Subscribe
Share
Share
Embed
The Horizontal vs Vertical AI Debate: Why This Ex-Meta AI Researcher Is Betting Big on Horizontal Web Agents
Should you build narrow (vertical) or go broad (horizontal) in AI? This episode unpacks why one PhD researcher abandoned his working vertical product to chase a much riskier horizontal bet - and why VCs leaning heavily into vertical AI might be missing something.
Abhishek Das is the co-founder and co-CEO of Yutori, which has raised over $15 million from Radical Ventures, Felicis, and prominent angels including Ali Gil, Sarah Guo, Scott Belsky, and Guillermo Rauch. Previously a research scientist at Meta's FAIR lab, Abhishek holds a PhD from Georgia Tech where he pioneered work on AI agents that can see, talk, and act starting in 2016.
In Today's Episode We Discuss:
00:53 - Why how we interact with the web hasn't changed in three decades and what will break that
02:27 - The coming shift from manual browsing to AI assistants performing tasks in the background
05:57 - What "agents" actually meant in ML research before the term became overloaded
06:14 - Why 90% accuracy per step creates catastrophic failure rates over multi-step workflows
08:46 - The behavior pattern humans nail intuitively that machines struggle with: backtracking from errors
10:11 - The DoorDash experiment: building an end-to-end food ordering agent that never shipped
12:58 - Why training on sinle websites leads to memorization instead of generalization
13:03 - The dopamine problem: some tasks users don't want automated
15:08 - Why capability-scoped beats website-scoped: the pivot to read-only horizontal agents
18:05 - Three criteria that drove the horizontal decision: research, user value, and data strategy
24:18 - Scouts API launch: why different channels have different risk appetites for web agents
26:30 - Flying close to the sun: how Yutori competes with hyperscalers on horizontal AI
30:32 - What VCs should actually test for in horizontal AI teams beyond founder horsepower
32:10 - Why three-month roadmaps are the only reasonable planning horizon in AI today
33:05 - The dogfooding ritual: every team member rotates through user feedback weekly
34:50 - Why research and product can't be siloed and how ideas flow both directions
36:03 - The uncomfortable truth: end users don't care about your research breakthroughs
37:32 - The Nintendo Switch 2 problem: aggregating individual feedback into systemic fixes
39:35 - Reframing web agents as "buyer's agents" that filter the internet on your behalf
40:59 - The simulation bet: training agents on cloned websites for high-stakes irreversible actions
43:05 - Why initial team skepticism about Scouts' value proposition was completely wrong
45:01 - How scout reports contextualize results with reasoning and ingest feedback over time
47:52 - The core insight test: where does your instinct lie across research, market, and domain?
49:36 - The hiring trap: why preemptively hiring sales leadership to impress VCs backfires
51:18 - The 12-year-old advice that still guides him: "Be a sponge when entering a new space"
53:05 - Non-negotiables: walking the dog with podcasts and personally reading every user email
54:49 - What founders actually need from VCs: direct and timely feedback, not just capital
Should you build narrow (vertical) or go broad (horizontal) in AI? This episode unpacks why one PhD researcher abandoned his working vertical product to chase a much riskier horizontal bet - and why VCs leaning heavily into vertical AI might be missing something.
Abhishek Das is the co-founder and co-CEO of Yutori, which has raised over $15 million from Radical Ventures, Felicis, and prominent angels including Ali Gil, Sarah Guo, Scott Belsky, and Guillermo Rauch. Previously a research scientist at Meta's FAIR lab, Abhishek holds a PhD from Georgia Tech where he pioneered work on AI agents that can see, talk, and act starting in 2016.
In Today's Episode We Discuss:
00:53 - Why how we interact with the web hasn't changed in three decades and what will break that
02:27 - The coming shift from manual browsing to AI assistants performing tasks in the background
05:57 - What "agents" actually meant in ML research before the term became overloaded
06:14 - Why 90% accuracy per step creates catastrophic failure rates over multi-step workflows
08:46 - The behavior pattern humans nail intuitively that machines struggle with: backtracking from errors
10:11 - The DoorDash experiment: building an end-to-end food ordering agent that never shipped
12:58 - Why training on sinle websites leads to memorization instead of generalization
13:03 - The dopamine problem: some tasks users don't want automated
15:08 - Why capability-scoped beats website-scoped: the pivot to read-only horizontal agents
18:05 - Three criteria that drove the horizontal decision: research, user value, and data strategy
24:18 - Scouts API launch: why different channels have different risk appetites for web agents
26:30 - Flying close to the sun: how Yutori competes with hyperscalers on horizontal AI
30:32 - What VCs should actually test for in horizontal AI teams beyond founder horsepower
32:10 - Why three-month roadmaps are the only reasonable planning horizon in AI today
33:05 - The dogfooding ritual: every team member rotates through user feedback weekly
34:50 - Why research and product can't be siloed and how ideas flow both directions
36:03 - The uncomfortable truth: end users don't care about your research breakthroughs
37:32 - The Nintendo Switch 2 problem: aggregating individual feedback into systemic fixes
39:35 - Reframing web agents as "buyer's agents" that filter the internet on your behalf
40:59 - The simulation bet: training agents on cloned websites for high-stakes irreversible actions
43:05 - Why initial team skepticism about Scouts' value proposition was completely wrong
45:01 - How scout reports contextualize results with reasoning and ingest feedback over time
47:52 - The core insight test: where does your instinct lie across research, market, and domain?
49:36 - The hiring trap: why preemptively hiring sales leadership to impress VCs backfires
51:18 - The 12-year-old advice that still guides him: "Be a sponge when entering a new space"
53:05 - Non-negotiables: walking the dog with podcasts and personally reading every user email
54:49 - What founders actually need from VCs: direct and timely feedback, not just capital
Creators and Guests
Host
Pascal Unger
Co-founder / Managing Partner at focal.vc We lead pre-seeds in North America with up to $1M: We exclusively back AI native software startups at the very start. Thereafter, we help them get off the ground better and faster, supported by a network > 200 GTM executives. That's our singular, unwavering focus.
What is focal podcast?
Pivotal early lessons of today's best startups.
Welcome to the focal podcast where we go deep with some of today's best founders and operators on ONE crucial lessons from their early days.
This podcast is not the usual "highlight reel" startup podcast that goes one inch deep across 20+ topics. Rather, we ask the questions you’d ask if you were sitting across from them. No fluff, just the real, actionable insights you’d get if these founders were mentoring you 1on1.
We cover topics including:
- What worked and why.
- Costly mistakes and how they fixed them.
- Frameworks that truly made a difference.
- Tactics to move faster.
- What they wish they’d known sooner.
- And much more!
"Only a fool learns from their own mistakes. The wise learn from the mistakes of others."