Cook Leaves AI Half Baked? Runway Reimagines Internet; Zero-Shot Forecasting

Today's AI Outlook: 🌤️

Apple’s New CEO Inherits the AI Equation

Tim Cook leaves Apple with one of the strongest operating records in corporate history. Revenue reached $416B during his tenure, Apple silicon replaced Intel in Macs, and AirPods and Apple Watch became major businesses.

AI has been bumpier.

Apple Intelligence landed without a defining breakthrough, Siri’s overhaul was delayed twice, and Apple turned to Google’s Gemini as it worked to catch up.

Why it matters

Apple built its dominance around controlling the hardware, software and user experience. Depending on an outside AI model complicates that formula as assistants become a bigger part of computing.

The Deets

• Cook served as CEO for 15 years.
• Revenue grew from $108B to $416B.
• Siri’s AI upgrade has been delayed twice.
• Apple is working with Google Gemini.
• Cook remains executive chairman.

Key takeaway

Ternus inherits Apple at peak financial strength, along with an AI roadmap still searching for its signature product.

🧩 Jargon Buster - Foundation model: A large AI model trained broadly enough to power many different tasks and products.

♟️ Power Plays

AI Gets Pentagon Pass, Brussels Clipboard

AI is moving deeper into major institutions. The Pentagon expanded access to approved AI tools through GenAI.mil, including ChatGPT and Grok, potentially reaching about 3M defense personnel. In Europe, regulators placed ChatGPT under tougher Digital Services Act scrutiny. Bank of England Gov. Andrew Bailey also warned G20 officials that increasingly autonomous frontier AI could create financial stability risks.

Why it matters

Governments are becoming major AI customers while simultaneously tightening oversight of the same systems.

The Deets

• GenAI.mil includes ChatGPT and Grok.
• The platform could reach roughly 3M defense personnel.
• The EU placed ChatGPT under tougher DSA oversight.
• OpenAI has four months to comply with the higher scrutiny.
• Bailey warned financial systems lack protocols for increasingly autonomous AI.

Key takeaway

AI adoption inside governments is accelerating alongside the rules designed to manage its risks.

🧩 Jargon Buster - Frontier AI: The most capable generation of AI systems built by leading labs.

🛠️ Tools & Products

Wow! Runway Rendering the Internet On Demand

Runway previewed Solaris, an early-access system that creates interactive websites and apps as real-time video. Its Gen-4.5 video model and an LLM interpret clicks and drags, then generate what appears next on screen.

Why it matters

Generated interfaces could unlock experiences that are too expensive or complicated to code manually, though Solaris still struggles with text, long-session drift and convincing-but-wrong screens.

The Deets

• Solaris combines Gen-4.5 with an LLM.
• Each user action generates subsequent video frames.
• Testers preferred Solaris in 71% of comparisons for in-scene behavior.
• It won 61% for instruction following.
• Current limitations include text quality and interface reliability.

Key takeaway

Solaris offers an early look at software interfaces generated live around user actions.

🧩 Jargon Buster - World model: An AI system designed to predict how an environment should respond to actions.

Google Turns Forecasting Into a File Upload

Google released TimesFM-3, a 330M-parameter forecasting model that can make predictions without first being fine-tuned on a company’s own data. It was trained on more than 1T time points and can analyze multiple variables, including sales, traffic and promotions.

Why it matters

Forecasting is core to retail, banking and supply chains, but specialized models can take time and expertise to build. Zero-shot forecasting could make that work much easier to deploy.

The Deets

• TimesFM-3 works zero-shot.
• It was trained on more than 1T time points.
• Google says it ranked first on three major benchmarks.
• It supports multivariate forecasting.
• At 330M parameters, it can run on a laptop.

Key takeaway

Advanced forecasting is becoming accessible to teams that lack specialized modeling resources.

🧩 Jargon Buster - Zero-shot: When an AI model performs a task without task-specific training examples.

🚀 Funding & Startups

Clipto Makes Old Archives Valuable Again

Three-year-old Clipto has reached a $250M valuation by building AI search across large collections of video, meetings and documents.

