The Great Divide in AI
One of the most consequential debates happening in artificial intelligence right now is not about what AI can do, but about who gets to access it and how. This is the open source versus closed source debate, and it touches everything from how quickly AI advances to who benefits from it to how safe it is.
If you have followed technology news at all, you have probably seen headlines about Meta releasing Llama, Google keeping Gemini closed, or startups building on open models. But what do these terms actually mean in the context of AI, and why does it matter?
What "Open Source" Means in AI
In traditional software, "open source" has a clear definition: the source code is freely available, anyone can inspect it, modify it, and redistribute it. Linux, Firefox, and WordPress are classic examples. You can download their code, change it however you want, and share your changes with others.
In AI, things are more complicated. An AI model is not just code. It involves at least three distinct components, and different models share different combinations of these.
The three components of an AI model
Model weights. These are the learned parameters, the billions of numbers that encode everything the model has learned during training. Sharing the weights means someone else can run the model without needing to train it from scratch, which would cost millions of dollars.
Training code. This is the software used to train the model: the architecture, the training procedures, the optimization techniques. Having the code allows researchers to understand how the model was built and potentially improve upon it.
Training data. This is the dataset used to train the model. Knowing what data went into a model is crucial for understanding its biases, capabilities, and limitations. It is also the most contentious component to share, since training data often includes copyrighted material or personal information.
The spectrum of openness
Very few AI models are truly open source by the traditional definition. Instead, there is a spectrum:
Fully open. Both the weights, training code, and training data are publicly available. This is rare. OLMo from the Allen Institute for AI (AI2) is one of the few models that has released all three components, making it a genuinely open project that researchers can fully replicate and study.
Open weights. The model weights are available for download, but the training code and data may not be. Meta's Llama models fall into this category. You can run them and fine-tune them, but you cannot fully replicate the training process because you do not have the exact dataset or all the training details.
Partially open. Some information about the model is published, perhaps a research paper describing the architecture and training approach, but the actual weights and data remain private.
Fully closed. Nothing is shared beyond the API. You can use the model through a service, but you have no access to weights, code, or data. You cannot run it on your own hardware or inspect how it works.
Major Open Models
Llama (Meta)
Meta's Llama family of models has become the most widely used open-weight model series. Meta releases the model weights for free, allowing anyone to download and run them. Llama models come in various sizes, from relatively small models that can run on a laptop to large models that rival closed-source competitors.
Meta's motivation is partly strategic. By making strong models freely available, they reduce the competitive advantage of companies that sell AI access, while benefiting from the community's improvements and applications built on top of Llama.
However, Llama comes with a license that includes some restrictions. Very large companies with over 700 million monthly active users need special permission to use it, and there are restrictions on using it to train competing models. This means it is not "open source" in the purest sense, though it is far more accessible than closed alternatives.
Mistral
Mistral, a French AI company, has released several open-weight models that punched well above their weight in terms of performance relative to their size. Their approach demonstrated that smaller, well-trained models could compete with much larger ones, which was an important finding for making AI more accessible.
Mistral has since moved toward a hybrid approach, releasing some models openly while keeping their most advanced models closed. This trajectory is common in the industry: companies start open to build reputation and community, then become more closed as the commercial stakes increase.
OLMo (AI2)
OLMo, from the Allen Institute for AI, stands out because it is one of the most genuinely open AI projects. AI2 has released not just the model weights but also the training data, the training code, the evaluation framework, and detailed documentation of their process. Their goal is to enable scientific research into how language models work, which requires full transparency.
OLMo may not always top the performance leaderboards, but its contribution to the field is significant because it allows researchers to study, replicate, and build upon every aspect of the model.
Major Closed Models
GPT (OpenAI)
OpenAI's GPT series, including GPT-4 and its successors, are among the most capable AI models available but are entirely closed. You can use them through OpenAI's API or through ChatGPT, but you cannot download the models, inspect their weights, or run them on your own hardware.
OpenAI has published research papers about earlier models, but the trend has been toward less disclosure over time. The company argues that keeping models closed is partly a safety measure, preventing misuse, and partly a business necessity.
Claude (Anthropic)
Anthropic's Claude models are also closed. Like OpenAI, Anthropic provides access through an API and consumer products but does not release model weights. Anthropic's stated focus on AI safety is part of their rationale: they argue that controlling access to the most capable models helps manage risks.
Gemini (Google)
Google's Gemini models are primarily closed, accessible through Google's products and API. Google has released some smaller open models (like Gemma), but their frontier models remain proprietary. Google's vast infrastructure and data resources give them advantages that would be hard for others to replicate even with open weights.
The Open vs Closed Debate
This debate generates strong opinions on both sides, and both sides make compelling arguments.
Arguments for open models
Democratization. Open models allow anyone, from a university researcher in a developing country to a small startup, to use, study, and build upon cutting-edge AI. Without open models, AI power concentrates in a handful of wealthy corporations.
Scientific progress. Science advances through transparency and reproducibility. When models are closed, the AI research community cannot fully understand what works, what does not, and why. Open models enable the kind of rigorous study that pushes the whole field forward.
Trust through transparency. If you can inspect a model's weights and training data, you can audit it for biases, understand its limitations, and verify claims about its capabilities. With closed models, you have to trust the company's word.
Competition and innovation. Open models create a more competitive ecosystem. Thousands of developers and researchers can experiment with open models, often producing innovations that a single company would never have discovered.
Avoiding vendor lock-in. When you build your application on a closed model, you are dependent on that company's pricing, availability, and decisions. Open models give you more control and more options.
