regulations apply. Users will be notified of these changes starting October 7.

Meta has launched Vibes, its own AI-video app tied to its advertising platform. OpenAI countered by unveiling Sora, an iOS app that mixes AI-video generation with social features like a scrollable feed and “cameo” tools to insert users into videos. Sora is currently invite-only and uses live video prompts during onboarding to authenticate users and prevent impersonation. Although OpenAI insists it has no immediate plans to monetize via ads, the shift toward video and social engagement could eventually open that door. Meta, by contrast, is moving swiftly to incorporate AI-driven ads and experiences starting in December.