What We Expect from NotebookLM: Day 4 — Advanced AI Functionalities
AI that thinks with you — From passive tool to active thinking partner
Today’s AI assistants are transactional — you ask, they answer. This model misses the deeper potential of human-AI collaboration. We need a thinking partner that learns our habits, anticipates needs, and grows more useful over time.
NotebookLM has the foundation for that role — it understands sources and creates insights — but it remains passive, waiting for explicit commands. It doesn’t adapt, automate patterns, or proactively highlight what matters.
The shift must be toward an active collaborator, through:
Personalized command systems that learn your language.
Context-aware citations tied to sources.
Multimodal intelligence that goes beyond text.
Feature 1: Custom command system
What we expect: Personalization of AI interaction through user-defined shortcuts and prompt templates that evolve with usage patterns.
The repetition problem in current interfaces
You find yourself typing the same instructions repeatedly. “Create an executive summary of this document, maximum 300 words, in bullet point format.” Three days later, working with a different document, you type nearly identical instructions again. And again. And again. The AI has no memory of your preferences, no understanding that you consistently want the same output format.
This friction compounds over time. Researchers develop favorite analysis patterns. Students prefer specific study formats. Professionals have standardized report structures. Yet every interaction starts from zero, requiring full specification of requirements.
Personalized shortcuts eliminate repetition
We expect a command definition system where users create their own shortcuts. Imagine uploading a simple markdown file that defines your personal NotebookLM language. The command “/sum” instantly triggers “create executive summary, maximum 300 words, bullet format.” The command “/eli5” means “explain like I’m 5 using simple language and analogies.” The command “/cite” generates “list all sources with page numbers in APA format.”
Analysis commands could handle complex requests. The shortcut “/compare [topic]” would trigger comparative analysis across all sources for any specified topic. The command “/timeline” would extract chronological event sequences. The shortcut “/gaps” would identify knowledge gaps in current sources. The command “/bias” would detect potential biases across your source collection.
Output format commands would provide instant transformation. The shortcut “/eo2” would rewrite any response in bullet points. The command “/table” would convert information to comparison table format. The shortcut “/flow” would create decision tree flowcharts. The command “/quiz” would generate ten quiz questions with detailed explanations.
Advanced capabilities through command composition
Simple shortcuts are just the beginning. We expect variable support where commands accept parameters. The shortcut “/sum 500 narrative” would create a 500-word narrative summary, while “/sum 200 technical” would produce a 200-word technical overview. Chained commands could combine operations, like “/cite + /academic” producing citations in academic format with proper bibliographic structure.
Conditional logic would add intelligence. A command could specify “if source_type equals research_paper then use academic tone, else use plain language explanation.” Custom parameters would enable complex analysis requests with structured input defining focus areas, depth levels, and comparison sources.
Discovery and learning mechanisms
The system needs discoverability. As you type “/” in the chat interface, a suggestion dropdown appears showing all available commands with brief descriptions. A command palette accessible via keyboard shortcut provides fuzzy search across all defined commands. Usage statistics reveal which commands you employ most frequently, helping optimize your personal command library.
Community command marketplaces would enable sharing. Researchers could publish their analytical command sets. Students could share study-optimized commands. Legal professionals could distribute citation and brief-writing commands. The best community contributions would rise through rating and adoption metrics.
Consider the legal researcher workflow. She analyzes case law daily, always needing citations in Bluebook format, always wanting chronological timelines of relevant decisions, always requiring bias checks for jurisdictional coverage. Instead of typing elaborate instructions for each case, she defines three commands: “/blue” for Bluebook citations, “/casetimeline” for chronological analysis, and “/jurischeck” for jurisdictional bias detection. What previously required typing 200+ characters now takes 6. Her cognitive load drops, her speed increases, and her focus stays on legal analysis rather than prompt engineering.
Feature 2: Context-aware citation enhancement
What we expect: Complete transparency in AI reasoning through intelligent citation systems that show exactly where information comes from.
The mysterious numbers problem
Current NotebookLM responses include citations like [1], [2], [3]. This is better than no citations. But it creates a frustrating puzzle. You read “Research shows productivity increases by 35% [1] when using spaced repetition [2], particularly in technical domains [1][3].” You see the numbers. You don’t immediately know which PDF is [1]. You have to scroll to a separate citation list, match the number, find the source, then return to reading. This friction breaks flow and reduces trust.
The problem intensifies with multiple sources. In a notebook with 50 PDFs, remembering that [1] refers to “Ebbinghaus_Memory_1885.pdf” while reading paragraph five of a long response requires working memory resources better spent on comprehension. The citation system should reduce cognitive load, not add to it.
Enhanced citation format provides instant context
We expect inline citations that immediately identify sources. Instead of [1], we see “[Ebbinghaus_Memory_1885 — p.47].” The filename appears directly in the citation. The page number shows exactly where to verify. No memory required. No scrolling to match numbers. Complete transparency in a single glance.
Get Mihailo Zoin’s stories in your inbox
Join Medium for free to get updates from this writer.
The full citation would read: “Research shows productivity increases by 35% [Ebbinghaus_Memory_1885 — p.47] when using spaced repetition [Bjork_Learning_2011 — p.112], particularly in technical domains [Ebbinghaus_Memory_1885 — p.89][Dunlosky_Practice_2013 — p.5].”
Interactive citation elements
Hovering over any citation should trigger a tooltip preview showing source title, full bibliographic information, the date it was added to the notebook, and quick action buttons to open the source or view surrounding context. This creates zero-friction verification without leaving the reading experience.
