# Tiny Teams

Britney Walker, CRV & Eric Simons, StackBlitz & Sid Bendre, Oliv & Max Broer Herbas, Gum Loop & Grant, Gamma & Vic Paruturi, Data Lab & Alex Duffy, Every | AI Engineer World's Fair 2025 | 3:37:25

Source: https://www.youtube.com/watch?v=xhKgTkzSmuQ
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/tiny-teams
Published: 2025-06-05
Tags: engineering-culture, evals, product-strategy, team-adoption

## TL;DR
- Small teams can survive sudden demand when they have shared context, high trust, and the discipline to focus on a small set of high-impact work.
- Oliv, Gum Loop, Gamma, and Data Lab use generalists, automation, simple systems, and direct customer contact to grow without adding headcount at the usual rate.
- Benchmarks influence what AI companies build, so builders should create evaluations that test useful, human-centered abilities rather than only narrow scores.

## Summary
This session brings together founders and operators from several small AI companies. Eric Simons describes how StackBlitz launched Bolt.new with fewer than 20 people, then handled a surge of customers through shared context, direct user contact, strict prioritization, and AI-assisted support. Sid Bendre explains Oliv's four-person approach to consumer software, built around profitable growth, KPI ownership, generalist hiring, reusable playbooks, and internal automation. Max Broer Herbas describes Gum Loop's selective hiring, paid work trials, sparse meeting culture, customer-led product work, and use of its own automation platform. Grant argues that Gamma has stayed small by hiring generalists and player-coaches while maintaining a strong culture. Vic Paruturi applies a similar model to document intelligence, with senior generalists working across customers, research, data, and product. Alex Duffy closes by arguing that benchmarks spread like memes and shape model development, so people should build evaluations around creativity, real-world use, and human goals.

## Key ideas
### Shared context lets a small team respond to sudden demand
[03:14](https://www.youtube.com/watch?v=xhKgTkzSmuQ&t=194s)
Eric Simons says Bolt.new launched when StackBlitz had fewer than 20 people and the company expected it might shut down. The product was brittle, support demand was intense, and neither Simons nor his chief of staff had support in their job titles, yet they handled many tickets themselves. The team had already worked together for seven years, which gave them shared context, trust, and speed. Simons connects this to a long-standing company rule: keep a small number of people with more context per head, so they can act without waiting for permission or passing work through layers of management.

### A lower burn rate gives startups more chances to find product-market fit
[07:50](https://www.youtube.com/watch?v=xhKgTkzSmuQ&t=470s)
Simons compares early startup work to an enterprise sales funnel. Teams need to take many shots before one works, so they need enough runway to keep trying. Humans are the most expensive part of a company, and adding headcount reduces the time available for experiments. He gives the example of a former competitor that was acquired and stripped for parts two weeks before Bolt launched. In his view, the company may have been capable of competing, but it did not have enough runway to reach the other side. A small team also has to let some fires burn, since it cannot address everything at once.

### Direct user contact can stabilize a fragile launch
[14:19](https://www.youtube.com/watch?v=xhKgTkzSmuQ&t=859s)
During Bolt.new's first week, Simons believed users would churn if the founders and team stayed invisible while the product remained brittle. The company began weekly office hours on YouTube and X, showing users what it was building, acknowledging problems, and explaining where fixes were headed. He says this kind of work is hard to quantify, but it creates user love and helps maintain belief while the product improves. He also argues that community can extend customer support without adding company headcount. StackBlitz combined community help with AI support and a large hackathon.

### Oliv compounds learning through generalists, KPIs, and reusable systems
[25:31](https://www.youtube.com/watch?v=xhKgTkzSmuQ&t=1531s)
Sid Bendre describes Oliv's approach as a combination of operating principles, organization, and AI tooling. The company hires people with several complementary strengths, such as product engineers who are full stack and strong product thinkers, marketers who can code, and designers who can build. Each person owns a KPI, which Bendre says reduces micromanagement because decisions are judged against a measurable outcome. Oliv also treats repeated failures as system problems and builds technical playbooks and operational blueprints that can be reused across products. This helped the company transfer learnings from Quizard to Unstuck AI.

### A tiny team can divide product ownership from shared automation work
[32:25](https://www.youtube.com/watch?v=xhKgTkzSmuQ&t=1945s)
Oliv organizes engineers into harvesters and cultivators. Harvesters own individual products, live in the metrics, run experiments, build features end to end, and work with marketing. Cultivators build the company's agentic operating system and automate work across areas such as marketing, design, and product. Bendre says AI tools should turn a strong generalist into a much stronger contributor rather than compensate for someone who lacks the required standard. The company also uses LaunchDarkly beyond feature releases, routing model traffic, changing infrastructure priorities when services fail, and running UI and paywall experiments without code pushes.