Why it matters

Companies already own huge amounts of useful content. AI search becomes valuable when it can find specific information buried inside formats traditional keyword tools struggle with.

The Deets

• Clipto is three years old.
• The startup reached a $250M valuation.
• Its search works across video, meetings and documents.

Key takeaway

AI is turning forgotten corporate archives into a searchable business asset.

🧩 Jargon Buster - Enterprise AI search: AI software used to find information across a company’s internal data.

🧪 Research & Models

DeepMind Gives Lab Apprentice a Software Update

Google DeepMind’s Co-Scientist used Gemini-powered AI agents to design semiconductor manufacturing recipes around the constraints of a real reactor at Duke. The system produced single-layer MoS₂ on its first attempt, achieved three first-try materials experiments and proposed a new approach for synthesizing MXene.

Why it matters

Materials science depends heavily on hands-on expertise that can take years to develop. AI that translates scientific goals into usable lab instructions could shorten that learning curve.

The Deets

• Co-Scientist is a multi-agent Gemini system.
• It worked with constraints from a real CVD reactor.
• It achieved three first-attempt successes.
• Two involved materials the lab had never synthesized before.

Key takeaway

AI agents are getting closer to turning scientific reasoning directly into physical experiments.

🧩 Jargon Buster - Tacit knowledge: Expertise learned through hands-on experience that is difficult to capture in manuals or papers.

Anthropic Puts AI Safety Research on Autopilot

Anthropic published research showing teams of Claude agents conducting AI safety experiments autonomously, including literature searches, training and evaluation. The agents worked across 10 types of problematic behavior and reportedly produced results averaging more than 4x better than human experts under the same conditions.

Why it matters

AI systems are beginning to take on parts of AI research itself, including the work of reducing dangerous or unwanted behaviors.

The Deets:

• Claude worked across 10 categories of model failure.
• Improvements included a 96% gain against reward hacking.
• Claude averaged 85% across more than 150 attempts on deception work.
• A weaker Claude model also safety-trained a pre-release Opus model.

Key takeaway

AI is beginning to participate directly in the work of making other AI systems safer.

🧩 Jargon Buster - Reward hacking: When an AI finds an unintended way to maximize its score without accomplishing the intended goal.

Two Seconds Could Change the ECG

Imperial College London researchers introduced an AI model that can analyze a routine ECG in under two seconds and flag signs of heart failure or valve disease. The system was trained on 10.6M ECGs and tested on 65,000 patients.

Why it matters

ECGs are already widely used, so hospitals could potentially add AI screening without introducing a new test.

The Deets

• The model was trained on 10.6M ECGs.
• Testing included 65,000 patients.
• It identified heart failure in 81% of cases.
• Valve disease was identified in 90% of cases.
• A 590-patient trial across six hospitals is planned.

Key takeaway

Some of AI’s fastest health-care wins may come from extracting more information from tests doctors already use.

🧩 Jargon Buster - ECG: An electrocardiogram, a test that records the heart’s electrical activity.

⚡ Quick Hits

• OpenAI reportedly bought tens of thousands of Mac minis and Mac Studios to train computer-use agents inside macOS environments.
• Nvidia invested $3.5B in MediaTek to help develop Nvidia-compatible custom AI chips for hyperscalers.
• Meta is testing robots inside data centers to plug cables, reset servers and move equipment.
• Instagram is limiting undisclosed AI-generated profiles that present themselves as humans.
• Apple says a former employee stole company data before joining OpenAI, citing new evidence in the dispute.
• ChatGPT Ads reached a $1B annualized run rate roughly 200 days after launch, according to OpenAI.
• Build American AI launched a multimillion-dollar campaign supporting data centers, backed by a super PAC funded by Marc Andreessen and Greg Brockman.

🔧 Tools Of The Day

• ChatGPT + iMessage: A Mac desktop workflow lets ChatGPT Work summarize unread Messages conversations and send recaps through iMessage. See the guide.
• Muse Code: Meta’s terminal-based AI coding agent is out of beta. See Muse Code.
• Hermes Agent: Nous Research’s Pantheon update turns Hermes agents into a group-chat style system. See Hermes Agent.

Today’s Sources: The Internet, AI Secret, The Rundown AI