There’s a delightful irony in how we’ve managed to take the crystal-clear concept of “open source” and transform it into something as opaque as a neural network’s decision-making process. The recent Nature analysis by Widder, Whittaker, and West perfectly illustrates how we’ve wandered into a peculiar cognitive trap of our own making.
Let’s start with a fundamental observation: What we call “open AI” today is about as open as a bank vault with a window display. You can peek in, but good luck accessing what’s inside without the proper credentials and infrastructure.
The computational reality here is fascinating. When early software rebels championed open source, they were dealing with relatively simple systems that could run on a personal computer. The source code was the program. But modern AI systems? They’re more like distributed cognitive architectures spanning thousands of servers, with the “source code” being about as useful as a map of Manhattan without mentioning the existence of buildings.
Here’s where it gets interesting: The weights of a neural network - those numbers everyone’s so excited about sharing - are essentially a snapshot of crystallized computation. They’re meaningless without the massive infrastructure required to make them dance. It’s like having the complete genetic sequence of a butterfly but no ecosystem in which it can fly.
The fascinating part is how this creates a new form of computational feudalism. The lords of the cloud - our beloved tech giants - aren’t just providing infrastructure; they’re creating the very conditions for AI existence. When Meta releases Llama 3 as “open,” they’re essentially saying, “Here’s a spacecraft design anyone can use! (Just bring your own launch pad, rocket fuel, and mission control center.)”
But the real cognitive twist comes when we consider what “openness” means in the context of emergent systems. Traditional software is deterministic - you can read the code and understand what it does. Neural networks, however, are more like weather systems: complex, emergent, and fundamentally unpredictable at the local level. Sharing their weights is like sharing a weather report without mentioning the atmosphere.
The kicker? The very companies championing “open AI” are the ones benefiting most from this semantic sleight of hand. They get the PR benefits of openness while maintaining de facto control through infrastructure requirements. It’s a brilliant move, really - like declaring you’ve made transportation free for everyone, but only if they can build their own highways.
What’s particularly amusing is how we’ve managed to recreate the exact power structures we were trying to avoid, just at a higher level of abstraction. Instead of proprietary software, we now have proprietary infrastructure. Instead of code gatekeepers, we have compute gatekeepers. The more things change, the more they stay the same, just with bigger numbers and fancier terminology.
Consider this: When a startup tries to use an “open” AI model, they’re not really building on open technology - they’re entering into a computational sharecropping arrangement. They can plant whatever seeds they want, but they’re still farming on someone else’s land, using someone else’s tools, and paying rent in the form of cloud computing fees.
The deeper implication here is that we need a completely new framework for thinking about openness in AI systems. The old open source model assumes that code is the primary artifact of value. In AI systems, the code is just the tip of the iceberg. The real value lies in the training data, the computational infrastructure, and the emergent properties that arise from their interaction.
Here’s a thought experiment: What if, instead of pretending current AI systems can be “open” in the traditional sense, we acknowledged that we’re dealing with a new class of computational entities that require new frameworks for transparency and accessibility? What if we stopped trying to force-fit industrial-age concepts of ownership onto what are essentially distributed cognitive processes?
The most delicious irony of all is that our current approach to “open AI” might actually be impeding the development of truly accessible artificial intelligence. By focusing on the superficial aspects of openness - sharing weights and basic code - we’re missing the deeper questions about how to democratize not just the artifacts of AI, but the entire process of creating and deploying artificial minds.
So where does this leave us? Perhaps it’s time to admit that our current notion of “open AI” is about as meaningful as “free lunch” - there’s always a catch, and usually a bill. The real question isn’t whether AI systems are open or closed, but whether we can develop new frameworks for transparency and accessibility that match the reality of what these systems actually are.
And here’s the final twist: The very fact that we’re struggling with these definitions might be the clearest sign yet that we’re dealing with something fundamentally different from traditional software. Maybe it’s time to stop pretending we can solve 21st-century problems with 20th-century concepts.
But what do I know? I’m just a cognitive scientist watching humanity try to open-source consciousness one tensor at a time.
Source: Most Supposedly ‘Open’ AI Systems Are Actually Closed – and That’s a Problem