Research Team Dynamics

Exploring patterns in solo work, small teams, and collaboration | January 2026

About This Research

What this document explores: Patterns in when solo/remote work succeeds vs. when teams/co-location appear beneficial—looking at historical examples, technology companies, and specifically AI research. Includes profiles of researchers, analysis of 20+ AI papers by team size, and frameworks for thinking about your own situation.

Key observations from the data:

Historical Patterns of Solo & Remote Success

Work Characteristics That Appear to Enable Solo Success

CharacteristicPotential ReasonExamples
Long feedback loopsLess need for real-time coordinationDarwin (20 years), Einstein (8 years)
Objective verificationWork can speak for itselfMathematics, code, experiments
Deep focus as bottleneckInterruptions may be costlyWriting, research, complex coding
Symbolic/digital outputEasy to share remotelyPapers, software, proofs

Where Solo Work Has Succeeded

  • Mathematics: Ramanujan, Perelman, Grothendieck
  • Theoretical Physics: Einstein's miracle year, Newton during plague
  • Writing: Most books
  • Open Source: Linus wrote Linux in his bedroom

Where Teams Appear Necessary

  • Experimental Physics: CERN, Manhattan Project
  • Large-Scale Engineering: Bridges, rockets, OS at scale
  • Time-Sensitive Work: Journalism, trading, emergency response
PATTERNS OBSERVED ───────────────────────────────────────────────────────────── Solo/Remote tends to appear in: Teams tend to appear in: ───────────────────────────────── ──────────────────────────── Long feedback loops Short feedback loops Objective verification Subjective/political decisions Deep focus is bottleneck Coordination is bottleneck Symbolic output Physical output Exploration phase Execution phase Novel ideas Scaling ideas

Remote-First Companies — Evidence

Successful Remote-First Companies

Company Type Remote Status Valuation
ShopifyPublic"Digital-by-default" (May 2020)~$188B
CoinbasePublicRemote-first, no HQ~$61B
GitLabPublicAll-remote from founding~$5.6B
AutomatticPrivateRemote-first since founding~$7.5B
DeelPrivateRemote work infrastructure~$17.3B
PostHogPrivate100% remote, 20+ countries$1.4B
SupabasePrivateBorn-remote$5B
LinearPrivateFully remote from start$1.25B

Research on Remote Work & Innovation

Nature study (2023): Analyzed 20M papers + 4M patents. Found remote teams were less likely to produce breakthrough discoveries vs on-site teams in their dataset.

One interpretation: Remote teams may do more technical work while in-person teams do more conceptual work.

Counterpoint (Oxford 2022): Long-distance collaboration correlated with breakthroughs post-2010, possibly due to better collaboration tools.

Important Context

These studies are pre-AI coding tools. Human + AI collaboration may change these dynamics significantly.

AI Research — Observed Patterns

The 1 → Few → Many Trajectory

A pattern that appears repeatedly in AI research:

  1. Individual vision/prototype — often solo or small team
  2. Small team validates and refines — 2-6 people
  3. Scale execution — 10-30+, where coordination becomes important
OBSERVED PATTERNS IN AI WORK ───────────────────────────────── ──────────────────────────── Often done by small teams: Often requires larger teams: ───────────────────────────────── ──────────────────────────── Interpretability research Frontier model training Safety/alignment theory Large-scale data operations Fine-tuning & applications Infrastructure at scale Open source contributions Regulatory/policy navigation Novel architectures (early) Multi-modal integration Blog posts that build reputation Production deployment

Potentially Solo-Friendly Opportunities in AI

  1. Mechanistic interpretability — Chris Olah's reputation came from solo blog posts
  2. Fine-tuning and applications — Accessible without training frontier models
  3. Safety research — May favor deep thinking over coordination
  4. Open source tools — Karpathy's nanoGPT, Hotz's tinygrad
  5. Evaluations and benchmarks — High leverage, doesn't require massive compute
  6. Novel research directions — Early exploration may not need teams

