CHAPTERS
- 0:00 – 2:35
Why AI product design is harder than it looks (and what this episode covers)
Aakash introduces Elizabeth Laraki (early Google Search/Maps designer) and frames the episode as a practical masterclass on designing AI-powered products. The throughline: AI is non-deterministic, so product design must adapt—beyond just “chat UIs.”
- •Elizabeth’s credentials: original designer on Google Search and Maps
- •Promise of a full-spectrum AI product design breakdown (features + tools)
- •Framing AI products as non-deterministic systems that require new UX thinking
- •Teaser stories: AI search, multimodal design, and a cautionary image-expander incident
- 2:35 – 4:08
How Google Search evolved: blending new result types and AI summaries
Elizabeth explains why early Google Search designs endured: the team focused on integrating new information types (images, video, maps) without redesigning everything. The conversation sets up a key principle—evolution through details and information architecture, not wholesale UI churn.
- •Early focus: diversify beyond “10 blue links” with images, video, maps
- •Search stays recognizable because designers iterated via nuanced improvements
- •Modern Search: rich modules + emerging AI summaries in results
- •Design lesson: keep core mental model stable while expanding content types
- 4:08 – 6:55
Google’s AI Search integration: competing with ChatGPT while managing hallucinations
They debate the value of AI answers inside Search. Elizabeth argues Google must meet users where they already are, while both acknowledge hallucinations and confidence thresholds as central product challenges.
- •Strategic rationale: prevent user migration to ChatGPT by embedding AI in Search
- •Hallucinations are ecosystem-wide; users need discretion like early internet days
- •Design implication: decide when to show AI answers based on confidence
- •Product teams must treat AI as non-deterministic and design guardrails accordingly
- 6:55 – 9:56
Designing multimodal AI help: why linear chat fails for image/video tasks
Using a ‘bike seat adjustment’ example, Elizabeth critiques chat’s linearity for hands-on problems. She proposes interfaces where the visual context stays central while conversation (audio/text) wraps around it—closer to a remote expert experience.
- •Chat is linear; real-world help is dynamic and iterative
- •Better model: video/photo remains central while guidance happens around it
- •UI opportunities: highlight bolts/parts directly rather than describing in text
- •Design can drive research: specify ideal interactions to guide model capabilities
- 9:56 – 13:12
The AI image expander disaster: unintended consequences in real workflows
Elizabeth shares a real incident where an image expander added sexualized details to a conference promo image. The story illustrates how innocuous tool chains can produce harmful outputs—and why AI products need extra scrutiny and transparency.
- •Workflow chain: crop to square → expand to portrait → AI “fills in” problematic details
- •User trust risk: output felt intentional/sexualizing despite being accidental
- •Tool variance: some expanders produced even worse results
- •Core lesson: AI can introduce surprising social/cultural bias and reputational harm
- 13:12 – 18:59
Safeguards + human-in-the-loop: training data, evals, and UI transparency
They shift from the incident to prevention: improving models (data + evals) and designing UI to clearly distinguish AI-generated portions from original content. The emphasis is on practical guardrails rather than perfect automation.
- •Model layer: training data quality and bias strongly shape outcomes
- •Evaluation layer: add checks around sensitive areas and edge cases
- •UI layer: clearly show what pixels/content were AI-generated vs original
- •Human review is essential; design should make review fast and obvious
- 18:59 – 19:20
A simple 3-step process for designing AI features
Elizabeth offers a high-level, repeatable process: define the product, design it, then build it. The simplicity is intentional—clarity of intent and user need precedes UI and model work.
- •Step 1: define what you’re building (who it’s for + what it does)
- •Step 2: design the experience and interaction model
- •Step 3: build and iterate with real usage signals
- •AI doesn’t remove fundamentals; it increases the need for crisp definition
- 19:20 – 25:21
AI products that are designed well: why ChatGPT became the default tool
Elizabeth explains why ChatGPT stands out: it’s broadly useful, uncluttered, and supports multiple interaction modes while keeping the core action simple. She also notes people use other models as “reference checkers,” similar to Uber vs Lyft behavior.
- •ChatGPT as an all-purpose default for Q&A, planning, translation, cooking, etc.
