Two Nature Papers Published the Same Day: How Robin and ERA Redefine the AI Scientist
On May 19, 2026, *Nature* simultaneously published two feature papers on AI scientists, a third on Google's Co-Scientist, and two cautionary editorials. Together the issue forms a deliberate hedge: showcasing frontier capability while reminding readers that AI remains a tool, not a replacement. The two flagship systems attack complementary bottlenecks of science—Robin handles the "thinking" (hypothesis generation), and ERA handles the "doing" (code and algorithms).
---
1. Robin: An "AI Research Group" That Reads Papers
Who built it
Robin comes from FutureHouse, a non-profit founded in 2023 and funded by former Google CEO Eric Schmidt. Its founders, Sam Rodriques and Andrew White, are physicists and chemical engineers. In late 2025, FutureHouse spun out Edison Scientific, raising $70M at a $250M valuation.
System architecture
Robin is a multi-agent pipeline built on three cooperating agents:
| Agent | Role | Core capability | |-------|------|-----------------| | Crow | Broad literature search | Answers "Has this question been studied?" | | Falcon | Deep literature synthesis | Maps what is known and unknown | | Finch | Data-driven discovery | Proposes hypotheses from experimental data |
Crow and Falcon handle "dry-lab" literature review; Finch handles "wet-lab" data analysis. The trio runs a closed loop: literature → hypothesis → experimental design → data → new hypothesis → new experiment.
Case study: A new drug candidate for dry AMD
The target was dry age-related macular degeneration (dAMD), a leading cause of blindness in developed countries with no approved therapy.
1. Literature scan — Crow and Falcon traced the pathology to impaired phagocytosis in retinal pigment epithelium (RPE) cells. 2. Hypothesis — Enhancing RPE phagocytosis could be therapeutic. 3. Drug screen — Among existing drugs, Robin identified ripasudil, an approved Rho-kinase (ROCK) inhibitor for glaucoma, never previously proposed for AMD. 4. Wet-lab validation — Human researchers ran Robin-designed assays (RNA-seq, flow cytometry). Phagocytosis rose 7.5×. 5. Mechanism analysis — Robin proposed follow-up RNA-seq, identifying ABCA1 upregulation as a potential new therapeutic target. 6. Manuscript — All hypotheses, experimental plans, analyses, and figures in the paper were generated by Robin.
Timeline: 2.5 months from concept to submission.
Why it matters
This is not "AI-assisted drug discovery." It is the first end-to-end demonstration of a closed scientific-discovery loop driven by AI. Humans only executed the physical wet-lab steps. FutureHouse notes the central tension: Robin can generate ten hypotheses per hour, while cell cultures still need weeks to grow.
---
2. ERA: An "AlphaGo" for Scientific Coding
Who built it
Google DeepMind + Google Research + Harvard SEAS, led by Michael Brenner (Harvard) and Shibl Mourad (DeepMind).
Core method: LLM + tree search
ERA's innovation is not prompt engineering but tree search—the same family of algorithms behind AlphaGo:
1. Input: a scoreable task (e.g., "forecast COVID-19 hospitalizations") and a quality metric (e.g., accuracy). 2. Gemini generates initial code. 3. Code runs in a sandbox; score is returned. 4. Tree search decides: exploit the current branch or explore a new one. 5. LLM edits the code—adding components, swapping algorithms, importing ideas from papers or textbooks. 6. Loop until scores plateau.
Crucially, ERA can ingest external research ideas and convert them into runnable code, not merely recombine existing snippets.
Validation across five domains
| Domain | Task | Result | Reference | |--------|------|--------|-----------| | Single-cell analysis | scRNA-seq batch integration | 40 new methods, every one beats the public leaderboard top human submission | Public leaderboard | | COVID-19 forecasting | Hospitalization prediction | 14 models outperform the official CDC ensemble and every single model | CDC ensemble | | Zebrafish neural prediction | Activity across 70,000 neurons | Beats all baselines; trains ~100× faster than video models | Current state of the art | | Geospatial analysis | Satellite-imagery reasoning | Expert-level, novel U-Net + Transformer hybrid | Human best | | Numerical integration | Complex integral evaluation | Expert-level | Math software | | Time-series forecasting | General-purpose | Best tree-search result on the GIFT-Eval leaderboard | Leaderboard |
The standout result is single-cell integration: 40 novel methods, each algorithmically innovative rather than mere tuning—one example is a three-tier hierarchical ensemble combining Holt-Winters, regional LSTM, and country-level LSTM. The COVID-19 result has acute public-health implications: AI-built models could outperform expert-designed forecasting during future crises.
Why it matters
ERA attacks the software bottleneck of modern science. The system is domain-general: the same architecture reached expert level across bioinformatics, epidemiology, neuroscience, geospatial, and mathematics—a universal scientific problem solver.
