How Replication Could Teach Machines What Good Science Looks Like — Edward Hughes

Summary

Edward Hughes, cofounder and chief scientist at Inherent, lays out a theory of scientific AI that centers on open-ended discovery, evaluation in hindsight, and collective intelligence rather than a single autonomous agent. He distinguishes innovation from creativity, argues that creativity depends on constraints, context, and field-level recognition, and frames replication as a creative act that reconstructs intent under incomplete information. The second half of the conversation turns to Replica, a benchmark built from redacted figures in real papers, and Faraday, a 27B Qwen model trained with GRPO-style RL and a frontier coding agent as a tool. Hughes says the system beats Codex, Claude, and GLM 5.2 on held-out replication tasks, while also surfacing practical issues like cheating, rubric design, and the limits of harness-heavy approaches. He closes by arguing that the real product is a recursive company: many humans and many agents improving one another across the scientific stack.

Notes

• Hughes frames the mission as building AI scientist systems that ask better questions, not just answer prompts.
• He says the next era is collective intelligence, not individual intelligence.
• Inherent has raised $50 million from Index Ventures and Radical Ventures at a $225 million post-money valuation.
• The team is small: 11 people.
• He says AI by itself is not the objective; the point is what AI enables in the wider scientific ecosystem.
• His background is theoretical physics, especially string theory scattering amplitudes.
• He used string theory dualities as a way to map hard calculations into easier geometric spaces.
• AlphaGo’s Move 37 was a major catalyst for his interest in discovery agents.
• He joined DeepMind in 2017 to pursue agents that could make discoveries.
• Multi-agent RL appealed to him because cultural evolution seems to outpace individuals.
• He describes cultural evolution as the fastest intelligence-generating process in the universe.
• His prior RL work included Adaptive Agents, a meta-RL system across a very large task space.
• Adaptive Agents was, at the time, a 500 million parameter trained-from-scratch RL agent.
• He led an AI Scientist team at DeepMind focused on applying coding agents to scientific research.
• Inherent is meant to be a horizontal intelligence layer for all of science.
• He argues the startup is needed because discovery requires interaction with the full scientific ecosystem, not just a single model.
• He says a weaker coding agent can use a stronger coding agent as a tool, and both weights and context change.
• He treats this as a mixed weights-and-context architecture, not a pure one or the other.
• The current stack has sticky abstractions, and AI scientist systems may be able to relax several constraints at once.
• He defines innovation as turning unknown unknowns into known knowns.
• Creativity, in his view, requires a metacognitive step: recognizing that something is creative.
• He says Move 37 was recognized by commentators, not by AlphaGo itself.
• Csikszentmihalyi’s creativity model has individual, domain, and field components.
• He argues generalizable discovery engines could transfer insights across domains, such as from weather modeling to protein design.
• He describes creativity as satisficing, not optimizing.
• You must break some constraints, but not all of them, or the result becomes white noise.
• Creativity comes from relaxing constraints on interpretation and tools so new observations become possible.
• He gives exaptation examples from penicillin, microwaves, and GPUs.
• He uses Wagner’s Tristan chord to show that creativity can come from context rather than novelty of the raw object.
• Human communication is high-noise and narrow-bandwidth, and that under-specification can foster creativity.
• He says science is a search for good explanations, and good explanations are hard to vary.
• He argues copying is creative because transmission requires reconstructing another person's idea.
• He says the 10,000-to-4,000-years-ago creativity burst reflects copying and field changes more than individual biology.
• Open weights models make model-weight copying easy, which he thinks can reduce creativity by removing constraints.
• He casts humans as constraint engineers who shape interfaces between systems to promote novelty.
• He says frontier scientific-AI capability may come first on culturally advanced problems.
• He thinks the frontier in AI science is evaluation, not just model design.
• For scientist agents, he wants hindsight evaluation rather than foresight reward functions.
• He says open-endedness cannot have one single global goal, though local goals are allowed.
• He links this to curiosity-driven search in under-specified problems.
• Scientific discovery requires deceptive goals because a perfect world model would only confirm expectations.
• He defines a world model as an action-conditioned forward model of the world.
• Open-ended systems should produce outputs that are both novel and learnable by humans.
• He says human legibility matters; discoveries that cannot be parsed by humans have limited value.
• Replica is a paper-replication task space built around redacted figures and reconstruction from text.
• Replication is framed as a public good and a source of hidden tacit knowledge.
• The benchmark redacts figures from PDFs and asks the agent to recreate them from the paper description, without accessing the original figure.
• Judging uses another frontier coding agent plus a rubric, and they validate the judge against human taste.
• Faraday is a Qwen 3.6 27B model post-trained on many Replica rollouts with a frontier coding agent as a tool.
• Replica training uses 242 tasks and 68 held-out tasks.
• The held-out tasks are AI-for-science papers, while training includes classic ML papers and older open-endedness work.
• Faraday outperforms the frontier tool model and baselines like Claude and GLM 5.2 on held-out replication tasks.
• He says prompt optimization on GPT 5.5 Codex improves only a tiny amount versus Faraday’s larger advantage.
• Scaling Replica across ICML, ICLR, and NeurIPS could expand the task set by about two orders of magnitude.
• Some papers have no sensible scale-down version, so the team also tests larger-resource replications out of distribution.
• They also evaluate with larger budgets like eight B300 GPUs and eight hours for selected papers.
• Replica uses Gemini to identify plot figures, redact them with a UNIX utility, and create gold figures for judging.
• Most papers yield one or two usable figures, though some yield up to 13.
• Per-task rubrics worked well for scoring scientific rigor, visual fidelity, and claim reproduction.
• GRPO training initially hit an RL crisis because the tasks are non-verifiable and the judge is noisy.
• The 27B model can use Unix utilities, Codex, internet access, and downloads inside long multi-turn rollouts.
• They stabilized training by scoring each rollout multiple times with the same judge and averaging.
• They also added per-turn credit assignment so the judge assigns weight to each step of the rollout.
• He says the model is especially learning to preserve the scientifically load-bearing turns.
• They deliberately avoid constructing harnesses in this work.
• A cofounder analysis found harness-based capabilities are often matched by base models about three months later.
• Harnesses may be more sample-efficient, so the choice depends on the tradeoff you want.
• A next step is to use harnesses during training and distill the gains back into model weights.
• Recursive self-improvement is treated as a company-level phenomenon rather than a single isolated agent.
• The company itself should improve through interactions between humans and agents.
• He thinks most current agent use is too one-to-one and underuses many-to-many interaction.
• He analogizes the organizational shift to electrifying factories with dynamos instead of steam turbines.
• He says Inherent wants to reinvent the AI research factory from the ground up with agents at the center.
• He notes that OKRs work for Google-style organizations but clash with open-endedness if goals are too rigid or too long-term.
• He sees an emerging market for AI-assisted, AI-accelerated R&D because ideas are getting harder to find.
• He argues researcher productivity, drug-development cost, and Nobel-age trends all point in the wrong direction.