Solo plus AI matches a team
But top results still need teams with AI. 776-person trial; 160,000+ Product Hunt launches.
Dell'Acqua et al. peer-reviewed Kim et al. preprint
2026–2030 · Lisbon · London · Rome · Zurich
The science of AI entrepreneurship
How AI changes who can found a company, and what they build. Five studies, 350 teams a year.
Scale350 randomised teams a year. Four sites.
Hard outcomes$10K in 30 days. $30K MRR in 60. Verified, not reported.
Your name on itOpen benchmarks. One pre-registered hypothesis a year.
Why now
Seven findings that define the gap, as causal structure.
2025–26 literature
+increases−decreases?untestedArrows show the programme's reading, not the papers' claims.
Value depends on where AI sits
Teaching 515 startups where AI fits raised revenue 1.9× and cut capital needs 39.5%.
Kim working paper
Long-horizon autonomy is unsolved
Agents derail in Vending-Bench. 3 of 12 models stay profitable in YC-Bench. Best completes ~30% of StartupBench.
Backlund & Petersson preprint He et al. (Collinear AI) preprint Zhu et al. preprint
Revenue is not a clean reward
Self-improvement works on benchmarks, but evaluators must co-evolve to stop reward gaming.
Zhang peer-reviewed
Human control is untested with real money
L1–L5 autonomy and oversight regimes are formalised, but never run with founders and live revenue.
“Comparing Human Oversight Strategies for Computer-Use Agents” preprint
AI ideation homogenises
New-to-the-world creation needs measurement and de-fixation.
Wenger & Kenett peer-reviewed
Physics can constrain business dynamics
Neural operators embed governing equations; neuro-symbolic frameworks impose economic constraints.
Ahmadi peer-reviewed “ARTEMIS: A Neuro-Symbolic Framework for Economically Constrained Market Dynamics” preprint “PhyWorld: Physics-Faithful World Model for Video Generation” preprint
At a glance
Five studies, one annual cycle. Each feeds the next.
| # | Study | Window | Units | Status |
|---|---|---|---|---|
| 1 | Course | Jul–Aug · 30 days | 350 teams, randomised | In design |
| 2 | Accelerator | Sep–Nov · 60 days | 15 teams + 15 controls | In design |
| 3 | Simulation | Dec–Mar | Off-cohort | In build |
| 4 | Novelty | Continuous | All cohort ideas | Running |
| 5 | FounderBench | Continuous | 350 teams a year | Running |
Jul–Aug · CourseAug–Sep · SelectionSep–Nov · AcceleratorNov–Dec · AnalysisDec–Mar · SimulationApr–Jun · Pre-registration
The studies
What each asks, where it stands, how to join.
Open a study
In design. First cohort July 2027. Join before March 2027 to shape it.
Question
Can a founder-controlled agent system reach $10K revenue in 30 days — and which control setup keeps the founder in charge?
Design
350 teams, randomised in three strata. Solo (100): CEO plus product and GTM agents. Pair (200): CEO and CPO, AI as CMO, 2×2 oversight by self-improvement. Trio (50): human CMO owns GTM. Agents never hold funds.
Endpoint
Verified revenue at day 30. Power d≈0.57 per 50-team arm.
References
- Dell'Acqua et al., “The Cybernetic Teammate”, Organization Science peer-reviewed
- Kim, Kim & Koning, “Mapping AI into Production”, INSEAD WP 2026/20/STR working paper
- Ju & Aral, “Collaborating with AI Agents”, arXiv:2503.18238 preprint
- Cai et al., “AI as ‘Co-founder’”, arXiv:2512.06506 preprint
- Kim et al., “Generative AI Fuels Solo Entrepreneurship, but Teams Still Lead at the Top”, arXiv:2605.10291 preprint
- “Comparing Human Oversight Strategies for Computer-Use Agents”, arXiv:2604.04918 preprint
- “Principles and Guidelines for Randomized Controlled Trials in AI Evaluation”, arXiv:2605.02050 preprint
- Becker, Rush, Barnes & Rein (METR), “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”, arXiv:2507.09089 preprint
Repos
- METR/Measuring-Early-2025-AI-on-Exp-OSS-Devs
- MAST
- NVIDIA/physicsnemo
- obastani/GenAICanHarmLearning
Data
- EIA cohort traces and payment exports
- Y Combinator directory (5,404 firms, CC0)
- Antler portfolio (~1,600 firms)
- Product Hunt API
- Census BFS, Eurostat
- PDL company data (32.3M rows)
In design. First cohort September 2027. 25 teams in 2028.
Question
Can founders hand agents more autonomy over 60 days while the revenue is real — and where is the control-value frontier?
