💡Faraday AI Agent Outperforms Giants on Paper Replication
Tiny Model Beats Big Players in Research Replication
TL;DR
London-based Inherent's AI agent, Faraday, outperformed larger models from Anthropic and OpenAI on scientific paper replication tasks using just 27 billion parameters. This could redefine research efficiency.
Inherent, a London startup founded by ex-DeepMind alumni, unveiled its AI agent Faraday, which outperforms much larger models in independently reproducing the findings of published scientific papers. The kicker? Faraday does this with just 27 billion parameters, compared to competitors' hundreds of billions. This matters for researchers and developers looking to automate complex tasks without bloated models. Inherent's model uses reinforcement learning to develop 'research taste,' identifying valuable experiments autonomously. With a $50 million seed round under its belt, the company is set to grow from 12 to around 20-25 employees by year-end.
Key Points
Faraday, Inherent's AI agent, uses reinforcement learning to replicate scientific findings without being told the answer in advance.
Inherent raised $50 million in a seed round, backing Faraday's development and future growth plans.
The model Qwen 3.6 has only 27 billion parameters, significantly smaller than competitors like Anthropic and OpenAI.
Faraday aims to demonstrate 'research taste,' identifying valuable experiments autonomously without explicit instructions.
Inherent is expanding its team from 12 to around 20-25 employees by the end of this year.
Why It Matters
If you're automating complex research tasks, Faraday's approach could save time and resources. Its reinforcement learning method for identifying valuable experiments without explicit instructions can streamline workflows. However, its reliance on OpenAI’s GPT-5.5 Codex means it leans heavily on existing software, limiting its standalone capabilities.
Frequently Asked Questions
Why does this matter?
If you're automating complex research tasks, Faraday's approach could save time and resources. Its reinforcement learning method for identifying valuable experiments without explicit instructions can streamline workflows. However, its reliance on OpenAI’s GPT-5.5 Codex means it leans heavily on existing software, limiting its standalone capabilities.
What happened?
London-based Inherent's AI agent, Faraday, outperformed larger models from Anthropic and OpenAI on scientific paper replication tasks using just 27 billion parameters. This could redefine research efficiency.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,305 builders reading daily.