Arguments for closed models
Safety and misuse prevention. The most capable models could potentially be used for harmful purposes, from generating convincing misinformation to aiding in the development of dangerous materials. Keeping models closed provides a layer of control over who can use them and how.
Alignment and guardrails. Closed model providers can implement and maintain safety guardrails. With open models, anyone can remove these guardrails, creating uncensored versions that could be used for harmful purposes.
Business sustainability. Developing frontier AI models costs hundreds of millions or billions of dollars. Companies need revenue to fund this research, and giving away the results for free makes it harder to sustain the investment.
Quality control. Closed model providers can monitor how their models are used, identify problems, and push fixes. With open models, there is no central point of control or quality assurance.
Safety Implications
The safety dimension of this debate deserves special attention because it is where the stakes are highest.
The case that open models are safer
Some experts argue that open models are ultimately safer because they allow the entire research community to study and identify problems. Bugs, biases, and vulnerabilities are found faster when thousands of people can inspect the model. This is the same argument that has made open-source software generally more secure than proprietary alternatives over time.
Additionally, concentrating AI power in a few closed companies creates its own risks. What happens if those companies make bad decisions, are acquired, or face pressure from governments? A diverse ecosystem of open models is more resilient than dependence on a few gatekeepers.
The case that closed models are safer
Others argue that the most capable models are genuinely dangerous in the wrong hands and that controlling access is a reasonable precaution. As models become more powerful, the potential for misuse grows. A model that can help with legitimate chemistry research might also help someone synthesize dangerous substances. Keeping the most capable models behind an API with usage monitoring provides a safety net.
The practical reality
In practice, both approaches have a role to play. Most experts agree that moderate-capability models can be safely released openly, while the very frontier of capability might warrant more careful access controls. The challenge is that the frontier keeps moving. What was considered dangerously capable two years ago is now available as an open model, and society has adapted.
Business Implications
The open versus closed question has profound business implications that affect everyone from solo developers to Fortune 500 companies.
For companies building with AI
If you are building a product that uses AI, open models offer significant advantages. You can run them on your own servers, avoiding per-query costs that add up quickly at scale. You can fine-tune them on your specific data, creating a model that is uniquely suited to your use case. And you are not dependent on any single provider's pricing or availability.
However, closed models often offer higher raw performance, especially at the frontier. They also come with managed infrastructure, support, and regular updates. For many companies, the convenience and capability of closed models outweigh the flexibility of open alternatives.
For AI companies
The business model question is existential for AI companies. If your model is open, how do you make money? Some approaches include selling hosted services (making the model easy to use even though it is freely available), offering enterprise support and customization, and using the model as a gateway to other products and services.
Meta's strategy with Llama is a good example: the model is free, but it drives engagement with Meta's cloud infrastructure and helps Meta stay relevant in the AI ecosystem.
The cost factor
One of the most important practical considerations is cost. Running a large open model on your own hardware requires significant investment in GPUs and infrastructure. For many use cases, the API pricing of a closed model is actually cheaper than self-hosting an open one, at least until you reach a certain scale of usage.
Running Models Locally
One of the most tangible benefits of open models is the ability to run them on your own hardware. This has become increasingly accessible thanks to tools and optimizations that make it possible to run capable models on consumer hardware.
Why run locally?
Privacy. When you run a model locally, your data never leaves your machine. For sensitive applications, whether medical records, legal documents, or proprietary business data, this can be a critical advantage.
Cost at scale. If you are making thousands of API calls per day, running your own model can be significantly cheaper than paying per query to a cloud provider.
Customization. Local models can be fine-tuned on your specific data and configured exactly to your needs, without the constraints of an API provider's terms of service.
Offline access. A local model works without an internet connection, which matters for applications in remote areas, air-gapped environments, or simply on a long flight.
What you need
The hardware requirements vary dramatically based on the model size. Small models with 7 or 8 billion parameters can run on a modern laptop with 16 gigabytes of RAM, though responses will be slower than what you get from a cloud API. Larger models with 70 billion parameters or more typically require dedicated GPU hardware costing thousands of dollars.
Quantization techniques, which reduce the precision of model weights to make them smaller, have made local deployment much more practical. A quantized model uses less memory and runs faster, with only modest quality trade-offs.
Tools for local deployment
Several tools have made running open models locally straightforward, even for non-technical users. Applications like Ollama, LM Studio, and GPT4All provide user-friendly interfaces for downloading and running open models on your own computer. The experience is not as polished as using ChatGPT or Claude, but the gap is narrowing.
Where Things Are Headed
The open versus closed debate is not going to be resolved anytime soon. Both approaches will likely coexist, with open models serving as the foundation for a vast ecosystem of applications and research, while closed models push the frontier of what is possible.
Several trends are worth watching. First, the gap between open and closed models is narrowing. Open models that would have been considered state-of-the-art a year ago are now freely available. Second, hybrid approaches are emerging, where companies release some models openly while keeping others closed. Third, governments and regulators are starting to weigh in, which could reshape the landscape significantly.
Understanding this debate is essential for making sense of AI news, because virtually every major announcement involves a choice about openness, and that choice has implications for who benefits, who is at risk, and how the technology evolves.
See This in the News
The open model ecosystem continues to evolve rapidly, with new architectures and approaches pushing the boundaries of what openly available models can achieve. For a look at how open AI research is advancing, see this coverage of AI2's latest work: AI2 OLMo: Hybrid Transformer-RNN Open Model. It illustrates how organizations committed to full openness are experimenting with novel architectures that combine the best of different approaches, potentially reshaping what open models can do.