---
3. Side-by-Side: "Thinking Deep" vs. "Doing Right"
| Dimension | Robin | ERA | |-----------|-------|-----| | Institution | FutureHouse (non-profit) | DeepMind + Harvard (industry + academia) | | Goal | End-to-end scientific discovery | Expert-level scientific software | | Core method | Multi-agent (literature + data) | LLM + tree search (code + optimization) | | Human role | Execute physical experiments | Provide initial prompts, validate outputs | | Output | Hypotheses, experimental designs, manuscripts | Runnable code, algorithmic innovations | | Validation | In vitro wet lab | Public leaderboards, literature comparison | | Speed | 2.5 months incl. experimental wait | Hours to days | | Key novelty | Repurposing ripasudil for AMD | 40 scRNA-seq methods + 14 COVID models |
The two systems are complementary: Robin asks "why"; ERA asks "how." Wired together, they would form a complete AI research loop—Robin proposes a target such as ABCA1, ERA codes the ABCA1 expression analysis, Robin updates the hypothesis, ERA iterates the code.
---
4. The Third Paper and Two Editorials
Co-Scientist (Google) — Gemini-based multi-agent system for drug discovery. It identified in vitro-validated repurposing candidates for acute myeloid leukemia and, in a single demonstration, recovered an Imperial College antibiotic-resistance hypothesis that took a human team a decade to develop—ERA accomplished it in days.
Two editorials bookend the bundle:
1. Nature editorial — "Why AI cannot do good science without humans": AI can run experiments and analyze data, but posing the truly important questions still requires human intuition, values, and curiosity. 2. Messeri & Crockett — "The uncritical adoption of AI in science is alarming": AI can produce plausible-but-wrong conclusions; reviewers and editors need new guardrails.
Nature's stance: AI is a tool, not a substitute.
---
5. What Gets Replaced—and What Doesn't
Replaced in the short term
- Literature review — WikiCrow has auto-generated 15,616 gene Wikipedia entries at ~8 min each, with ~9% error rate (lower than human-written pages).
- Code implementation — ERA compresses months of coding into hours.
- Data analysis — Finch auto-selects statistics, plots figures, writes result descriptions.
- Hypothesis generation — Within known knowledge, AI systematically traverses possibilities faster.
- Asking the truly important questions — breakthrough science comes from questions outside existing frameworks.
- Value judgments — deciding which of 100 candidates warrants a $100M clinical trial.
- Tacit experimental knowledge — cell-line quirks that never make it into the literature.
- Error detection — the scientific community's skepticism is the last line of defense.
- Cross-disciplinary intuition — the substrate of breakthrough insight.
- FutureHouse / Edison Scientific: $70M raise, $250M valuation; Kosmos AI Scientist handles 1,500 papers and 42,000 lines of analysis code per run.
- Google: Co-Scientist and ERA sit inside a broader AI-for-Science stack alongside AlphaFold and Google Health.
- Eli Lilly + NVIDIA: $1B joint AI lab.
- Merck + Google Cloud: 10-year, up-to-$1B strategic partnership.
- AstraZeneca: acquisition of Modella AI for oncology R&D.
- Robin proves AI can "think" — it surfaced ripasudil for AMD, a non-obvious repurposing hit validated in vitro. That is creative hypothesis generation, not database search.
- ERA proves AI can "do" — across five unrelated domains it produced expert-level code, including 40 leaderboard-beating methods. That is algorithmic innovation, not hyperparameter tuning.
- Ghareeb, A.E. et al. (2026). A multi-agent system for automating scientific discovery. *Nature*. https://doi.org/10.1038/s41586-026-10652-y
- Aygün, E. et al. (2026). An AI system to help scientists write expert-level empirical software. *Nature*. https://doi.org/10.1038/s41586-026-10658-6
- Gottweis, J. et al. (2026). Accelerating scientific discovery with Co-Scientist. *Nature*. https://doi.org/10.1038/s41586-026-10644-y
- FutureHouse: https://www.futurehouse.org/
- Harvard SEAS: https://seas.harvard.edu/news/2026/05/ai-system-automates-coding-scientific-research
- Nature editorial (2026). Why AI cannot do good science without humans. *Nature* 653, 650.
- Messeri, L. & Crockett, M.J. (2026). The uncritical adoption of AI in science is alarming. *Nature* 653, 675–676.
Unlikely to be replaced
A new role emerges: the scientist as AI coach and referee—setting direction, vetting hypotheses, choosing which experiments to run, and integrating AI output into larger theory.
---
6. Industry Signal: The Commercial Race
AI-driven research has moved from academic curiosity to commercial arms race.
---
7. Bottom Line
Their shared limitations are equally clear: both still need human hands for physical experiments or initial prompts; both operate within existing scientific paradigms; neither has been stress-tested at scale over time.
The Nature editorials are correct: AI cannot do good science without humans. But humans cannot do good science at AI speed or scale either. The likely outcome is not *"AI replaces scientists"* but "scientists who use AI replace scientists who don't."
---