A citation sidebar could display all sources referenced in the current response. For each source, it would show how many times it was cited, which specific pages were referenced, and offer to display the relevant quoted passages. This overview helps readers assess the evidential basis of AI-generated content.
Export with proper bibliography
When exporting any AI-generated response, the system should automatically generate a bibliography section. Format options would include APA, MLA, Chicago, or Harvard styles. Citations within the exported document would be hyperlinked, allowing readers to jump directly to the bibliography entry. A separate reference list would provide full citation details in professionally formatted text.
The academic integrity scenario shows why this matters. A graduate student uses NotebookLM to analyze 40 papers for a literature review. The AI generates insightful synthesis across sources. With current citation systems, the student must manually track down which claims came from which papers, verify each assertion, and create proper citations for the final document. This process takes hours and introduces error risk. With enhanced citations, every claim links directly to its source with page numbers. Export generates a properly formatted bibliography. The student focuses on analysis and writing rather than citation archaeology. Academic integrity improves while time investment drops dramatically.
Feature 3: Image and multimedia support
What we expect: Expansion beyond text-only to full multimodal intelligence that understands images, audio, and video as naturally as written documents.
The text-only limitation
Knowledge doesn’t exist only in words. Researchers photograph whiteboards after brainstorming sessions. Students take pictures of handwritten lecture notes. Professionals screenshot diagrams and charts. Podcast listeners want to analyze spoken content. Video tutorials contain valuable information in visual and auditory channels.
Current NotebookLM handles text beautifully but ignores everything else. You can’t drag a photo of handwritten notes into the chat. You can’t upload a lecture recording for transcription and analysis. You can’t reference a diagram from a YouTube video. This artificial limitation forces users to manually transcribe visual and auditory information into text, wasting time and losing fidelity.
Image processing capabilities
We expect drag-and-drop image support directly in the chat interface. Drop a photo of handwritten notes, and optical character recognition extracts the text for searching and analysis. Upload a whiteboard photo from a brainstorming session, and the system converts scattered ideas into structured notes with relationship mapping. Screenshot an infographic, and data extraction pulls numbers and facts into analyzable format. Share a diagram, and the vision model interprets the visual logic, explaining what the flowchart or schematic represents.
Visual question answering would enable natural interaction. You could ask “What does this flowchart show?” after uploading a process diagram. The system would analyze the image, identify components and connections, and explain the represented workflow in clear language. This transforms static images into interactive knowledge sources.
Audio integration transforms spoken content
Audio file uploads would trigger automatic transcription with multilingual support handling diverse languages naturally. Speaker diarization would identify who spoke when during multi-person recordings, creating organized transcripts that preserve conversational structure. Timestamp navigation would allow jumping to specific moments, so referencing “the explanation at 23:45” links directly to that position in the audio. Key point extraction would summarize hour-long recordings into essential insights with timecode references.
The interview researcher scenario demonstrates the power. She conducts 30 expert interviews for a qualitative study. Each interview runs 60–90 minutes. Traditionally, she would pay for professional transcription services or spend hundreds of hours transcribing manually. Then she would read all transcripts, manually coding themes and tracking which expert said what. With audio integration, she uploads all 30 recordings. The system transcripts everything, identifies each speaker, extracts key themes, and allows her to ask questions like “What did experts say about regulatory challenges?” The AI synthesizes across all 30 hours of audio, citing specific speakers and timestamps. Her analysis time drops from weeks to days.
Video processing combines visual and auditory understanding
YouTube video support and uploaded video file processing would analyze content through dual channels. Visual analysis would detect slide changes in presentations, extract on-screen text from educational videos, and identify important visual moments. Audio transcription would convert speech to searchable text. Combined understanding would integrate what’s shown with what’s said, creating richer comprehension than either channel alone. Keyframe extraction would automatically screenshot important moments, creating a visual summary of hour-long videos.
Multimodal chat enables sophisticated queries
The real power emerges in combination. You could ask “Compare this diagram [uploaded image] with the concept from [specific PDF source].” The AI would interpret the diagram through its vision model, search for relevant concepts in the specified PDF, generate comparative analysis highlighting similarities and differences, and create a unified visualization combining insights from both sources.
Consider the engineering student workflow. She’s learning circuit design. Her textbook contains circuit diagrams. Her professor sketches variations on the whiteboard during lectures. YouTube tutorials show animated explanations. Currently, these exist in separate mental spaces. With multimodal support, she uploads the textbook PDF, photographs the whiteboard sketches, and links to the YouTube tutorials. When she asks “How do these three approaches to amplifier design differ?” the AI analyzes the textbook diagram, interprets her whiteboard photos, processes the YouTube video content, and generates a comprehensive comparison across all three sources. Learning accelerates because the AI bridges modalities she previously had to connect manually.
Tomorrow: What we expect — Interfaces that adapt to you
In Day 5, we explore what we expect from user experience revolution. NotebookLM’s interface should learn how we work and adapt accordingly, not force us to adapt to rigid structures designed for average users who don’t exist.
Tomorrow’s preview covers adaptive interfaces that reconfigure based on usage patterns, cognitive load reduction through intelligent hiding of complexity, role-based presets optimized for different professions, and collaborative features that transform individual work into team intelligence.
About this series:
Day 4 of our six-day exploration of what we expect from NotebookLM’s future. Today’s focus: advanced AI capabilities that transform passive tools into active thinking partners.
- Day 1 (October 8): Intelligent source organization
- Day 2 (October 9): Automation of repetitive processes
- Day 3 (October 10): Export and interoperability
- Day 4 (October 11): Advanced AI functionalities
- Day 5 (October 12): User experience revolution
- Day 6 (October 13): Intelligent analysis and insight generation