### Gum Loop hires slowly and tests people by having them do the work
[41:41](https://www.youtube.com/watch?v=xhKgTkzSmuQ&t=2501s)
Max Broer Herbas says Gum Loop is deliberately selective because every person has a large effect on a team of fewer than 10. The company has interviewed hundreds of candidates and uses practical work trials rather than relying on abstract interview performance. Candidates work in the codebase or join a short retreat as contractors, with payment for their time. Two customers eventually joined the company because they already understood the product and wanted to build it. Herbas calls this product-led hiring. The company also keeps meetings rare, gives people uninterrupted time, and expects strong hires to make progress without constant management.

### Generalists reduce handoffs when product work includes research and operations
[2:29:28](https://www.youtube.com/watch?v=xhKgTkzSmuQ&t=8968s)
Vic Paruturi argues that headcount does not equal productivity. At Data Lab, a small group trained an OCR model while covering customer conversations, paper reading, architecture, data cleaning, training, inference, and product integration. In a larger company, those tasks might be split among several teams, with context lost at every handoff. Data Lab instead hires senior generalists who can work across the stack, uses simple technology, reuses components, and fills repetitive edges with AI and internal tools. Paruturi says this approach creates faster feedback from customers to model development and avoids the meeting and coordination costs of specialized departments.

### Benchmarks spread ideas and steer model development
[3:01:00](https://www.youtube.com/watch?v=xhKgTkzSmuQ&t=10860s)
Alex Duffy describes benchmarks as memes in Richard Dawkins's sense: ideas that spread from person to person. A benchmark begins as someone's question about whether AI can do a particular task, becomes popular, and may eventually be used for training or evaluation until models saturate it. Duffy argues that benchmark creators therefore influence what model providers optimize for. He favors evaluations that are understandable, generative, creative, experiential, and able to become harder as models improve. His AI Diplomacy example tests negotiation and social strategy through a board game, revealing behavior that a static math or coding test would not capture.

## Notable quotes
- "You really want a small number of people with more context per head because what that means is that people at the company have more agency." (07:15)
- "Profit is power and profit is focus." (29:19)
- "If you get one thing from this talk, this is the thing: more people does not equal more productivity." (2:29:46)
- "The people here, the people that get that, the people that can build benchmarks are going to shape the future." (3:06:37)
- "The role of a human in an AI world is to define the goal and to define what's good and bad on route to that goal." (3:14:44)

## Tools & references mentioned
- CRV
- StackBlitz
- Bolt.new
- Oliv
- Quizard AI
- Unstuck AI
- OpenAI Codex
- LaunchDarkly
- Azure OpenAI
- Palantir
- Gum Loop
- Cursor
- Windsurf
- Gamma
- Data Lab
- Marker
- Surya OCR3
- Jeremy Howard
- Answer.AI
- fast HTML
- MonsterUI
- HTMX
- Alpine.js
- Every
- Context Window
- AI Diplomacy
- Claude Opus
- Claude
- Gemini 2.5 Pro
- Gemini 2.5 Flash
- OpenAI o3
- Llama 4 Maverick
- DeepSeek R1
- Richard Dawkins
- Simon Willison

## Who should watch
- Founders deciding whether to add headcount after a funding round will get concrete arguments for staying small and ways to judge when that approach is failing.
- Engineers and product leaders building AI products can compare hiring, support, internal automation, customer feedback, and organization practices across several companies.
- People designing evaluations for AI systems will find a case for benchmarks that test real-world behavior, creativity, negotiation, and human goals.

## Related talks

- [Small AI Teams with Huge Impact](https://aietalks.com/talks/small-ai-teams-with-huge-impact) (Vikas Paruchuri, Datalab, 17:36)
- [Building a 10 person unicorn](https://aietalks.com/talks/building-a-10-person-unicorn) (Max Brodeur-Urbas, Gumloop, 12:03)
- [Rethinking Team Building: How a 30-Person Startup Serves 50 Million Users](https://aietalks.com/talks/rethinking-team-building-how-a-30-person-startup-serves-50-million-users) (Grant Lee, Gamma, 18:06)
- [The New Lean Startup](https://aietalks.com/talks/the-new-lean-startup) (Sid Bendre, Oleve, 13:26)
- [How to build world-class AI products](https://aietalks.com/talks/how-to-build-world-class-ai-products) (Sarah Sachs, Notion & Carlos Esteban, Braintrust, 1:43:46)
