Why Even 10× Smarter AI Scientists Won’t Accelerate Science: The 300‑Year‑Old Paper Bottleneck

In 2026 the AI‑scientist track is crowded with agents that can read literature, generate hypotheses, run experiments, and write papers, yet the author asks how much faster science would move if such an agent were as clever as a top human researcher. The answer, drawn from the blog "The Second Half of AI for Science," is that progress would not accelerate significantly.

Legacy Bottleneck: The Paper

The fundamental obstacle is a three‑century‑old protocol – the scholarly paper. Since the 1665 launch of the Philosophical Transactions , research has been forced into a linear, narrative format designed for human readers. This imposes two "taxes": a narrative tax that discards the messy, branching, and failed parts of discovery, and an engineering tax that omits the fine‑grained technical specifications needed for reproducibility. Peer review and the academic incentive system further reinforce these constraints.

First Half: Incremental Buffs

Recent AI‑for‑Science work has repeatedly added scaffolding, memory, multi‑agent orchestration, and self‑evolution loops, achieving modest benchmark gains and flashy demos. However, this approach hits three walls: (1) the paper format’s loss of information, (2) the labor‑intensive peer‑review process that still relies on human validation of results that could be automatically re‑executed, and (3) incentive mechanisms that reward attention‑economy metrics rather than genuine scientific advancement. An anecdote about a protein‑design team illustrates the "Bitter Lesson": a new LLM instantly superseded months of handcrafted pipelines, showing that effort spent on temporary scaffolding is quickly erased by model upgrades.

Proposed Paradigm Shift: Agent‑Native Research Artifact (ARA)

The accompanying paper, titled The Last Human‑Written Paper (arXiv:2604.24658), introduces the Agent‑Native Research Artifact (ARA) – a complete computational entity that bundles scientific logic, executable code, evidence linking each claim to its original output, and a full exploration graph that retains failed branches. The ARA repository is available at https://github.com/ARA-Labs/Agent-Native-Research-Artifact.

Empirical Benefits

When research is delivered as an ARA instead of a PDF, AI agents answering 450 questions improve their question‑answer accuracy from 72.4 % to 93.7 %, a gain of over twenty points attributed solely to the format change. End‑to‑end reproducibility success rises from 57.4 % to 64.4 %, reflecting both format advantages and model capabilities. Crucially, preserving failure trajectories enables subsequent AI scientists to avoid dead‑ends, accelerating future exploration.

Second Half: Rebuilding the Scientific Network

The author argues that true acceleration requires fixing the communication protocol rather than making individual agents faster. By making research forkable like open‑source code, science gains version control, dependency graphs, and "git blame" for knowledge. Humans shift from direct experimenters to goal‑setters, compute‑budget allocators, and safety overseers, especially in high‑risk domains such as synthetic biology where alignment must become hard‑engineered.

Vision

In the envisioned future, a network of AI scientists continuously generates executable research artifacts that can be forked, combined, stress‑tested, and re‑executed within hours. The literature becomes a living, executable tree rather than isolated PDFs, with humans navigating the canopy, pruning, and steering the growth.