On September 23, 2026, Anthropic published something more interesting than another benchmark.
Its Claude agents had searched more than 200,000 reverse transcriptases, generated thousands of candidate systems, narrowed them to a small set of compelling candidates, and eventually surfaced something Anthropic calls array-associated reverse transcriptases (ART).
The search involved roughly 950 agents, 21 hours, and 210 million tokens.
But the most important part of the story is not the number of agents.
It is what one of them noticed.
A strange pattern.
A repeat array sitting next to an unusual reverse transcriptase.
Something that did not simply match an existing annotation.
Something that looked wrong enough to investigate.
And that leads to a much bigger question:
What if the next generation of AI for science is not primarily about predicting what is known, but discovering what does not fit?
From prediction to anomaly
Most AI-for-science systems are optimized around questions such as:
- What is the function of this protein?
- Which molecule is likely to bind this target?
- Which structure is most probable?
- Which candidate has the highest predicted activity?
These are powerful problems.
But scientific discovery often begins somewhere else.
It begins with:
“Why is this here?”
CRISPR itself is a beautiful example of this pattern.
An unusual repeat structure was observed in bacterial genomes long before its full biological significance became clear.
The history of science contains countless examples where the signal was not a known category.
The signal was an anomaly.
And anomalies are exactly where conventional classification systems can struggle.
The real opportunity: Anomaly-to-Mechanism
Imagine a system whose primary objective is not:
Predict the label.
Instead:
Find the thing that does not make sense.
Then force the system to explain why.
I call this architecture:
Anomaly-to-Mechanism Engine (A2M)
Its workflow would look something like:
Massive Scientific Data
↓
Pattern Detection
↓
Anomaly Discovery
↓
Hypothesis Generation
↓
Alternative Explanations
↓
Scientific Red Team
↓
Information-Rich Experiment
↓
Human Validation
↓
Evidence Update
↓
Reproducible Discovery
The key difference is philosophical.
The AI is not rewarded for producing the most convincing explanation.
It is rewarded for producing a hypothesis that survives attempts to destroy it.
The Scientific Skeptic
This is where today's AI scientist architectures could become much more interesting.
For every promising hypothesis, create adversarial agents:
The Advocate
Why could this be important?
The Prosecutor
What evidence suggests this is an artifact?
The Archaeologist
Has somebody already discovered this under another name?
The Statistician
How surprising is this pattern under an appropriate null model?
The Experimental Skeptic
What is the cheapest experiment that could prove this hypothesis wrong?
Now the system has a completely different objective.
Not:
“Make the hypothesis sound intelligent.”
But:
“Try to kill the hypothesis.”
If it survives, increase its priority.
If it fails, archive it.
And preserve the reason.
Discovery should have a data structure
One of the biggest problems with agentic science is that conversations disappear.
A model searches.
It reasons.
It generates hypotheses.
Another model critiques them.
Someone runs an experiment.
Then six months later:
Why did we believe this in the first place?
We need something more structured.
I call it a:
Discovery Card
Every candidate scientific discovery should carry:
Discovery ID
Observation
↓
Detected anomaly
↓
Primary hypothesis
↓
Alternative hypotheses
↓
Supporting evidence
↓
Contradicting evidence
↓
Literature status
↓
Uncertainty
↓
Falsification conditions
↓
Proposed validation
↓
Experimental result
↓
Updated belief
↓
Complete provenance
Now a discovery is not merely a paragraph generated by an LLM.
It becomes a versioned scientific object.
HypothesisOps: GitHub Actions for Science
This leads to another idea.
Software engineering has Git.
Machine learning has MLOps.
Scientific discovery needs something analogous.
Call it:
HypothesisOps
Every hypothesis gets:
- a unique identifier
- version history
- evidence
- counter-evidence
- model provenance
- dataset provenance
- agent traces
- competing hypotheses
- experimental outcomes
- human decisions
- confidence updates
Imagine opening a scientific hypothesis six months later and seeing:
HYP-004821
Created: 2026-09-23
Evidence:
████████░░ 78%
Alternative explanations:
3
Falsification tests:
2
Human validation:
Pending
Evidence added:
+4
Evidence rejected:
-2
Model provenance:
17 agents
Data provenance:
Genome datasets A/B/C
Current status:
UNDER REVIEW
This is not another chatbot.
It is an epistemic infrastructure layer for scientific discovery.
The Missing Metric: Information Gain
There is another problem.
Suppose an AI generates 1,000 hypotheses.
Which one should humans investigate?
The obvious answer is not necessarily:
The hypothesis with the highest probability.
Instead:
Which experiment would teach us the most?
This changes the optimization target.
Imagine four competing hypotheses:
H1 → defense mechanism
H2 → replication mechanism
H3 → regulatory mechanism
H4 → evolutionary artifact
The best next experiment is not necessarily the experiment that confirms H1.