Design
15 teams, ranks 16–30 as a comparison group. Stepped-wedge L2 to L4 over ~120 team-weeks, crossover on self-improvement cadence, weekly derailment audit. MRR fitted as an acquisition-churn model.
Endpoint
$30K verified MRR at day 60, 20+ customers, churn under 15%.
References
- Backlund & Petersson, “Vending-Bench”, arXiv:2502.15840; Vending-Bench 2 preprint
- He et al. (Collinear AI), “YC-Bench”, arXiv:2604.01212 preprint
- Chen, Narasimhan & Liu, “CEO-Bench”, arXiv:2606.18543 preprint
- Zhu et al., “StartupBench”, arXiv:2608.17800 preprint
- Cemri et al., “Why Do Multi-Agent LLM Systems Fail?”, NeurIPS 2025 D&B peer-reviewed
- Zhang, Hu, Lu, Lange & Clune, “Darwin Gödel Machine”, ICLR 2026 peer-reviewed
- Kwa et al. (METR), “Measuring AI Ability to Complete Long Tasks”, NeurIPS 2025 peer-reviewed
- “ARTEMIS: A Neuro-Symbolic Framework for Economically Constrained Market Dynamics”, arXiv:2603.18107 preprint
Repos
- collinear-ai/yc-bench
- zlab-princeton/ceobench-src
- jennyzzt/dgm
- codelion/openevolve
Data
- EIA 8-wave panel and control group
- Baremetrics Open Startups, Indie Hackers
- MAST-Data (1,642 traces)
- YC-Bench, CEO-Bench run logs
- Census and Eurostat survival rates
In build now. First training run December 2026.
Question
Can a world model calibrated on real cohorts reproduce their outcomes and evolve agents that transfer to real customers?
Design
Six work packages. World-model training with one site held out. FounderSim. The L5 counterfactual. Sim-evolved agents in a pre-registered arm. Synthetic customers. The physical-product branch.
Endpoint
Docking: simulated effect sizes fall inside real confidence intervals.
You get
The world model is the licensable asset. Compute partners are named in the physical-product branch.
Compute is the binding constraint on physical products.
References
- Chu et al., “Agentic World Modeling”, arXiv:2604.22748 preprint
- “World Models: A Comprehensive Survey”, arXiv:2606.00133 preprint
- Hafner, Yan & Lillicrap et al., “Training agents inside of scalable world models” (Dreamer 4), arXiv:2509.24527 preprint
- Assran et al., “V-JEPA 2”, arXiv:2506.09985 preprint
- Agarwal et al., “Cosmos World Foundation Model Platform for Physical AI”, arXiv:2501.03575 preprint
- Binz et al., “A foundation model to predict and capture human cognition” (Centaur), Nature 644 peer-reviewed
- Teutloff, “Synthetic Founders”, arXiv:2509.02605 preprint
- Ahmadi, Cao, Humphrey & Karniadakis, “Physics-Informed Machine Learning in Biomedical Science and Engineering”, Annual Review of Biomedical Engineering peer-reviewed
Repos
- marcelbinz/Psych-101
- facebookresearch/vjepa2
- nvidia-cosmos/cosmos-predict2.5
- QwenLM/E-CommerceBench
Data
- EIA course and accelerator traces
- PDL, YC and Antler directories
- Census, Eurostat, OECD
- Amazon Reviews 2023, Product Hunt
- Psych-101 (10M+ choices)
- The Well (~15 TB), PDEBench
Running now against public corpora, ahead of 2027.
Question
Are AI-assisted teams converging on the same products, and does idea diversity predict revenue?
Design
Novelty locked on days 1–5 against the cohort, prior cohorts and a frozen LLM idea bank. Weekly convergence, feasibility-constrained novelty, blinded expert ratings, a de-fixation arm.
Endpoint
Outcome-linked, feasibility-constrained novelty and diversity metrics.
You get
Measures AI-driven homogenisation across ~350 real teams a year, with novelty locked before build.
Open needs: blinded raters, patent data access.
References
- Wenger & Kenett, “Large language models are homogeneously creative”, PNAS Nexus 5(3) peer-reviewed
- Sourati, Ziabari & Dehghani, “The Homogenizing Effect of LLMs on Human Expression and Thought”, arXiv:2508.01491 preprint
- “Examining and Addressing Barriers to Diversity in LLM-Generated Ideas”, arXiv:2602.20408 preprint
- Azad & Baten, “Ex Ante Evaluation of AI-Induced Idea Diversity Collapse”, arXiv:2605.06540 preprint
- “Anchorless Diversification for Parallel LLM Ideation”, arXiv:2605.30150 preprint
- “Epistemic Diversity and Knowledge Collapse in LLMs”, arXiv:2510.04226 preprint
- Bellemare-Pepin et al., “Divergent creativity in humans and LLMs”, Scientific Reports 16 peer-reviewed
- “IDEAFix”, arXiv:2606.00875 preprint
Repos
- arguslab-duke/llms_are_homogeneous
- SakanaAI/ShinkaEvolve
- phyworld
- keon/awesome-physical-ai
Data
- EIA idea and product corpus
- YC and Antler descriptions
- AI market maps, YC RFS taxonomy
- Product Hunt launches
- USPTO PatentsView, Google Patents
- Kickstarter corpus
Panel open now. First EIA labels 2027, v1 in 2028.