- •Secondary tools (Claude/Gemini) used for verification rather than first choice
- •Design win: minimal clutter; simple core workflow with advanced “level-ups”
- •Potential downside: the blank-page problem for new users (onboarding tension)
- 25:21 – 28:26
Descript/Riverside: ‘baking AI into the cake’ for end-to-end workflows
They discuss why AI features feel powerful when embedded into the actual job-to-be-done of video editing. The tools reduce intimidation by turning a complex workflow into approachable, modular steps.
- •Transcript-based editing changes the mental model (edit text → edits video)
- •Remove filler words and smooth edits; not perfect but high leverage
- •Eye contact correction and clip/title generation make production accessible
- •Design principle: integrate AI across the workflow, don’t bolt it on as a gimmick
- 28:26 – 33:34
Midjourney and AI image generation UX: onboarding, cost friction, and output quality
They unpack why Midjourney succeeded despite earlier Discord friction, and why moving toward a web interface matters for adoption. Elizabeth emphasizes consistent output quality and the economic reality that limits casual tool experimentation.
- •Discord-based UX was a barrier; web access is a key unlock
- •AI tools are costly to run, pushing subscriptions and limiting experimentation
- •Success driver: fast idea-to-4-variations loop with consistently decent quality
- •Room to improve: broader critique of image-gen tool UX (teased for later)
- 33:34 – 38:37
Designing AI voice interfaces: context-first, not screen-first
Elizabeth explores voice UX through examples: ChatGPT voice in the car, Limitless as an always-on coach, and Meta Ray-Bans translating menus poorly. The key is matching interaction style to context and adopting a more human dialog strategy than “screen reader mode.”
- •Voice context: often no one is looking at a screen (e.g., driving)
- •ChatGPT voice can feel like ‘another person in the car’—natural conversation
- •Limitless value: feedback/coaching from ambient conversation data
- •Meta glasses example shows failure mode: reading menus linearly vs dialog-driven summarization
- 38:37 – 42:31
Beyond chat: canvas-based AI and co-creation instead of linear transcripts
They argue chat should be a tool, not the whole interface—especially for tasks needing structure and stability (like travel itineraries). Elizabeth highlights Cove as a canvas approach that supports parallel threads and editable artifacts.
- •Chat constraints: linear output and poor referencing of earlier context
- •Co-creation needs a stable artifact (e.g., a ‘document’) not endless regeneration
- •Non-determinism makes deterministic tasks harder without a structured UI
- •Cove-style canvas: multiple workstreams/documents instead of one chat stream
- 42:31 – 45:11
AI design tools for designers: productivity at the edges, taste at the center
Elizabeth is candid: she doesn’t use many AI design tools because the output rarely meets professional taste standards. The recommendation is to use AI for speed and scaffolding, while relying on human judgment for high-quality craft.
- •AI tools can help amateurs reach ‘good enough’ faster
- •For expert designers, current tools often fall short of quality expectations
- •Where AI helps: specs, first-pass layouts, prototypes, peripheral tasks
- •Core principle: design taste and judgment remain the bottleneck
- 45:11 – 57:38
Live design exercise: decomposing ‘LinkedIn for AI’ into product strategy and system design
In a live whiteboarding-style segment, Elizabeth shows how to turn a vague prompt into a concrete product direction. She maps possible interpretations (matchmaking, certification, networking, content) and then drills into matchmaking as a two-sided marketplace with AI-driven attributes.
- •Start with definition: clarify objective before jumping to pixels
- •Explore product directions: matchmaking vs certification/training vs networking/content
- •Matchmaking deep dive: job seeker UI + employer UI with AI ‘magic in the middle’
- •Borrow patterns from other matching domains (dating, college admissions, personality testing)
- 57:38 – 1:12:47
Google Maps redesign and the India landmark-navigation story: timeless user research principles
Elizabeth recounts simplifying Google Maps from multiple tabs/search boxes to a single search box—controversial at the time but foundational now. She then explains how research in India led to landmark-based directions, showing that strong product design still starts with understanding how people navigate in the real world.
- •Maps redesign: feature growth created clutter; solution was simplifying around core use cases
- •Design architecture: organize actions by global vs task-specific needs
- •Critique of newer Maps UI: colder aesthetics and renewed clutter without true cleanup
- •India launch: field research revealed landmark-based navigation; system integrated ‘pass by X’ and verification landmarks