It may be the experiment that most efficiently separates H1, H2, H3, and H4.
That gives us another optimization objective:
Maximum Information Gain per Unit Cost
The AI scientist therefore becomes less like a prediction engine and more like an experimental strategist.
What Anthropic's ART discovery really demonstrates
There is an important distinction here.
Anthropic says the underlying reverse transcriptase had already been identified in previous work.
The interesting observation was the larger architecture around it: the associated repeat array and accessory protein.
And Anthropic explicitly says that the function of ART is not yet known. Human scientists performed the laboratory work and are continuing to investigate it.
That distinction matters.
The achievement is not:
“AI solved biology.”
It is closer to:
AI searched an enormous biological space and surfaced an unusual structure that humans considered worth investigating.
That is arguably more interesting.
Because it suggests that AI may be able to expand the search frontier of science.
The Dark Biology Atlas
Now scale this idea.
There are enormous numbers of biological sequences whose functions remain poorly characterized.
Instead of simply creating a database of unknown proteins, imagine a:
Dark Biology Atlas
For every unknown biological system:
Evolutionary conservation
Genomic neighborhood
Structural similarity
Repeat architecture
Ecological distribution
Potential interactions
Literature coverage
Novelty
Alternative explanations
False-positive risk
Experimental tractability
But there is one crucial difference:
The system doesn't ask only:
“What don't we know?”
It asks:
“Which unknowns are unusually informative if we investigate them?”
That creates a map of scientific opportunity.
Can AI Learn Scientific Taste?
Anthropic describes an especially interesting part of its workflow.
Because Claude generates huge numbers of hypotheses, the researchers began studying the hypotheses themselves:
Why are some worth testing while others are discarded?
The answers can then influence future agent instructions.
This suggests a fascinating research question:
Can Scientific Taste Be Learned?
Maybe scientific intuition is not magic.
Maybe part of it can be decomposed into measurable properties:
- unexpectedness
- internal coherence
- independent evidence
- falsifiability
- experimental tractability
- distance from existing knowledge
- potential explanatory power
- information gain per experiment
That could become a new form of pre-evidence hypothesis selection.
Not:
“Is this hypothesis true?”
But:
“Is this hypothesis worth spending scientific attention on?”
AURORA
Put all of this together and you get something bigger.
I call the architecture:
AURORA
Autonomous, Uncertainty-Resolved, Open, Reproducible Agentic Science
Its mission:
Turn anomalies into falsifiable discoveries.
AURORA could contain:
SCOUT
Find unusual patterns
GENOMIC GRAMMARIAN
Understand structural context
TAXONOMIST
Map known and unknown families
ARCHAEOLOGIST
Search scientific history
SKEPTIC
Attack hypotheses
STATISTICIAN
Measure surprise and false discovery
EXPERIMENT PLANNER
Maximize information gain
SAFETY GOVERNOR
Control sensitive outputs
PROVENANCE AUDITOR
Track everything
HYPOTHESISOPS
Version the entire epistemic state
And the output isn't:
“Here is the answer.”
It is:
“Here is something strange. Here is why it might matter. Here are the competing explanations. Here is why we might be wrong. Here is the evidence. And here is the smallest next step that can teach us something.”
This is bigger than biology
The architecture doesn't fundamentally depend on DNA.
Replace the input layer and the same system could search for anomalies in:
Materials science
Unexpected crystal behavior.
Physics
Measurements inconsistent with current models.
Neuroscience
Unexpected relationships between neural dynamics and behavior.
Climate science
Patterns that violate existing models.
AI research
Model behaviors that existing theories cannot explain.
Medicine
Patient-level patterns that don't fit established phenotypes.
The domain changes.
The epistemic engine remains.
The real AI Scientist
Maybe we have been asking the wrong question.
The question isn't:
“Can AI become a scientist?”
A better question is:
“Can we build machines that systematically increase the rate at which humans encounter high-value, falsifiable scientific anomalies?”
That is a much more interesting engineering problem.
Because science doesn't advance only when someone knows the answer.
It advances when someone notices that the existing answer is insufficient.
From AI that answers to AI that notices
The next generation of scientific AI may therefore have a different architecture.
Not:
Data → Prediction
But:
Data → Anomaly → Hypothesis → Attack → Experiment → Evidence → Update → Discovery
And perhaps the most valuable capability of an AI scientist will not be its ability to say:
“I know.”
It will be its ability to say:
“I don't know what this is—but this pattern is too strange to ignore.”
That is where scientific discovery begins.
The opportunity is not to build an AI that replaces the scientist.
It is to build an epistemic operating system that makes humanity better at noticing, challenging, testing, and preserving the things we would otherwise miss.
HypothesisOps
From raw data to falsifiable discovery, with provenance.
created by Seyed Alireza Alhosseini Almodarresieh
Top comments (0)