Thesis

What do they actually do

Thesis is an applied research lab building AI systems that automate parts of the machine‑learning research workflow (e.g., planning experiments, training/evaluating models, and iterating on designs) to speed up discovery. Public materials emphasize agentic ML R&D capabilities and report a state‑of‑the‑art result on OpenAI’s MLE‑Bench, achieved quickly and with relatively small compute, as an early proof point (YC profile, website, MLE‑Bench, OpenAI post).

They describe a longer‑term plan to generalize these tools beyond ML into areas like computational biology and wet‑lab hypothesis prioritization, but current public artifacts focus on autonomous ML research workflows and benchmarks (research page).

Who are their target customer(s)

• Academic principal investigators in ML, computational biology, or chemistry: Need publishable results and grant wins but face slow experiment cycles and limited lab/compute budgets; seek faster, lower‑cost ways to find promising ideas or models (YC profile).
• Small biotech or pharma R&D teams: Must choose which molecular or experimental hypotheses to run when wet‑lab validation is expensive and slow; want better in‑silico prioritization to reduce wasted experiments (YC profile).
• ML research engineers at startups: Trial‑and‑error model design and long training iterations drive up compute costs; want automation to explore designs and hyperparameters faster, guided by benchmarks like MLE‑Bench (YC profile, MLE‑Bench paper).
• Independent researchers and deep‑tech startups: Lack large R&D teams and need tooling to run systematic experiments and track results without heavy headcount (YC profile).
• Corporate research groups at large labs: Internal productization and process constraints make it hard to explore risky ideas; want a rapid experimentation sandbox to de‑risk early research (YC profile).

How would they acquire their first 10, 50, and 100 customers

• First 10: Leverage founders’ network and YC introductions to run tightly scoped, outcome‑driven paid pilots with a few PIs, independent researchers, and 1–2 small biotechs; trade discounts for co‑authorship or shared IP and cite early MLE‑Bench results as proof (YC profile, research page, MLE‑Bench paper).
• First 50: Systematize outreach with conference demos/workshops and publish step‑by‑step case studies showing time/compute savings; offer a grant‑aligned pilot for PIs and recruit via lab mailing lists, GitHub, and conference attendees, using MLE‑Bench credibility to open doors (MLE‑Bench paper, YC profile).
• First 100: Productize repeatable pilot templates (e.g., model search, molecule prioritization, experiment triage) into a semi‑self‑serve offering; add a technical sales lead and partner with university core facilities/CROs to extend reach, anchored by published pilots and ROI stories (research page).

What is the rough total addressable market

Top-down context:

Broadly adding adjacent markets yields a top‑down pool of roughly $35–45B today: AutoML + AI developer tools (~$3–10B), bioinformatics (~$25.8B), lab automation (~$8.3B), and AI in drug discovery (~$1.9–3.6B) (AutoML, AI code tools, bioinformatics, lab automation, AI drug discovery, GMI).

Bottom-up calculation:

Near‑term, the serviceable market aligns with AutoML/AI research tools. As a working envelope, 1,000–3,000 likely early adopters (labs, startups, and R&D teams) paying $20k–$100k ARR implies roughly $20M–$300M of reachable spend over 1–3 years, contingent on validation and sales capacity.

Assumptions:

• Focus on ML researchers first; biology/wet‑lab features come later.
• Academic adoption will be slower and lower‑ACV than industry; mix skews toward startups/biotechs for revenue.
• Pricing resembles other research tooling (pilot/seat/usage tiers), with ACV concentrated in higher‑touch programs.

Who are some of their notable competitors

• Atomwise: AI‑powered small‑molecule discovery and virtual screening services; overlaps if Thesis prioritizes molecules/targets before wet‑lab validation (overview).
• Synthace: Cloud platform (Antha) to design, schedule, and automate complex biological experiments; overlaps on experiment planning/automation for wet‑lab work (PLOS blog).
• Benchling: Cloud R&D platform (ELN/LIMS/workflows) used by labs and biotechs; competes for the same institutional buyers organizing experimentation workflows (product overview).
• Weights & Biases (W&B): ML experiment tracking, visualization, and sweeps; overlaps with Thesis on automating and managing ML research loops (docs).
• BenchSci: Uses AI to surface relevant reagents and evidence to pick better experiments; overlaps with Thesis’ goal of reducing wasted wet‑lab runs.