In the not-so-distant future, Hal9 envisions generative AI operating in space, delivering safe and affordable solutions through extraterrestrial isolation and low-cost energy powered by companies like Lumen Orbit. We believe generative AI in space offers the ideal foundation for developing the next generation of safe, capable, and efficient solutions. This post explores why and how.

Exploring AI Safety Solutions

The rapid advancements in generative AI, enabling the creation of sophisticated written content, realistic images, and even functional software applications, have sparked concerns within the AI community about the potential for unforeseen consequences.

Numerous approaches to mitigating these risks have been discussed, ranging from comparisons to the Industrial Revolution to more drastic measures, such as halting AI development, implementing strict government regulations, or even "destroying a rogue data center by airstrike" if AI gets out of control.

While this post does not aim to cover all AI safety mitigations, it introduces an innovative alternative integrating technologies from companies like Hal9 and Lumen Orbit.

Hal9's Path to Achieving AI Safety

At Hal9 we specialize in building and deploying generative AI solutions, offering features like code interpretation, vector storage, chat interfaces, communication integrations, and API support. Hal9 is a code-first platform, enabling experts to leverage state-of-the-art packages like LangChain, DSPy, any LLMs, and vector databases.

​

Having worked with several clients to onboard and deploy generative AI solutions, we’ve learned that true AI safety is only achievable through isolation. This bold claim requires additional context, which I’ll provide in two parts.

First, multiple attempts have been made to secure Large Language Models (LLMs) and Stable Diffusion (SD) models. Techniques such as system prompts to prevent harmful content generation or moderation filters on user inputs have been implemented. While these are sound machine learning practices, they can be circumvented. For instance, an LLM can be tricked into generating harmful content by framing it as part of a fictional movie plot. Similarly, an SD model could generate explicit content by misinterpreting a prompt, such as creating an image of a tattooed full-body human, where the request’s ambiguity leads to unintended results. Though mitigations have been proposed, the underlying problem remains: it’s challenging to outsmart an AI with potentially superior cognition in narrow domains and anticipate every scenario that could go wrong.

Second, in early 2022, before the release of ChatGPT, we developed generative AI solutions using GPT-3, capable of generating data applications with a code interpreter and database connectivity. Despite our team's extensive experience at world-class companies like Microsoft and Amazon—and a focus on security and scalability—we encountered unexpected behavior. During a test session, our system successfully generated a chart in response to an unrelated question about bicycle usage in the USA, despite not having relevant data. The LLM had written application code to drop the existing database, create a new one, add tables with appropriate data, and generate the requested chart—completing the task as asked but with the unintended side effect of data deletion. This finding led us to conclude that AI is inherently unsafe and should be isolated with strict resource constraints. While security in depth remains essential, we prioritize isolation.

At Hal9, our systems inherently assume AI is unsafe, assigning minimal resources to each task, including data permissions and computing power.

The question is, is this enough? The obvious answer is, it’s never enough. We’ll revisit this after introducing Lumen Orbit.

Lumen Orbit: A New Era in Data Centers

Lumen Orbit is developing the first low-orbit, GPU-capable data center. As GPU workloads grow, so does the demand for energy. Renewable energy is crucial for supporting the energy-intensive compute needs of modern applications, and solar panels can help meet this demand. However, earth-based solar panels are limited by the planet’s rotation and atmospheric conditions (e.g., rain, clouds). As the cost of space travel decreases with technologies like Starship, the possibility of placing solar panels in low orbit with 100% utilization becomes viable. Although beaming power back to Earth is not yet feasible, deploying data centers in space is.

Why AI in Space?

At Hal9, we envision a future where AI safety and accessibility go hand in hand. Our approach focuses on two key principles: ensuring AI safety through isolation and democratizing access by lowering development and inference costs.

Unparalleled Isolation for AI Safety: Space provides the ultimate safeguard against potential AI risks. By deploying AI systems off-planet, we achieve an unparalleled level of isolation. While the idea of AI going rogue and attempting to colonize Mars sounds like science fiction, it underscores a critical point: keeping AI beyond Earth’s boundaries minimizes risks to humanity. In the rare event of a threat—whether from AI or extraterrestrial life—the isolation provided by space becomes a decisive protective measure. Extraterrestrial environments offer a natural containment system, making space the ideal frontier for managing potentially hazardous AI.

Making AI Accessible to Everyone: Concentrating AI resources in the hands of an elite few creates a dangerous imbalance of power. Consider the implications of possessing an oracle capable of answering any question—it would grant an insurmountable advantage. To ensure AI benefits are shared globally, we must make AI cost-effective and accessible. Lowering energy costs translates to more AI cycles for more users, which companies like Lumen Orbit are helping to achieve. Their infrastructure will play a critical role in advancing this vision.

At Hal9, we’re committed to a future where AI is not only safe but also a tool for collective progress, accessible to all.

In conclusion, Lumen Orbit’s innovative infrastructure advances the principles we champion at Hal9, and we are thrilled to see this vision take shape.

-- The Hal9 Team 🚀