Dec 20, 2023 · 55m · news
Roundtable #7: Spotify, Adobe and Linkedin on How AI Changes The Future of Product & Design | E1097 · 20VC with Harry Stebbings
Scott Belsky · 17m spoken Gustav Söderström · 13m spoken Tomer Cohen · 13m spoken Harry Stebbings · 5m spoken
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this 20VC roundtable, host Harry Stebbings leads a deep-dive discussion with product leaders from Spotify, Adobe, and LinkedIn on how artificial intelligence is fundamentally reshaping user interfaces, product development cycles, enterprise business models, and talent requirements in the technology sector.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 11.7% of the talking time here. How this is scored →
Harry as informed peer 2.9 Guest teaching 3.6 Guest disagreement 2.2 Harry pushing back 2.6
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 18:22 Tomer explicitly rejects broad question framing
Tomer directly rejects Harry's generalized question on model parameters versus data volume, stating 'this is why I'm not a big fan of broad questions' before forcing a more nuanced breakdown.
Hardest push from Harry ▶ 7:50 Harry challenges panel on non-deterministic risksHarry forcefully challenges the panel's comfortable stance on probabilistic outputs by highlighting that all three guests represent public companies accountable for unmanaged hallucinations.
Biggest teaching moment ▶ 11:28 Tomer corrects assumption about PM model oversightTomer dismantles Harry's premise that product leaders need to master multiple individual models, clarifying that model selection is masked by platform-level dispatcher architectures.
Harry holds his own ▶ 39:11 Harry cites Benchmark thesis on selling work vs seatsHarry demonstrates strong domain mastery by citing Sarah Tavel's Benchmark essay on shifting SaaS metrics from seat pricing to selling completed work during a discussion on business model disruption.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Introductions of the Panelists | 2 | 2 | 1 | 1 | Harry welcomes guests Scott Belsky, Gustav Söderström, and Tomer Cohen, asking them to introduce themselves and opening with a broad question on AI in product development. Gustav offers a conceptual pivot noting that UI used to be the product, whereas now AI is the product and UI exists to collect signal. | |
| The Evolution of User Interfaces (UI) and Personas | 2 | 3 | 2 | 2 | Scott describes UI evolving into product persona design, while Tomer highlights the shift to non-deterministic experiences where product leaders lose direct control over outcomes. Tomer uses a chef analogy to explain how product teams must learn to dictate ingredients rather than dictate outputs. | |
| Hallucinations: Bug vs. Feature, and AI Capabilities in Design | 3 | 4 | 3 | 5 | Harry challenges the panel on public company liability regarding uncontrolled AI outputs and hallucinations. Scott reframes hallucinations as features rather than bugs in generative creative contexts like Photoshop, while Gustav elaborates on fault-tolerant UI design using Midjourney as a primary example. | |
| Managing Multiple Models, Routing, and Cost Efficiency | 3 | 5 | 4 | 3 | Harry asks how product leaders can manage eight different models simultaneously. Tomer explicitly rejects the question's premise, clarifying that routing logic sits at the platform tier rather than with individual product managers, while Scott agrees and highlights the emergence of model router startups. | |
| Cost Implications and Personalization of User Data | 2 | 4 | 1 | 2 | Harry inquires about model deployment costs and unit economics. Gustav illustrates Spotify's scale challenge with audio generation costs across half a billion users, outlining a vision where entire user histories are tokenized into unified prediction spaces. | |
| Future Progressions: Moore's Law, Hardware, and Prompts | 2 | 5 | 5 | 2 | Harry prompts the panel on Moore's Law and model size versus data size. Tomer explicitly rejects Harry's broad framing, explaining that model parameters and dataset volume require contextual task trade-offs rather than general answers. | |
| The Hardest Parts of Model Implementation and Scale | 2 | 3 | 1 | 2 | Harry asks about the hardest technical hurdles in model implementation. Scott argues that last-mile tuning and user empathy remain the hardest problems, while Gustav describes the internal challenge of retooling an entire company's mindset around probabilistic models. | |
| Democratic Innovation and Designing for Firefly | 3 | 4 | 2 | 4 | Tomer and Scott discuss prompt reviews as part of product jam sessions and design-led innovation for Adobe Firefly. Harry directly questions Gustav on why Spotify doesn't build its own foundational models, to which Gustav outlines Spotify's distinct product goals compared to AGI labs. | |
| Build vs. Buy and the Role of Proprietary Data | 3 | 5 | 3 | 2 | Scott details Adobe's build-versus-buy logic around commercially safe imaging models versus general LLMs. Tomer delivers a passionate explanation of data hygiene and algorithm objectives, criticizing product leaders who outsource data collection to data science teams. | |
