Science is changing independent of AI

Number of Researchers are Growing

International R&D spending

PhD Researchers

Intellectual bottlenecks are growing

📝 Increasing paper count ($\approx$15M per year)

🧬 Larger data sets from cheaper experiments (genome at $200 per person, $1 / GB of sequencing)

🔍95% decline in disruptive papers since 1980

AI Progress

Model intelligence doubles every 7 months

Effect is Visible in Economy

Labbench 2

1,912 questions across accessing data, reading patents, reading figures, protocol understanding, etc.

Deletion of which residues from C. elegans protein COSA-1 would most likely affect the ability of COSA-1 to recruit MSH5 and ZHP3?

Residues 31–40

Labbench 2 Accuracy

FutureHouse Timeline

Edison Scientific

• Spinout from FutureHouse formed in 2025
• AIxBio Research Lab
• 50 employees
• $70M in seed financing

Introduction to AI Models

What is an agent?

How are agents trained?

1. Pre-training — broad knowledge via next-token prediction
2. Supervised Fine-Tuning (SFT) — learn from expert instruction–response pairs
3. Reinforcement Learning (RL) — optimize through environment interaction with verifiable rewards

RL expands capabilities beyond what supervised data can teach

Verifiable Rewards

Computational checks (code execution, tool outputs) produce objective reward signals — no human feedback needed

LabBench: verifiable tasks as RL training signal

Learning vs Frontier Models

Training Curve

FutureHouse Agents

Automating research of scientific literature

Measurement: LAB-Bench (2024)

Approximately what percentage of Drosophila with a H3.3K36R mutation finish developing and enclose?

1. 80%
2. 19%
3. 50%
4. 37%
5. 6%
6. 94%

Better at answering questions than PhD biology experts

Improving over time

Better than human written Wikipedia articles

PaperQA3 (2026)

Agents for Data Analysis

Evaluation: Can it reproduce papers?

... Calculate Spearman correlations of the resulting log-fold change (logFC) values across conditions. Perform hierarchical clustering. Plot and visualize the clustering result as a heatmap to show how different ASD forms cluster together as development progresses

Creates a reproducible R notebook

Side-by-Side

Side-by-Side

BixBench

Scientific Reasoning Models

Improving Models

Reasoning scaling

Can we build scientific reasoning models?

chemistry reasoning model

Works with molecular structures, but reasons in English

Start from base LLM and teach it chemistry

What can a reasoning model do?

Q:Propose a 1-step synthesis path that uses only commercially available reagents

Q: Propose a modification to this molecule to increase its solubility by about 1 LogS unit without affecting its scaffold.

data

Training Stages

Can learn from zero accuracy

Results vs humans and frontier models

Predictions

Progress in Agents

• Since early 2022, agents have mostly been stagnant. ReAct is still a powerful baseline
• Call tools, keep conversation history, use special prompts, each agent is special
• Progressive disclosure (MCP, skills) has been main innovation
• Training, good infra, tools, and data has been main differentiators

Uniagents

• Only write and execute code
• No special tools: new integrations are mostly just code to write
• No MCP-style context bloat: discover code like an unknown codebase
• No special harness: complexity is reduced and training is easier

Demo Tax

• There is a growing gulf between demo and production
• Scaling low-latency sandbox execution is challenging
• Secrets management and arbitrary code execution create complex security

Beyond Stateless Chat

• OpenClaw showed a different form factor
• A colleague that works independently, not just an assistant
• Requires long-running reasoning