Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research
A British AI lab founded by former DeepMind researchers has unveiled an autonomous agent capable of replicating complex scientific research with higher fidelity than leading models from Anthropic and OpenAI. The system, named Faraday, successfully reproduced the results of published machine learning papers — including code, experiments, and analysis — without human intervention, marking a potential inflection point in the automation of scientific discovery.
The US government’s Anthropic models ban was never about… Exhibit at TechCrunch Disrupt 2026: Get in front of… Railway secures $100 million to challenge AWS with AI-native… Apple accuses OpenAI of destroying evidence
Background
Inherent, a London-based startup operating in stealth until now, was formed by alumni of DeepMind’s research and engineering teams. The company’s founding thesis centers on a bottleneck in modern AI development: while foundation models excel at pattern recognition and generation, they struggle with the rigorous, multi-step reasoning required to validate and extend scientific work.
Faraday is designed as an "AI teammate" rather than a chatbot or coding assistant. It operates by ingesting a research paper — typically a PDF with supplementary materials — then autonomously reconstructing the computational environment, writing execution code, running experiments, analyzing outputs, and comparing results against the original claims. The agent iterates on failures, debugs environment mismatches, and adjusts hyperparameters until replication succeeds or a definitive limitation is documented.
In benchmark testing, Inherent evaluated Faraday against Claude 3.5 Sonnet and GPT-4o using a curated set of 50 recent machine learning papers from top venues including NeurIPS, ICML, and ICLR. The test suite covered diverse architectures — transformers, diffusion models, reinforcement learning agents — and required handling of GPU workflows, distributed training, and custom CUDA kernels. Faraday achieved a 78% full replication rate, compared to 42% for Claude and 35% for GPT-4o. Partial replication — where core results were reproduced but secondary claims failed — reached 91% for Faraday versus 63% and 58% respectively.
The system does not rely on a single large language model. Instead, it orchestrates a pipeline of specialized sub-agents: a planner that decomposes the paper into executable tasks, a coder that writes and refactors implementation, an executor that manages compute environments, and a verifier that statistically compares outputs. This modular design allows Inherent to swap components as better models emerge, while maintaining a consistent replication workflow.
Why It Matters
Replication is the bedrock of scientific progress, yet it remains chronically underfunded and underperformed. A 2016 Nature survey found that over 70% of researchers had tried and failed to reproduce another scientist’s experiments, and more than half had failed to reproduce their own. In AI, the problem is acute: rapid publication cycles, complex dependencies, and compute-intensive workflows make replication prohibitively expensive for most labs.
Faraday changes the economics. What once required a PhD student weeks of effort — tracking down dependencies, reverse-engineering undocumented preprocessing, debugging silent failures — can now be attempted in hours, at cloud compute cost. For companies building on published research, this means faster validation of vendor claims, quicker integration of academic advances, and reduced risk of deploying unverified methods.
The implications extend beyond AI. The same replication pipeline — ingest paper, reconstruct environment, execute, verify — applies to computational biology, climate modeling, materials science, and any field where results depend on code and data. Inherent has signaled intent to expand Faraday’s scope, but its near-term focus remains on ML research, where the evaluation infrastructure is most mature.
For AI adopters in enterprise, the signal is clear: the gap between "model reads paper" and "model reproduces result" is narrowing. Organizations that treat published research as a procurement channel — not just inspiration — gain a competitive edge. Faraday automates the due diligence that technical teams currently perform manually, or skip entirely.
What's Next
Inherent has not announced pricing, availability, or a public roadmap. The company says it is working with a small group of design partners — academic labs and industrial research groups — to stress-test Faraday on proprietary and pre-publication work. Key challenges remain: handling closed-source dependencies, navigating licensing restrictions on datasets, and scaling the verification layer to handle non-deterministic outputs like generative model samples.
There is also the question of trust. An AI that replicates a paper perfectly may still inherit the original authors’ errors, biases, or flawed assumptions. Faraday flags statistical anomalies and missing controls, but it does not yet critique experimental design. Inherent acknowledges this limitation and positions the tool as an accelerator, not a replacement, for human judgment.
Competitors are unlikely to stay idle. Both Anthropic and OpenAI have published work on tool-using agents for scientific tasks, and open-source projects like AutoGPT and MetaGPT have demonstrated multi-agent research workflows. But Inherent’s domain-specific focus — replication as a first-class task, not a demo — gives it a structural advantage in reliability and evaluation.
Takeaway for Business Owners
If your organization builds products on published AI research, you now have a tool that can validate those foundations before you invest engineering months. Faraday won’t replace your ML team, but it can eliminate the "trust but verify" gap that slows adoption of academic advances. Watch this space: the first companies to systematize research replication will move faster, fail less, and ship more reliable AI.
Source: Original Article