We Hosted Falcon 40B on Our Private Cloud in One Hour. Here Is Why That Matters.

The rapid evolution in AI is astounding. We never thought we would be here already, just seven months after the GPT-3 breakout.

Meet Falcon 40B: an open-source alternative, available on Hugging Face, that outperforms GPT-3 and Meta's LLaMA in some areas. Even more notable, it allows commercial use under the Apache 2.0 licence.

It took us exactly one hour to host this model on our private cloud infrastructure.

Let that sink in. State-of-the-art language AI, running entirely on infrastructure we control, with no data leaving the building. For companies with privacy constraints, regulated data, or simple reluctance to ship their knowledge to an external API, this changes the calculation completely.

This opens an enormous range of opportunities for business automation. At the same time, it is sobering to see how easily it can be misused. The built-in guardrails we know from GPT are missing, which allows the generation of harmful content or misinformation. Anyone deploying open-source models carries that responsibility themselves.

Both sides of that coin, the opportunity and the responsibility, will define how these technologies reach the workplace. We intend to keep testing them hands-on as they do, because credible advice about this technology can only come from using it ourselves.