DeepMind Alumni’s Startup Inherent Says Its ‘AI Teammate’ Beats Frontier Models at Reproducing Research

Inherent, a new startup founded by former DeepMind researchers, claims its “AI teammate” outperforms frontier models from OpenAI and Anthropic on a demanding task: reproducing scientific research. In this setting, the system reads papers and experimental write‑ups, then autonomously writes code, processes data and checks outputs to see whether it can match the original results.

According to reports summarized in recent AI news digests, Inherent’s agent‑style system breaks complex experimental protocols into smaller steps, calls external tools as needed, and iterates when runs fail. Two advertised strengths are its ability to search for additional papers when the original description is vague, and to reason over logs to adjust parameters without human guidance.

So far, however, performance has only been demonstrated on internal benchmarks designed by the company. Outside researchers have not yet independently validated the claims. If systems like this do scale, they could dramatically speed up literature review and experiment design across fields such as biology and materials science. At the same time, the idea of AI rapidly “rubber‑stamping” flawed methods raises concerns, making transparent, community‑agreed evaluation standards essential before such tools are widely relied upon in science.