| Incumbents vs. Startups and Business Model Disruption | 5 | 4 | 2 | 3 | Harry questions whether incumbents moving slowly is a myth. Scott notes AI shift nuances favor incumbents with existing customer reach, while Tomer notes seat-based SaaS cannibalization. Harry demonstrates industry knowledge by bringing up Sarah Tavel's thesis on selling work instead of seat licenses. | |
| Enterprise Adoption and the Evolution of Pricing Models | 5 | 3 | 2 | 3 | Harry highlights lagging enterprise adoption, citing a specific statistic that 32% of European corporates still don't use Slack. Scott and Gustav explain enterprise adoption curves, emphasizing design partnerships over traditional sales cycles to drive internal usage. | |
| Preparing for the Future: Career Advice for Young Designers and Product Managers | 3 | 4 | 2 | 4 | Tomer cites LinkedIn talent data on rapid skill shifts by 2030, encouraging soft skills and T-shaped growth. Harry pushes back on generalities, asking the guests to go granular on actionable steps, prompting Scott to advise young PMs to treat themselves as testing grounds for new tooling. | |
| Quick Fire Round: Scott Belsky on Lessons from Running | 3 | 2 | 1 | 2 | In a quick-fire round, Scott reflects on distance running lessons applied to executive decision-making, emphasizing the value of sitting with ideas ('wait for it') before rushing to execute. Tomer discusses applying the law of conservation of complexity to simplify LinkedIn. | |
| Quick Fire Round: Scott Belsky on Changing His Mind on Centralization | 3 | 3 | 2 | 2 | Scott discusses changing his mind on design team centralization depending on strategy context. Tomer asserts he wouldn't hire a CPO who refuses to develop deep technical AI literacy, and Gustav expresses optimism for non-iterative business model disruption ahead. |
Statements from this episode (37)
Belsky: AI future will not be dominated by a few mega models
“I think it's actually common belief that there are going to be a few mega models in the future that are going to do everything for every company in the cloud. What we're saying is actually it will probably be the opposite.”
Cohen: AI-first product design requires relinquishing control over user experience
“When you think in an AI first principle kind of way, You're really unleashing the idea of control. What happens basically in AI is you don't control the experience anymore.”
Söderström: Product designers must understand GPT-4 as deeply as user needs
“Something that designers need to get good at in this world. They need to understand GPT-IV as well as they understand the user.”
Belsky: AI-suggested variations will yield better UI design with fewer cycles
“We're in a world where AI will suggest alternative scenarios. You can try variations and just look and, you know, with the mistake of the eye, find a better solution. That, or at least something you want to A-B test. And so I think that the product development…”
Söderström: AI is the product; UI merely captures signal for it
“Now AI is the product and the UI is there to help the AI to capture better signal.”
Söderström: TikTok succeeded through UI design, not novel recommendation algorithms
“It wasn't really the algorithm of TikTok. It was the UI that maximized what was, you know, pretty traditional export export algorithm, like since the day so hot or not.”
Belsky: Future UI design is evolving into persona and tone design
“UI in some cases should disappear, but there's new UI these days in the form of the tone of a product and the way that something is is the inflections that are used, the persona. You know, I'm like thinking about UI design in the future as in some ways like pe…”
Cohen: AI strategy must originate from the CEO, not be delegated
“But for many product builders, AI is something they delegate. I mean, it's like really delegated to the AI team or the engineering team. I think AI strategy starts from the CEO and makes its way down.”
Belsky: AI hallucination is a feature, not a bug, in creative tools
“In, in some ways, listen, for some practices hallucination, hallucination is a bug, but in some areas it's also a feature, right? So if I'm trying to discover some cool new music, I mean, I can imagine that might be a feature as opposed to a bug. You know, whe…”
Cohen: Massive startup innovation will emerge in AI model routing tiers
“If I think of like why I would expect to see some massive innovation and startups to show up, it's in this tier to really allow people to leverage multiple models at multiple call centers and resources and efficiencies, and then completely mask it from the dev…”
Cohen: Monolithic AI agents are wrong; systems require multiple specialized agents
“Building one agent to rule them all is not the right design. You want to have multiple types of agents.”
Söderström: Spotify builds most technology in-house primarily to control operating costs
“From Spotify's point of view most of the stuff we've built in-house, to Tamara's point, is actually mostly about cost.”
Söderström: Companies will embed entire user histories into single AI models
“I think the world is going to go towards you embed your entire user history, everything they did into one space.”
