16 Oct 2025

Action! AI Video Race Is On; Your New Bae Is Bai; AI Cancer Fighter

Google Upgrades Veo As The AI Video Race Shifts From Hype To Control

Google rolled out Veo 3.1, positioning it as a filmmaker’s tool rather than a viral post generator. The update adds new creative controls - multi-image prompts (up to three for character consistency), start/end frame guidance for scene matching and continuous scene extension for clips up to one minute long.

The model now lives across Flow, Vertex AI, and Gemini, with both fast and standard modes.

After OpenAI’s Sora 2 dominated attention weeks ago, Veo 3.1 feels more like a quiet evolution. It improves realism and smooth transitions but still caps clips at 8 seconds by default, lacks full-HD output, and struggles with audio continuity when stitching segments. In short, Google prioritized production control over viral spectacle - betting that professional creators, not meme-makers, are its long-term users.

🧩 Jargon Buster: A “text-to-video” model like Veo takes a written prompt (e.g. “woman walking through a foggy forest”) and generates a full moving video clip without filming - essentially turning words into cinematic frames.

Read more: The Rundown AI, AI Secret

**OpenAI’s “Adult Mode” Opens The Floodgates For Synthetic Intimacy**

OpenAI will roll out an Adult Mode in December, allowing verified users to engage in erotic or emotionally explicit conversations through ChatGPT. CEO Sam Altman framed it as “treating adults like adults.” Critics called it a dangerous precedent for mental-health risk and synthetic dependency.

The move drags OpenAI into territory previously owned by Replika and Character.AI, where AI companionship meets emotional monetization. Regulators are likely to follow quickly with new “AI intimacy” rules. As one analyst put it: “Altman just drew the line between AI for work and AI for desire.”

🧩 Jargon Buster: “Synthetic intimacy” refers to emotional or sexual relationships with AI systems that simulate human connection — raising ethical, legal, and psychological challenges.

Anthropic Debuts Claude 4.5 Haiku - Small, Fast And Surprisingly Strong

Anthropic launched Claude 4.5 Haiku, the newest lightweight sibling in its Claude 4 family. Haiku runs at triple the speed and one-third the price of its larger Sonnet 4 model while rivaling its accuracy in coding, math and reasoning. It even approaches OpenAI’s GPT-5 on select benchmarks.

At just $1 per million tokens (vs. $3 for Sonnet), Haiku reflects how fast intelligence is getting cheaper. Enterprises can now deploy multiple Haiku “worker agents” in parallel, coordinated by a larger Sonnet 4.5 instance for complex workflows. It’s already accessible to all users through Claude.ai, the Claude Code development tool, and the API.

🧩 Jargon Buster: A “token” is a small unit of text - roughly a few characters or a word - that AI models process. Token pricing determines how much you pay per chunk of generated text or code.

Read more: The Rundown AI, AI Secret

Google, Yale Discover A New Cancer Pathway Using Gemma-Based AI

Researchers from Google DeepMind and Yale University built C2S-Scale 27B, a model based on the Gemma architecture that analyzes cellular data like language.

The AI identified silmitasertib, a drug previously unrelated to oncology, as a compound that helps the immune system better recognize cancer cells. Tests showed a 50% increase in tumor visibility when combined with immune therapies.

Clearly this isn’t just another drug-discovery claim but rather proof that biological models now follow scaling laws similar to language models: the bigger the network, the more novel insights it can uncover. Critics once doubted AI could make true discoveries. C2S-Scale 27B just rewrote that assumption in the lab.

Google’s $15B India Bet Rewires Global Compute Map

Google announced a $15 billion investment in a 1-gigawatt AI data-center campus in Visakhapatnam, India, its largest infrastructure project outside the U.S. Built with AdaniConneX and Airtel, the site will power 1,000 megawatts of compute capacity and serve as a hub for India’s national AI mission.

This effectively shifts a chunk of the global AI supply chain eastward, giving India a new role as the “AI foundry” for the Global South. It also helps Google hedge against U.S.–China trade friction while challenging AWS and Microsoft Azure to match local scale.

Tools & Launches

🪄 Emergent is a new multi-agent app builder that turns plain-language prompts into full web or mobile applications — backend, UI, and hosting included. Think of it as a conversational IDE for non-coders.

🎬 Supercut creates instant, chaptered video messages with captions, summaries and auto-formatting for social or stakeholder updates. Useful for async comms or personalized outreach.

🧪 KaneAI by LambdaTest lets QA engineers describe tests in natural language. Generates cases, executes them and returns structured bug reports. Designed for GenAI-native software teams managing complex test cycles.

🎥 Flask is a solo-built video collaboration suite that simplifies creative workflows with built-in AI rough-cutting, versioning, and team notes — tailored for indie creators.

Quick Hits

MIT’s Recursive Language Models let AI call itself recursively to handle long documents, outperforming GPT-5 mini by 114% on long-context tests.
Apple unveiled the M5 chip, claiming 4× AI throughput for on-device tasks across Mac, iPad, and Vision Pro.
Runway introduced Apps, modular editing tools for dialogue replacement, element removal, and re-shoot generation.
The International AI Safety Report (2025) flagged rising oversight gaps as adoption accelerates.
Meta announced a $1.5B, 1-GW data center in El Paso, Texas, expanding its AI infrastructure footprint.
OpenAI expanded ChatGPT Go to 89 countries, widening access to its low-cost agentic tier.

Today’s Sources: AI Secret, Anthropic Blog, Google, Meta AI, Microsoft Blog, The Rundown AI.