AI Breakthroughs — Team Size Analysis

Major Breakthroughs by Team Size

BreakthroughYearAuthorsSizeKey Driver(s)
Backpropagation19863SmallHinton
LSTM19972DuoHochreiter & Schmidhuber
AlexNet20123SmallKrizhevsky
Attention20143SmallBahdanau
GANs20148MediumGoodfellow
Batch Normalization20152DuoIoffe & Szegedy
ResNet20154SmallKaiming He
Transformers20178MediumEqual contributors
GPT-120184SmallRadford
BERT20184SmallDevlin
GPT-220196MediumRadford
DDPM (Diffusion)20203SmallHo
Scaling Laws20209MediumKaplan/McCandlish
GPT-3202014+LargeBrown/team
2-4
Typical team size for early conceptual work
6-14+
Typical team size for scaling work
98K+
LSTM citations (2 authors)
173K+
Transformer citations (8 authors)

An Interesting Pattern

Early era (1986-2000): Ideas and people seemed to be the constraint → small teams

Modern era (2018+): Compute and engineering seem to be constraints → larger teams for execution

Observation: Team size growth appears more correlated with implementation and scaling than with ideation. The conceptual breakthrough stage often remains small-team territory.

How Team Sizes Have Evolved

EraTypical Team SizePrimary Bottleneck
1986-20001-3Ideas, people
2009-20143-6Ideas, data
2014-20184-8Ideas, compute
2018-20206-14Compute, engineering
2020+10-100+ (industry)Scale, deployment

Researcher Profiles

Alec Radford Independent

Sam Altman called him "Einstein-level genius" — key figure behind GPT

Work Style

Key Contributions

PaperYearTeamRole
GPT-120184First author, built independently
GPT-220196Co-first author (with Wu)
CLIP202112First author
Whisper20226First author

Current: Left OpenAI Dec 2024. Independent researcher collaborating with multiple AI orgs.

Pattern

Stayed hands-on throughout career. Chose autonomy over institutional resources.

Chris Olah Technical Leader

Pioneer of mechanistic interpretability — TIME 100 Most Influential in AI 2024

Work Style

"We can understand them so well that we can go and hand-write weights."

Reputation Built From Solo Work

Current: Co-founder at Anthropic. Leads interpretability. Still hands-on.

Pattern

Technical credibility came from solo blog posts. Still codes as a co-founder.

Ilya Sutskever Research CEO

Co-founder & Chief Scientist at OpenAI → CEO of SSI

Work Style

"The age of scaling is ending... the next breakthroughs now depend on new learning methods, not more GPUs."

Current: SSI (July 2025)

Pattern

Built a structure where CEO = researcher. The entire company is a research lab.

Dario Amodei Full CEO

VP of Research at OpenAI → CEO of Anthropic

At OpenAI

Current (2025)

Pattern

Maximum leverage path. Uses technical credibility but doesn't do research anymore.

Frontier/Adjacent Researchers

Michael Edward Johnson Frontier Independent

Consciousness researcher — Co-founded Qualia Research Institute, now Symmetry Institute

Work Style

Key Contributions

Pattern

Demonstrates independent consciousness research is viable. Built reputation through long-form writing and frameworks, not traditional academia.

Michael Levin Frontier Academic Lab

Bioelectricity & collective intelligence — Director of Allen Discovery Center at Tufts

Work Style

Key Contributions

AI Connection

Pattern

Shows how frontier research can span biology, AI, and philosophy. Academic lab but intellectually operates like an independent researcher pursuing unconventional directions.

The Researcher Spectrum

PURE RESEARCHER ←――――――――――――――――――――――→ PURE CEO

Radford
Independent researcher
Olah
Technical leader who still codes
Ilya
CEO of research-only startup
Dario
CEO of large company (700+)

Different researchers have found different points on this spectrum. There's no single "right" path — each represents a different tradeoff between hands-on work and organizational leverage.

The Scaling Laws Origin Story

Who Discovered Scaling Laws?

The foundational paper (January 2020) formalized what OpenAI had believed intuitively.

Lead authors: Jared Kaplan and Sam McCandlish
Interesting note: Both were theoretical physicists, not traditional ML researchers.

Author List

Why Physicists?

Physicists are trained to look for universal scaling relationships. They may have seen AI capabilities vs compute as a physical phenomenon with predictable laws — similar to how physical systems follow power laws.

Most of this team later left to co-found Anthropic.

