The Twenty Minute VCRoundtable #7: Spotify, Adobe and Linkedin on How AI Changes The Future of Product & Design | E1097
CHAPTERS
- 0:00 – 0:42
What product leaders want you to know about AI (setup + key themes)
Harry sets up a roundtable with product leaders from Adobe, Spotify, and LinkedIn, framing the discussion around how AI changes product and design. The opening clips preview core tensions: mega-models vs many models, losing deterministic control, and designers needing deep model literacy.
- •Roundtable focus: AI’s impact on product development and design
- •Previewed debate: centralized “mega models” vs a long tail of specialized models
- •Shift from deterministic UX to probabilistic AI-driven experiences
- •Designers must understand model capabilities like they understand users
- 0:42 – 2:09
Panelist backgrounds: Adobe, Spotify, LinkedIn leadership perspectives
Each leader introduces their role and career context, grounding their viewpoints in the types of products they build—creative tools, media personalization, and professional network utility. Their backgrounds foreshadow different AI priorities: IP safety, personalization at scale, and workflow/task completion.
- •Scott Belsky: Adobe strategy/emerging products; founder of Behance
- •Gustav Söderström: Spotify long-tenured leader; now co-president
- •Tomer Cohen: LinkedIn CPO; engineering roots through semiconductors to AI
- •Different company contexts shape AI product constraints and opportunities
- 2:09 – 4:52
How AI changes product development: faster iteration, AI as the product, letting go of control
The panel moves from coding productivity to broader product-building changes: AI accelerates exploration and reduces cycles to find better solutions. Gustav reframes the shift as ‘AI is the product’ and UI becomes a mechanism to capture better signals; Tomer emphasizes AI-first as a leadership and mindset shift that requires surrendering deterministic control.
- •AI improves engineering throughput (coding, testing, bug-fixing)
- •AI enables rapid UI/feature variation and experimentation
- •Reframe: AI becomes the product; UI supports data/signal collection
- •AI-first must not be delegated to a single ‘AI team’
- •Probabilistic outputs force product leaders to relinquish full control
- 4:52 – 9:08
The future of UI: from screens to persona, tone, and conversation design
Harry challenges the idea that AI makes UI redundant, and Scott argues UI evolves rather than disappears—toward tone, inflection, and persona design. Gustav shares Spotify DJ as an example where brand and personality decisions are core UI choices in an AI-native experience.
- •UI doesn’t vanish; it morphs into persona/tone and conversational UX
- •Spotify DJ illustrates “personality as interface” decisions
- •Design becomes more holistic: UX + brand + interaction style
- •AI experiences require new design primitives beyond buttons and menus
- 9:08 – 11:02
Designing for imperfect models: tolerances, staged output, and AI literacy
Gustav argues designers must understand models as deeply as users to design around model latency and error rates. Using Midjourney, he explains how experience design can compensate for slow generation and low hit rates through batching, previews, and progressive refinement—turning model constraints into usable workflows.
- •Designers need model literacy (capabilities, failure modes, latency)
- •Midjourney example: four low-res options fast vs one slow, fragile result
- •UI should be ‘fault-tolerant’ to probabilistic outcomes
- •Recommendation: match UI structure to model hit-rate (e.g., show more options)
- 11:02 – 14:05
Many models, not one: routing layers, dispatchers, and agent teams
The conversation shifts to how products will orchestrate multiple models without forcing every product person to master each one. Tomer and Scott describe platform layers—routers/dispatchers—that select models based on task, quality, and cost, plus the application-level pattern of multiple specialized agents coordinated by a dispatcher.
- •Product teams shouldn’t manually juggle models; platforms should abstract it
- •Two routing layers: application-level agent dispatcher and infra-level model router
- •‘One agent to rule them all’ is the wrong approach; build a team of agents
- •Opportunity area for startups: routing/dispatch and abstraction layers
- 14:05 – 16:40
Cost becomes product viability: why routing and in-house build decisions matter
Harry probes unit economics, and Gustav highlights cost as a gating factor for shipping AI at Spotify scale (e.g., generating billions of minutes of audio). The panel links routing and optimization to margins, noting that performance gains and cost-efficiency improvements often arrive together.
- •AI compute cost can make or break consumer-scale features
- •Spotify example: voice generation at massive scale demands extreme efficiency
- •Routing isn’t only about ‘best model’—it’s about cheapest adequate model
- •Expectation: costs decline as models and systems improve
- 16:40 – 18:04
Moore’s Law, neural hardware, and the push-pull of bigger vs smaller models
The panel debates whether AI cost declines will follow Moore’s Law dynamics. Gustav predicts continued progress via specialized neural hardware and notes the simultaneous trends of scaling up frontier models while rapidly distilling capabilities into smaller models; Tomer adds that verticalized chip/software stacks will accelerate efficiency.
- •Moore’s Law may shift forms, but neural hardware is early-stage
- •Trend 1: bigger models keep improving; Trend 2: smaller models catch up fast
- •Vertical integration (chips + software) increases efficiency and lowers cost
- •Closed, optimized stacks from hyperscalers and device ecosystems may dominate some layers
- 18:04 – 19:36
Data vs model size: specialization, training quality, and proprietary advantage
Harry asks what matters more: bigger models or better data. Tomer argues it depends on the task: large assistants may need large models, but specialized, well-trained models can outperform larger under-trained ones; Gustav adds that long-term advantage may come from high-fidelity user data and understanding, not just model scale.
