HigherEd AI Daily: April 4, 2026 – Google’s Free AI Now Runs on Any Device, OpenAI Ends Sora on April 26, AI Learns While It Reads

If you have been waiting for AI to become accessible enough to use freely at your institution, that moment arrived this week.

Dr. Ali Green

Google Opens Its Best AI to Everyone. Gemma 4 Is Free, Offline, and Legally Unrestricted.

Google DeepMind released Gemma 4 this week, a family of four open-source AI models ranging from phone-sized to desktop-grade. For the first time, Google published the models under an Apache 2.0 license, removing every legal barrier to educational use. The smallest variants run fully offline and handle voice input, text, and images without any internet connection.

Why this matters for your teaching

You can now run a capable AI model entirely within your institution's own systems, with no data leaving your environment and no ongoing subscription cost.

Read more: Google DeepMind Gemma 4 Release

Pulled from: The Rundown AI

OpenAI Is Shutting Down Sora. You Have Until April 26 to Save Your Work.

OpenAI announced this week that it will discontinue Sora, its text-to-video generation tool, on April 26, 2026. The company cited high operational costs and is redirecting those computing resources to enterprise and coding products. The Sora API will also close on September 24. If Sora has been part of your course design or lecture demonstrations, you have a narrow window to export or archive anything you built with it.

Why this matters for your teaching

Alternatives like Runway, Kling, and ByteDance's Seedance 2.0 have already surpassed Sora in quality benchmarks and remain available today.

Read more: The Batch by Andrew Ng (Full Coverage)

Pulled from: The Batch by Andrew Ng

Researchers at Stanford and Berkeley Teach AI to Learn From Long Documents While It Works

A team from Stanford, UC Berkeley, Nvidia, and the Astera Institute published a paper introducing Test-Time Training End-to-End, a method that allows a language model to update its own weights while reading a long document rather than after training. The approach keeps inference speed constant even as documents grow to 128,000 tokens. Early benchmarks show it outperforms standard transformer models on long-context comprehension tasks.

Why this matters for your teaching

This research moves us closer to AI tools that can reliably process full journal articles, textbook chapters, or multi-hundred-page datasets without losing accuracy as the material grows longer.

Read more: The Batch by Andrew Ng (April 3 Issue)

Pulled from: The Batch by Andrew Ng

Try something new today

Prompt of the Day

"I am designing a [course subject] course for [level] students at a college or university. Help me write three measurable learning outcomes for the semester using Bloom's Taxonomy action verbs at the analysis level or above. For each outcome, suggest one formative assessment activity and one summative assignment that would give students meaningful practice and clear evidence of mastery before the final grade."

Tool of the Day

Google AI Studio

Test Google's newest Gemma 4 models directly in your browser with no setup or installation required. A practical first step before deciding whether to run the model locally at your institution.

Free