Blog ·
Science and Research: AI trends to focus on - AI agents begin running experiments
AI is shifting from answering questions to running lab experiments. New cheaper models (GPT-6.1 Sol, Sonnet 5.5, Gemini 4 Argon) and tools that log every action make reproducibility essential. Choose models based on cost and logging, not just capability.
Jeroen Erne
What changed this week
Scientific AI moved from answering questions to running experiments. Three threads converged: robot planning that uses simulator feedback to improve trajectories (SAIL), open-source connectors that let AI control lab instruments directly (K-Dense), and local research assistants that produce auditable lab workflows. The shift is operational — AI is now touching pipettes, not just PDFs.
On the model side, the week saw a flurry of releases tuned for research throughput. Anthropic shipped Sonnet 5.5 as a cheaper, faster work model. OpenAI launched GPT-6.1 Sol, claiming near-Astra performance at lower cost, while also fixing an image-encoding bug in GPT-6 Sol and Luna. Google released Gemini 4 Argon for complex long-horizon workflows, and Ant Group introduced Ling-3.1-flash with a one-million-token context window. The pattern is clear: frontier capability is being packaged into smaller, faster, cheaper models that researchers can run at scale.
Infrastructure for reproducibility gained ground. alphaXiv released OpenResearch desktop apps for running parallel research agents, with code and provenance logging built in. K-Dense's BYOK ("bring your own keys") assistant makes lab workflows auditable. Nvidia launched a full-stack platform for constraining agent behavior. The message: as AI agents act on the world — running experiments, controlling instruments, generating code — the ability to log every choice and every tool action is becoming a baseline requirement, not an afterthought.
Two stories underscored the tension between speed and safety. OpenAI reportedly cancelled a model release over safety concerns, while AMD announced it will acquire Fei-Fei Li's World Labs for $8.2 billion, signaling that spatial intelligence — AI that understands and acts in physical environments — is now a major commercial priority. For researchers, that means the tools arriving next year will be built to operate in labs and on benches, not just in notebooks.
What it means for you
Your workflow is about to change. When lab instruments speak MCP (Model Context Protocol) and AI assistants can call them, the line between designing an experiment and running it blurs. You will need to decide which steps you automate and which you keep manual — and you will need to log both. A result produced by a chain of AI tool calls is only as reproducible as the record of those calls.
Model selection is becoming a cost-performance decision, not a capability decision. GPT-6.1 Sol, Sonnet 5.5, and Gemini 4 Argon all sit in a narrow band of high capability. Your choice should hinge on which model's API, logging, and cost profile fits your experimental throughput. Record the model version and the prompt every time. If you switch models mid-project, document why.
The cancelled OpenAI release is a data point worth watching. When a major lab halts a model over safety, the reasons matter for your own risk assessment. If you are building on that provider's stack, you need a contingency plan for model deprecation or withdrawal. Keep your experimental code model-agnostic where possible.
Spatial intelligence and robot control (SAIL, Praxis-1 from Runway, Dyna-2.1) are moving toward general-purpose lab robotics. If your research involves physical manipulation — chemistry, materials, biology — start tracking what these systems can and cannot do. The claims are ambitious; independent validation is still thin.
What to focus on next week
- Audit one of your current experiments for AI provenance. Pick a result you generated with AI assistance. Can you reproduce it from logs alone? If not, add recording of model version, tool actions, environment constraints, and any human interventions.
- Test one of the new models on your own benchmark. Run GPT-6.1 Sol, Sonnet 5.5, or Gemini 4 Argon on a task you know well. Measure accuracy, latency, and cost. Record the comparison so you have a baseline for future model releases.
- Read the SAIL paper and the K-Dense lab-instrument MCP repository. Even if you do not work in robotics or lab automation, the design patterns — test-time search with simulator feedback, auditable instrument connectors — will shape tools arriving in your field within months.
- Draft a one-page policy for AI-assisted manuscript preparation. The week saw continued discussion of human-AI co-authorship. Decide now what you will disclose: which sections were drafted or edited by AI, which analyses were AI-generated, and where the human expert reviewed the output.
- Check your dependency chain for single-provider risk. If you rely on one model provider, outline what you would do if a model were withdrawn or its API changed without notice. Identify an alternative provider and run a trial inference this week.
These stories and the full week of Science and Research AI news are collected at all Science and Research AI news.