Hello from The AI Night,
Today in AI:
Google Launches Free Full Length SAT Practice Exams in Gemini App
Anthropic Releases a New "Constitution" for Claude AI Models
Google & Khan Academy Set to Bring Gemini into AI Tutoring Tools
Image Source: Google blog
Here's the deal: Google launched practice tests SAT directly inside the Gemini app, starting with the SAT. The feature is live now and available at no cost, with content sourced from The Princeton Review.
The Breakdown:
Full-length, on demand SAT exams built into Gemini's interface
Content grounded in Princeton Review's vetted test prep material
Immediate scoring with feedback on strengths and weaknesses
Users can ask Gemini to explain any incorrect answers in real time
Generates customized study plans based on identified knowledge gaps
Additional standardized tests planned for future release
Announced at the BETT education technology conference
The bigger picture: This positions Gemini as a direct competitor to paid test prep services by bundling high quality SAT preparation into a free consumer product. For students, it removes the cost barrier to structured exam prep. For Google, it's a sticky use case that brings younger users deeper into the Gemini ecosystem during a high stakes moment in their lives college admissions.
Image Source: Anthropic blog
Here's the deal: Anthropic released a new, comprehensive constitution for Claude, a foundational document that defines the AI's values, priorities and behavioral guidelines. It's published under Creative Commons CC0, meaning anyone can use it freely.
The Breakdown:
Shifts from a list of standalone principles to a holistic document explaining why Claude should behave certain ways, not just what to do
Written primarily for Claude itself intended to help the model generalize and exercise judgment in novel situations
Establishes a clear priority order: safety > ethics > Anthropic guidelines > helpfulness
Includes "hard constraints" (absolute rules) for high stakes behaviors like bioweapons related requests
Claude uses the constitution to generate its own synthetic training data
Sections cover helpfulness, ethics, safety and Claude's uncertain nature (including questions of consciousness)
The bigger picture: This is a transparent move with teeth, it lets researchers, competitors and users see exactly what behaviors are intended vs. bugs. As AI influence grows, publishing training principles openly sets a precedent others may follow or critique.
Google x Khan Academy
Google & Khan Academy Set to Bring Gemini into AI Tutoring Tools
Image Source: Google blog
Here's the deal: Google announced a partnership with Khan Academy to integrate Gemini models into new AI tutoring tools focused on reading and writing skills. The first tool, Writing Coach, launches today.
The Breakdown:
Writing Coach guides students through outlining, drafting, and refining essays rather than generating finished text for them
Available for grades 7-12 (beta for grades 5-6) in the U.S., covering persuasive, expository and literary analysis essays
Teachers can toggle between full interactive mode or feedback only mode
Reading Coach launches later this year for grades 5-12, enabling teachers to assign interactive reading experiences with comprehension tracking
Schoolhouse.world, Khan's peer tutoring platform, now uses Gemini to coach human tutors and simulate student interactions for practice sessions
The bigger picture: This positions Gemini as the backbone for one of the largest free education platforms globally. The design philosophy guiding process over generating answers signals how Google wants AI deployed in classrooms: augmenting teachers rather than replacing instructors.
What else you need to know:
Anthropic released its performance engineering take home test as an open challenge after redesigning it three times because successive Claude models kept matching or exceeding top human candidate scores within time limits.
Adobe Acrobat introduced AI tools that generate presentations and podcast style audio summaries from uploaded documents plus natural language commands for common PDF editing tasks.
Baidu launched ERNIE 5.0, a 2.4 trillion parameter MoE model with native omni modal capabilities that uses under 3% active parameters per inference and is now available via ERNIE Bot and Qianfan platform.
Comet upgraded its browser agent to use Anthropic's Opus 4.5 by default, citing improved reasoning and better handling of complex tasks for Perplexity Max subscribers.
Cognition released Devin Review, a free code review tool that intelligently organizes diffs, provides codebase aware chat and flags potential bugs to help developers understand complex PRs from humans or AI agents.
That’s it for today’s edition of The AI Night.
Our goal is to cut through the noise, surface what actually changed, and explain why it matters.
If this was useful, you’ll get the same signal here tomorrow.