Question
Can 30- and 60-day outcomes be predicted before founding — prospectively, without leakage?
Design
Prospective labels, hybrid predictors from LLM embeddings and fitted dynamics, baselines from mentor scores, LLM judges and a human VC panel. Annual leaderboard with a human-control constraint.
Endpoint
Pre-registered AUROC gap over judges in 2027.
You get
Day-30 revenue and day-60 MRR labels for ~350 teams a year, plus 6/12/24-month follow-up. Not obtainable from historical data.
Join the leaderboard before the first labels land in 2027.
References
- Chen et al. (Oxford / Vela Research), “VCBench”, arXiv:2509.14448 preprint
- Maarouf, Feuerriegel & Pröllochs, “A fused large language model for predicting startup success”, EJOR 322(1) peer-reviewed
- Maarouf, Bakiaj & Feuerriegel, “Predicting Startup Success Using LLMs”, arXiv:2601.16568 preprint
- “PHBench”, arXiv:2605.02974 preprint
- Griffin et al., “Random Rule Forest”, arXiv:2505.24622 preprint
- Benhenda, “YC Bench: A Live Benchmark for Forecasting Startup Outperformance in Y Combinator Batches”, arXiv:2604.02378 preprint
- “Beyond Isolated Investor: Roleplay-Based Collective Agents”, arXiv:2512.22608 preprint
- Ihlamur, “When Career Data Runs Out”, IEEE IDS 2026 peer-reviewed
Repos
- vcbench.com
- benstaf/ycbench
- collinear-ai/yc-bench
- NVIDIA/physicsnemo
Data
- EIA prospective panel
- VCBench (9,000 profiles)
- YC directory, Antler portfolio
- Product Hunt, PHBench
- Crunchbase, Dealroom
- Census, Eurostat
World model
Predicts how demand, product, go-to-market and business move together under founder and agent actions.
Public priors, private posteriors. Public datasets pre-train and constrain the model. EIA's randomised cohort traces calibrate it.
Components
- EIA traces plus public priors
- State encoder: open LLM plus structured features
- Physics-informed dynamics core
- Evaluator calibrated on verified revenue
- Control policy: when to ask the founder
The loop the agents run, and the loop Studies 2 and 3 measure. Each turn is gated by the founder.
- Observe
- Roll out
- Score
- Founder gate
- Act
- Retrain
Demand
Structure: Diffusion ODEs, saturation, price elasticity
Learned: Neural ODE with residual loss
Priors: Product Hunt, Amazon Reviews, Census entry series
Product
Structure: Quality-adoption coupling; PDE feasibility for hardware
Learned: Neural operators, video world models
Priors: The Well, PDEBench, automotive CFD sets
Go-to-market
Structure: Funnel conservation, capacity queueing, channel saturation
Learned: Constrained sequence model
Priors: YC and Antler descriptions, Product Hunt metadata
Business
Structure: Cash conservation; MRR = acquisition minus churn
Learned: Hybrid grey-box model
Priors: Open Startups and Indie Hackers revenue series
Physics inside
Every model carries priors from known laws, so modest data still generalises.
Partner on compute- Demand obeys diffusionA model that violates saturation or price elasticity is rejected before a founder sees it.
- Business obeys conservationCash conservation and the MRR identity are imposed, not learned.
- Hardware is admissibleNeural-operator surrogates give structural and thermal feasibility, so physical products enter from 2027.
- Compute unlocks itThe physical branch needs pre-training on ~15 TB of simulation data. Our largest compute dependency.
How to join
Nine routes. What you give, what you get.
Benchmark
- Give
- An agent stack, plus compute or credits
- Get
- Live benchmark runs, co-authorship, licensed access
- Studies
- 1, 2, 3, 5
Sponsor
- Give
- Sponsorship and tool access for one stratum
- Get
- One pre-registered hypothesis a year, results under embargo
- Studies
- 1
Co-PI
- Give
- Method expertise, replication or PhD supervision
- Get
- Co-PI role, authorship, embargoed data
- Studies
- 1, 2, 3, 4, 5
Data
- Give
- Payments, analytics, cloud or simulation data
- Get
- Named register entry, visibility, anonymised derivatives
- Studies
- 1, 2, 3, 4, 5
Funding
- Give
- A Horizon, EIC or national-council slot
- Get
- Work packages, host universities, open outputs
- Studies
- 1, 2, 3, 4, 5
Cohort
- Give
- 60 days with your own new product line
- Get
- A controlled agent system. You keep product, customers, revenue
- Studies
- 2
University
- Give
- Ethics cover, PhD students, a host site
- Get
- Co-supervision, data-governance seat, publication plan
- Studies
- 3, 4
Corporate team
- Give
- An intrapreneur team and a real budget
- Get
- Autonomy protocol, control findings, derailment audit
- Studies
- 1, 2
Early access
- Give
- Data-use terms, optionally portfolio labels
- Get
- Prospective 30/60-day benchmark and leaderboard
- Studies
- 5
Data
Public priors, private posteriors.
