DeepMind Alumni Build Faraday AI to Automate and Verify Scientific Research

British artificial intelligence startup Inherent, founded by Google DeepMind alumni, launched Faraday—an autonomous AI lab agent designed to systematically replicate complex scientific research. Benchmarks reveal Faraday outperforms frontier models from OpenAI and Anthropic in automated paper replication, signaling a major shift toward automated scientific discovery and verification.

Automating the Gold Standard of Scientific Truth

When Demis Hassabis and his team built Google DeepMind, their overarching mission was solving intelligence to advance science. Now, a new cohort of DeepMind alumni operating under the banner of Inherent is taking that vision a step further by automating the foundation of scientific discovery itself. The London-based artificial intelligence lab released Faraday, an autonomous AI lab agent engineered specifically to read, execute, and verify scientific research papers at scale.

The global scientific community faces a persistent, multi-billion-dollar bottleneck known as the replication crisis. Thousands of published papers across biotechnology, materials science, and machine learning contain subtle code errors, omitted hyperparameter values, or undocumented dependency setups that prevent other laboratories from reproducing their results. Inherent's Faraday directly addresses this systemic weakness by acting as an autonomous software engineer and empirical reviewer.

In head-to-head performance evaluations, Faraday successfully parsed technical research manuscripts, set up complex computational execution environments, repaired broken code repositories, and generated matching empirical results at a significantly higher success rate than Anthropic's flagship Claude models and OpenAI's latest reasoning systems. While general-purpose Large Language Models (LLMs) frequently hallucinate library dependencies or fail during multi-step command-line sequences, Faraday maintains continuous task execution state across long-horizon scientific workflows.

This performance gap highlights a crucial transition in autonomous AI deployment. Rather than providing conversational summaries of technical PDF documents, Faraday operates directly inside sandboxed virtual machinery—executing terminal commands, troubleshooting missing software drivers, and evaluating output data against published figures without human intervention.

Inside Faraday: How DeepMind Alumni Outpaced Silicon Valley Giants

The technical architecture powering Faraday reflects Inherent’s deep roots in reinforcement learning and high-performance scientific computing—disciplines mastered by its founders during their time at Google DeepMind. Traditional generative models treat scientific papers as pure text completion challenges. In contrast, Faraday approaches academic manuscripts as executable software blueprints.

When presented with a complex scientific paper, Faraday executes a structured, four-phase verification pipeline:

• Methodology Ingestion: The agent parses raw PDF documents, extracting mathematical formulas, algorithm logic, hyperparameter configurations, and explicit hardware requirements.
• Environment Synthesis: Faraday constructs an isolated virtual container, automatically installing exact CUDA versions, Python packages, and legacy compiler tools specified or implied by the authors.
• Iterative Debugging: If execution crashes due to missing data paths or deprecated library functions, the agent analyzes system stack traces, writes code patches, and re-executes scripts in real time.
• Empirical Verification: The agent compares generated charts, loss curves, and benchmark metrics against original figures in the published paper, outputting a detailed reproducibility score.

By prioritizing terminal execution accuracy and mathematical rigor over general chatbot conversation, Inherent created a specialized architecture that broader Silicon Valley tech giants overlooked during their pursuit of general-purpose models.

Democratizing R&D from London to Emerging Tech Hubs

The capability to instantly verify academic literature carries immediate utility for commercial research departments, pharmaceutical firms, and university laboratories worldwide. Replicating a single state-of-the-art machine learning model or drug discovery benchmark typically requires weeks of intensive effort by senior post-doctoral researchers or software engineers. Faraday reduces this effort from hundreds of billable hours down to minutes.

For research institutes and engineering firms operating across South Asia and the Gulf—where access to massive compute infrastructure and dedicated venture capital has historically trailed Silicon Valley—Faraday serves as an immediate force multiplier. Engineering teams in Lahore, Riyadh, or Dubai can now validate and build upon cutting-edge research published in London or San Francisco without wasting engineering hours on broken codebases.

Furthermore, early-stage venture capital firms and institutional investors can utilize Faraday to conduct automated technical due diligence. Instead of accepting pitch deck metrics or unverified benchmark claims at face value, investment teams can run published codebase claims through Faraday to confirm whether a tech startup's underlying technology performs as advertised.

By converting static scientific literature into dynamic, executable code, Inherent is altering the infrastructure of human knowledge generation. As Faraday moves into commercial deployment, the delay between reading a breakthrough paper and deploying its technology has effectively collapsed.

Frequently Asked Questions

What is Inherent's Faraday and how does it function?

Faraday is an autonomous AI agent developed by British startup Inherent to replicate and verify scientific research papers. It ingests PDF manuscripts, builds code execution environments, debugs code errors, and compares generated results against published benchmarks.

How does Faraday compare to models from OpenAI and Anthropic?

While OpenAI and Anthropic focus on general-purpose conversational models, Faraday is purpose-built for scientific computing and software execution. Benchmark testing showed Faraday outperformed both competitors in automated code replication and scientific error correction.

Why is automated paper replication useful for research institutions?

Automating paper replication resolves the scientific reproducibility crisis by verifying published experimental results in minutes rather than weeks. This allows research institutes and technology companies to validate algorithms without wasting engineering time on broken codebases.