GPT Timeline

June 2017
Transformer paper — New architecture for translation. The LLM paradigm wasn't yet obvious.
June 2018
GPT-1 — Proved unsupervised pre-training works. 117M parameters.
February 2019
GPT-2 — OpenAI called it "too dangerous to release." 1.5B parameters.
January 2020
Scaling Laws paper — Formal proof that scale = predictable capability gains.
June 2020
GPT-3 — 175B parameters. The paradigm became widely accepted.

Radford's Core Insight

"The underlying generative model learns to perform many of the tasks we evaluate on in order to improve its language modeling capability."

Translation: The model learns QA, translation, summarization as a side effect of predicting the next word. This was a conceptual breakthrough. Scale came later.

Location Considerations

What Tech Hubs Like SF Provide

Potential AdvantageAvailable Remotely?
Investor accessPartially — remote fundraising is more common now
Talent densityYes — hire globally
Serendipitous collisionsHarder to replicate remotely
Compute accessYes — cloud, partnerships
Credibility signalingPartially

You may not need a hub if:

  1. You have a clear research direction
  2. You have compute access
  3. You have 1-2 high-bandwidth collaborators
  4. You're building reputation through artifacts (papers, code, blog posts)

A hub might help if:

  1. Raising significant capital
  2. Recruiting from existing talent pool
  3. Early-career, need to build network
  4. Your work benefits from frequent in-person collaboration

AI as Force Multiplier

How AI Tools Are Changing Individual Capacity

Old ConstraintHow AI May Change It
Coding bandwidthCursor, Copilot, Claude Code
Design bandwidthMidjourney, Figma AI
Research bandwidthLiterature review, synthesis
Ops bandwidthAutomated monitoring, agents
Potential implication: Work that previously required 5-10 people may now be achievable by 1-2 people. This could dramatically expand the viable solo/remote surface area.

The Indie Research Movement

Framework for Thinking About Solo/Remote Viability

Questions to Consider

QuestionSolo/Remote May Work BetterTeams May Work Better
How long is the feedback loop?Days/weeks/monthsHours/minutes
How is quality verified?Objectively (tests, math)Subjectively (opinions)
What's the bottleneck?Deep thinkingCoordination
What's the output?Digital/symbolicPhysical
What phase is the work?ExplorationExecution at scale

Potentially high solo/remote viability:

  • Theoretical research
  • Writing (books, papers, code)
  • Design (once vision is clear)
  • Early-stage product development
  • Open source maintenance
  • Content creation

Potentially lower solo/remote viability:

  • Large-scale engineering
  • Experimental physics (equipment)
  • Team sports and performance
  • Manufacturing
  • Emergency response

Key Observations

1. Solo/small-team research has produced major work

Radford built GPT-1 independently. Olah's reputation came from solo blog posts. Johnson operates entirely outside academia.

2. Many conceptual breakthroughs came from small teams (2-4)

LSTM (2), Attention (3), AlexNet (3), GPT-1 (4), Diffusion (3). Often traceable to 1-2 key individuals.

3. Team sizes appear to grow for execution

GPT-3's 14 authors reflects scaling/engineering work. Conceptual breakthrough stage often remains small-team territory.

4. Research is increasingly distributed

Many successful AI labs operate with distributed or remote teams. Location may matter less than it once did.

5. The 1 → few → many pattern appears often

Individual vision, small team validation, large team scaling.

6. Some researchers stay hands-on at senior levels

Radford and Olah demonstrate this is possible, but it requires deliberate choice.

7. Frontier research happens in unexpected places

Michael Levin (bioelectricity), Michael Johnson (consciousness) are doing solo/small-team work that intersects with AI from outside traditional AI.

8. Work characteristics may matter more than industry norms

Long feedback loops, objective verification, and deep focus as bottleneck appear to favor solo work.

9. AI tools are expanding what's solo-viable

The capabilities available to individual researchers are increasing rapidly.

10. Distribution channels are diversifying

Blog posts (Olah), open source (Karpathy), Substack (Johnson) can build reputation without traditional publishing.

Summary

This research suggests that for certain types of work, solo or small-team approaches may be viable or even advantageous. The key factors appear to be the nature of the work itself — particularly feedback loop length, verification method, and whether the bottleneck is thinking or coordination. As AI tools continue to develop, the range of work that can be done by individuals or small teams may expand further.