Licence stated per source
EIA cohort tracesPrivate · Studies 1, 3, 4, 5 · Consent
EIA accelerator panelPrivate · Studies 2, 3, 5 · Consent
Y Combinator directory5,404 firms · Studies 1, 2, 4, 5 · CC0
Antler portfolio~1,600 firms · Studies 1, 2, 4, 5 · Public
PDL company data32.3M rows · Studies 1, 3, 4, 5 · Free
Product HuntLaunches · Studies 1, 3, 4, 5 · Public API
US Census BFSOfficial stats · Studies 1, 2, 3, 5 · Public domain
Eurostat, OECDOfficial stats · Studies 1, 2, 3, 5 · Open
Open Startups, Indie HackersMRR series · Studies 2 · Public
The Well, PDEBenchPhysics corpora · Studies 3 · Open
VCBench, PHBenchBenchmarks · Studies 2, 5 · Public
Psych-101Human choices · Studies 3 · Open
USPTO PatentsViewPatents · Studies 4 · Open
Crunchbase, DealroomFunding · Studies 5 · Licensed
Goals 2027–2030
What we expect to be judged against.
-
2026
in progress
Control layer and trace schema
Priors fitted on YC, Antler, Product Hunt and Census
-
2027
planned
World model v0.1
Beats mentor scores and LLM judges on day-30 revenue
-
2028
planned
World model v1 and FounderSim
Simulated cohort reproduces real 2028 effect sizes
-
2029
planned
Planning agents
Beat non-model agents on verified revenue
-
2029
planned
Control policy
Matches the best fixed oversight regime on value and safety
-
2030
planned
World-model status
5+ external stacks adopt FounderBench; 6+ peer-reviewed papers
Team
Lisbon · London · Rome · Zurich
- Principal investigator
- ML engineer, agents
- ML engineer, scientific ML
- Simulation engineer
- Data and ethics manager
- PhD students
Papers
Our outputs, and the field work we track.
Ours
- Founder Control Layer v1 and trace schemaTechnical report · expected Dec 2026planned
- Course Study 2027 pre-registrationRegistered report · expected Jun 2027planned
- Course and Accelerator working papersWorking paper · expected Nov 2027planned
Tracked
- Dell'Acqua et al., “The Cybernetic Teammate”, Organization Sciencepeer-reviewed
- Kim et al., “Generative AI Fuels Solo Entrepreneurship, but Teams Still Lead at the Top”, arXiv:2605.10291preprint
- Kim, Kim & Koning, “Mapping AI into Production”, INSEAD WP 2026/20/STRworking paper
- “Comparing Human Oversight Strategies for Computer-Use Agents”, arXiv:2604.04918preprint
- “Principles and Guidelines for Randomized Controlled Trials in AI Evaluation”, arXiv:2605.02050preprint
- He et al. (Collinear AI), “YC-Bench”, arXiv:2604.01212preprint
- Chen, Narasimhan & Liu, “CEO-Bench”, arXiv:2606.18543preprint
- Zhu et al., “StartupBench”, arXiv:2608.17800preprint
Verified 2026-09-14.
Funding
Amounts are sent with the first reply.
Compute
Give: Credits sized to a cohort or a season
Get: Benchmark runs, register entry, co-authorship
Hypothesis
Give: Sponsorship and tool access
Get: One pre-registered hypothesis a year, embargoed results
Licence
Give: Annual fee
Get: World model and FounderBench access
Consortium
Give: A Horizon, EIC or council slot
Get: Work packages, host universities, open outputs
Join a study
1 · AI builds, founder controls
Course In design
In design. First cohort July 2027. Join before March 2027 to shape it.
- Asks
- Can a founder-controlled agent system reach $10K revenue in 30 days — and which control setup keeps the founder in charge?
- Measures
- Verified revenue at day 30. Power d≈0.57 per 50-team arm.
- Runs
- Jul–Aug · 30 days · 350 teams, randomised
Thank you
Nothing else is needed from you now. If you want to add context, reply to our email.
Monthly update
One email a month: what we published, what we found, what is next.