- •Model size (parameters) matters, but only relative to the goal
- •Under-trained large models can lose to well-trained smaller models
- •Specialized ‘role’ agents (e.g., job seeker coach) can outperform generalists
- •Proprietary user data and high-fidelity signals may be the enduring moat
- 19:36 – 26:43
Hardest parts of implementation: the ‘final mile,’ org retooling, and prompt-driven prototyping
Scott calls out the “final mile” tuning—experience finesse only product teams with deep customer knowledge can do—and warns against over-fixating on tech over UX. Gustav and Tomer describe retooling organizations: aligning mindsets, setting AI-first bets, and even requiring prompts as artifacts in product reviews, enabling designers/PMs to prototype without engineering bottlenecks.
- •Implementation difficulty shifts ‘up the stack’ over time (talent → data → product/design)
- •The ‘final mile’ is experience tuning driven by customer empathy
- •Org change: educate teams to calibrate expectations (avoid under/over-estimating AI)
- •New product rituals: bring prompts to jam sessions; critique prompt quality
- •AI democratizes prototyping for designers and PMs
- 26:43 – 29:32
Build vs partner: where each company should own models (and why)
Gustav explains Spotify builds models aligned to its goals (cost-efficient media experiences) rather than competing for AGI, and partners for general capabilities. Scott frames the decision rule: build models only if you can be best in the world for that domain (e.g., Adobe for imaging), and partner for general LLMs; Tomer emphasizes differentiation through integration layers and domain understanding rather than chasing frontier LLM parity.
- •Spotify: build where goals differ from AGI labs; partner for commoditized capabilities
- •Adobe: build imaging models as a core competitive advantage; partner for LLMs
- •LinkedIn: invest in integration/routing to apply models to its graph and workflows
- •Key heuristic: own what you can uniquely do best; buy/partner for the rest
- 29:32 – 35:21
Data strategy and governance: quality signals, IP safety, and AI enabling new business models
Tomer argues product leaders must treat data as ‘oxygen’—owning algorithm objectives and data collection rather than outsourcing it. Scott details Adobe’s approach to commercial safety: respecting IP, not training on customer creations without consent/compensation, and using licensing as a trust differentiator; Gustav adds that AI’s biggest disruptions come when it enables new business models, which is only beginning.
- •Product leaders must define algorithm objectives and ensure high-quality data collection
- •Adobe stance: do not train on customer creations by default; compensate when opted-in
- •IP/licensing and trust become strategic differentiators in generative AI
- •AI’s major impact may arrive through business-model innovation, not just better tech
- 35:21 – 44:18
Who wins and how fast incumbents can move: platform shifts and adoption curves
Harry challenges the ‘incumbents can’t move fast’ narrative. Scott argues AI differs from cloud/mobile shifts because incumbents can inject AI into existing workflows and customer touchpoints; Gustav insists disruptive change usually requires a business-model break, and expects AI adoption to follow a sharper S-curve because teams can trial AI in parallel rather than fully migrating stacks.
- •AI platform shift may favor incumbents with distribution and customer insight
- •Injecting AI into existing products can be faster than rebuilding from scratch
- •True disruption often requires a new business model that breaks incumbents
- •Enterprise adoption likely accelerates via step-function references and parallel trials
- 44:18 – 49:16
Career advice for designers and PMs: AI literacy, T-shaped skills, and being the change agent
The panel closes with concrete guidance for early-career builders: jobs and skill requirements are changing fast, so cultivate a growth mindset and AI literacy. Tomer recommends broader (T-shaped) capability plus domain expertise and human strengths; Scott advises adopting tools personally, piloting new practices at work, and becoming the teammate who introduces modern workflows; Gustav stresses education is more accessible and democratic than ever.
- •LinkedIn data: skill requirements shifting rapidly through 2030
- •Develop AI literacy + growth mindset; expect titles/tasks to morph
- •Become T-shaped: broad skills + deep specialty in a domain/industry
- •Use tools personally first; run pilots; be the team’s early adopter and educator
- •Human strengths (judgment, interpersonal skills, imagination) remain critical
- 49:16 – 55:54
Quick-fire round: running lessons, LinkedIn complexity, org centralization, and AI expectations for leaders
A rapid Q&A surfaces practical leadership heuristics: Scott’s ‘wait for it’ lesson from running, Tomer’s focus on reducing product complexity with AI, and Scott’s evolving stance on centralization based on strategy. Tomer is blunt that product leaders must be willing to go deep on AI principles, and Gustav shares excitement about returning to a high-change era reminiscent of the mobile shift.
- •Running as a metaphor: letting ideas mature before acting improves outcomes
- •LinkedIn: AI can simplify complex multi-use-case products without adding features
- •Leadership playbooks aren’t permanent; centralize/decentralize as strategy changes
- •CPOs/PM leaders must understand AI principles (probabilistic UX, data, velocity)
- •AI era feels like the beginning of a major new platform shift