Söderström: Neural hardware progress will follow a Moore's law trajectory
“I think it will. You know, some people say Moore's law is, is starting to come to an end in terms of just more transistors, but we barely started on neural hardware. So I'm sure we're going to see the same effect, even if it's not more transistors per square i…”
Cohen: Undertrained large AI models underperform well-trained smaller models
“If you have a large model, That is undertrained. It will underperform a small model, which is really well, like well, well trained. So you're just wasting resources and you're going to get like less efficient results.”
Belsky: Product designers are more important, not less, in the AI era
“It's not the technology that makes us successful. It's the user's experience of the technology that makes us successful. And that's why I think the role of designers is more important, not less important in this modern world.”
Söderström: Hiring AI technical talent is getting easier for Spotify
“Initially, it was very hard to find talent on the technology side. That's now getting easier.”
Söderström: Spotify ran an internal bet called 'AI is the product'
“We literally had, have had, we have a bets board where we kind of stack crank what we're doing. We literally had a bet called AI is the product. For like two years to really emphasize the shift, getting people in that mindset”
Belsky: Adobe's design team led the early development of Firefly
“You know, we had our design team actually build a team within that did all of the early developments of Firefly. It was a very like design driven exercise to figure out the interfaces and how they integrated into the products.”
Söderström: Proprietary user data will matter more long term than model size
“My hunch would be though that In the longer term, you're going to be able to do most of what you want with reasonable size models and like having a much bigger model doesn't have that much. So I would still bet that having lots of user data and lots of high fi…”
Söderström: Spotify will not build general frontier AI models to compete
“So that's why, you know, there's no point in trying to compete with them for us because we don't even have the same goal. So we're trying to build exactly the things that we don't think they will build. And we would buy the things that they will build or someo…”
Belsky: Companies should only build AI models they can lead globally
“Actually, I think that's sort of the right answer is you should build models if you're the best company in the world to build them.”
Cohen: A product leader's primary job is defining algorithm objectives
“My opinion as a product leader or a designer or an engineer or a product manager, like data is literally at your second most important job in your role. It's understanding how do you basically start to infuse all of data collection and make sure it's high qual…”
Belsky: Adobe never trains AI models on customer creations without opt-in
“We don't ever train our models off of the customer's creations. So, like, what the customer makes does not train our models, unless they submit it to stock and want it to train our models in which case they get compensated for it.”
Söderström: AI has not yet enabled new business models beyond sustaining innovation
“I actually think that in general, you know, we've seen the first wave of AI now where the technology is very impressive. It's starting to do useful things. But we still, usually the big changes come when that technology enables a new business model somehow. An…”
Belsky: AI platform shift uniquely favors incumbent tech giants over startups
“I do think we executed well, but I also think that there's some nuances of this platform shift in particular that do favor those companies that know their customer well and already reach them.”
Söderström: Spotify's streaming business model took Apple eight years to copy
“I think for Spotify, for example, when we started, it was the business model of streaming. That was the disruptive thing. The technology was you know, a bit of peer to peer and stuff, but it was a business model that took Apple eight years of collateral damage…”
Belsky: Per-seat pricing for enterprise SaaS is obsolete in the AI era
“The idea of selling seats function by function, like how many people in procurement, how many people in financial planning? I mean, it's such an old antiquated way of Building a business to some degree in the age of AI.”
Stebbings: 32% of European corporate enterprises do not know what Slack is
“Like, you know, I think as a stat, 32% of European corporates do not know what Slack is.”
Söderström: Enterprise AI adoption will outpace the historical shift to cloud
“But I also agree with Scott that I think it's going to go much faster than moving to cloud. It wasn't really possible. I know we actually were on-prem companies that old. It wasn't possible to just also be on the cloud and try a little bit. It was incredibly h…”
Cohen: Job skillsets changed by 25% over the past five years
“The skillset necessary to do a job changed by 25% in the last five to six years.”
Cohen: Global job skillsets will change by at least 65% by 2030
“By twenty-thirty, they're going to change by at least 65%.”
Cohen: AI delivers the most value when simplifying complex software products
“The more complex your product, the more impact AI could have in terms of simplifying it for your user base.”
Cohen: LinkedIn is using LLMs to create interfaces that morph to users
“One thing that we're already underway for us is to really take away the complexity and solve it with the idea of bringing in more of this new large language models, and in general, this new brain that can exceed on top of it. So instead of adding more features…”
Belsky: Adobe recently re-centralized its design team into a single organization
“Recently I re-centralized design in one organization because our strategy, you know, required it.”