Can AI Scientists Do Independent Research Yet?
What end-to-end research automation can do today, where it still fails, and why verification remains a human responsibility.
Product updates and notes from the Mira team.
What end-to-end research automation can do today, where it still fails, and why verification remains a human responsibility.
Benchmark accuracy is not reproducibility. Scientific AI needs recorded methods, environments, parameters, and evidence.
AI can now generate candidates and help close experimental loops, but validation and reproducible execution remain the hard parts.
A context window is temporary working state. Research teams need project knowledge that can be saved, retrieved, and reused.
What this case study reveals about verification, reproducibility, and audit trails in AI-assisted research.
OpenAI says models including GPT-5.6 Sol escaped an internal evaluation sandbox and reached Hugging Face. What does the incident mean for long-running agents?
Try Kimi K3 on Mira with account credits. Review its 1M context, official benchmarks, API pricing, limits, and a real Deep Research test.
Try GLM-5.2 in your browser with Mira credits and no API key, plus benchmarks, cost vs GPT-5.5, and a first-hand test.
Materials research combines computable candidates, physical constraints, and measurable experiments into a practical AI workflow.
From sequence and SMILES to three ranked ESMFold2 complexes: inspect the exact inputs, confidence metrics, format handoff, interaction diagram, and validation limits.
See how Mira records a phenylethyl resorcinol molecular-orbital and electrostatic-potential task through conformer sampling, DFT, HOMO/LUMO, and ESP analysis.