Welcome back to LLM Decode 👋
AI is moving beyond screens and software into two very different frontiers: space and biology. Google is testing whether AI compute can operate in orbit, while NVIDIA and the University of Washington are opening up a model that can help scientists design entirely new proteins.
Here’s what matters today.
Google Is Taking AI Compute Into Orbit
Google's Project Suncatcher has reached its first real-world test. On October 1, a prototype satellite developed with Planet launched aboard SpaceX's Transporter-18 rideshare mission, carrying Google AI hardware into orbit. Google has confirmed contact with the satellite and says it is operating as expected.
The experiment is testing whether Google's Tensor Processing Units (TPUs) can withstand the radiation, vibration and extreme thermal conditions of space. Cooling is a particular challenge because conventional airflow isn't available in a vacuum, so the team is testing heat pipes and radiators instead.
The long-term vision is much more ambitious: Google is exploring interconnected, solar-powered satellites that could eventually operate as scalable machine-learning infrastructure. Future designs could use high-bandwidth laser links to connect satellites and distribute AI workloads across an orbital network.
Why it matters
AI infrastructure could eventually extend beyond Earth: Space offers near-constant sunlight and avoids some terrestrial constraints.
This is still experimental: Google is testing fundamental engineering questions before considering large-scale deployment.
AI compute is becoming an infrastructure-design problem: Where and how models are powered could become as important as the chips themselves.
Satellite-to-satellite networking will be crucial: Large-scale orbital AI would require extremely fast communication between compute nodes.
AI Is Learning to Design New Proteins
NVIDIA and the University of Washington's Institute for Protein Design have introduced RFD4-Proteina, an open AI model designed to generate targeted protein designs. Rather than simply predicting existing molecular structures, the model helps researchers search for proteins with specific properties and functions.
Researchers can give the model molecular inputs, constraints and scoring objectives, allowing it to generate and refine potential designs. NVIDIA says the model showed higher computational design success rates and generation throughput than evaluated baselines across protein-binding, small-molecule binding, enzyme-scaffolding and DNA-binding benchmarks.
The model is being made available through partners including Hugging Face, Google Cloud, OpenAI, Anthropic, Benchling and Databricks, allowing researchers to incorporate protein design into broader scientific and agentic workflows. The designs still need to be validated experimentally in the laboratory.
Why it matters
AI is moving from prediction to generation in biology: Researchers can explore designs that don't already exist in nature.
Open models could accelerate scientific collaboration: More researchers can experiment with the technology and build specialized workflows.
The laboratory remains the final test: Computationally promising designs still need real-world validation.
Drug discovery could become more iterative: AI can generate candidates, scientists can test them, and experimental results can guide the next design cycle.
4 AI Workflows to Try
Search, don't just generate: Use AI to explore large possibility spaces - materials, molecules, code or research ideas, rather than asking for a single answer.
Connect AI to specialist tools: The biggest gains can come when a model can call domain-specific software, databases and simulations instead of working from text alone.
Keep humans in the validation loop: For scientific, financial or operational decisions, treat AI outputs as candidates to test—not automatically correct conclusions.
Watch the “AI + physical infrastructure” trend: Space, laboratories, factories and autonomous vehicles are becoming new environments where AI can operate.
Ready to level up your AI skills?
Explore Our Courses →That’s it for today.
The AI space doesn’t slow down - and neither should your thinking.
See you in the next drop.