Here’s what’s on our plate today:

• 🧪 Open-source vs. closed: Who wins the AI future?

• 🧠 MIT job loss study, EU-US tensions, and AI fanfic.

• 💡 What the tip is about: Using model openness to stress‑test your stack.

• 🗳️ Should AI models be open-source or closed-source?

Let’s dive in. No floaties needed…

The Laboratory

Open or closed? Inside the debate reshaping the future of AI

In historical circles, there has been a long-standing debate on what matters more. The stories of the rise and fall of empires, kingdoms, and kings, or the stories of the rise, collapse, and metamorphosis of cultures and civilizations.

While on the surface the debate may appear like an intellectual exercise, the two different approaches present two very distinct points of view. Some historians argue that political events set the framework for cultural change, while others say culture is the engine that ultimately shapes history, no matter who sits on a throne.

The debate continues to this day because both sides offer valuable insights and neither completely explains the past on its own.

A somewhat similar debate also exists in the minds of technocrats. The debate: should software be open-source or closed-source?

For the uninitiated, open-source software is software whose inner workings are visible to everyone. The ‘source code’, which is the set of instructions that makes the program run, is open for people to look at, learn from, change, and share.

Closed-source software is the opposite. Its inner workings are kept secret by the company or developer who owns it. People can use the software, but they cannot see how it works or change it.

Though this debate has been going on since the late 1970s, when computers were just becoming popular outside research labs, it has taken on a new dimension with the emergence of AI.

The rise of open-source models

When ChatGPT was first launched in late 2022, proprietary models completely dominated. At the time, open-source projects had neither the necessary funding nor the access to powerful chips needed to build advanced models.

This was in stark contrast to the finances available to companies like OpenAI, which had raised $11 billion, and Anthropic, which had managed to raise $7.3 billion.

However, the dominance of closed-source AI did not last.

In early 2023, Meta released the first LLaMA models, which were much smaller yet performed impressively well. By mid-2024, Meta released Llama 3.1 405B and called it the first open-source model that could match or beat GPT-4. Around the same time, the French company Mistral released small models that outperformed much larger proprietary ones.

The pace only increased in 2025. DeepSeek R1 reached the same level as OpenAI’s o1 model while costing only $5.6 million to train, which was cheaper than previous systems.

China also embraced open-source AI as a way to close the gap with U.S. leaders, and many Chinese models now match the most advanced systems.

OpenAI joined the trend in 2025 by releasing two powerful open-weight models, gpt-oss-120b and gpt-oss-20b, under a fully open Apache 2.0 license.

The funding also started coming for open-source projects, with Meta investing $20 billion in AI-focused data centers to gain independence from third-party AI vendors.

Mistral also raised nearly $1 billion in funding, allowing it to grow from 35 employees in early 2024 to 350 employees in 2025, demonstrating the rapid scaling of open-source challengers.

Infrastructure platforms emerged to make open models accessible. Model developer platforms like Fireworks.ai, Together.ai, and Modal enhanced the accessibility and usability of open-source AI for enterprises, while Hugging Face became the de facto hub for model sharing and collaboration, hosting thousands of models and datasets.

Why big firms embrace openness

Companies are using open-source AI for reasons very different from those that drove earlier open-source software.

For Meta, releasing AI models openly helps spread its technology everywhere. When thousands of startups and developers build tools with Llama, Meta’s influence grows without having to sell anything directly.

This also brings indirect benefits: As Llama gets better, the AI assistants inside WhatsApp, Instagram, and Messenger improve too. Better assistants lead to more user engagement, which boosts Meta’s advertising business. Meta also benefits from cloud partnerships and rising demand for hardware used to train these models, even though the models themselves are free.

Similarly, for Mistral, open-sourcing is about positioning Europe as a serious competitor to the United States in AI.

The company has become a symbol of European technological independence, earning political support such as French President Macron promoting its chatbot, Le Chat. Mistral also secured government contracts, including a project to improve chatbots used by French civil servants.

For Chinese companies, open-source is a way to expand their global reach amidst increasing geopolitical tensions.

Alibaba opened its Qwen models to help spread AI more widely and encourage developers to build new applications, strengthening its cloud business. Open-sourcing also helps Chinese companies avoid U.S. export restrictions while building their own developer ecosystems.

Open-source also creates long-term cost advantages. Companies like OpenAI, Google, and Anthropic must pay huge computing costs every time someone uses their proprietary models through an API.

Meta, on the other hand, saves money as more people run Llama themselves. Organizations that use Llama on their own hardware avoid API fees entirely, making open-source even more attractive.

However, while companies have taken the path of open-source, not everyone agrees with their approach. Partially, because AI is unlike any software that was developed by human coders through years of hard work.

There is now a major argument about what ‘open’ really means in the world of AI.

The fight over what ‘open’ means

Unlike traditional open-source software, where the code and everything behind it are fully available, AI companies disagree on how much they need to share for a model to be considered open. The biggest disagreement is about training data.

Every major company releases model weights but keeps the datasets secret. The Open Source Initiative has tried to set rules for what counts as open-source AI, but some purists argue that without the original training data, a model can’t truly be open because users cannot fully understand or recreate it. OSI counters that the weights matter more, since they are what users actually work with.

It is important to note here that model weights are the learned numerical parameters that turn inputs into outputs; changing them changes behavior. On the other hand, training data is used to teach the model; changing it changes what the model can learn.

As such, to be genuinely open-source, others must study, verify, reproduce, and improve the system.

Weights alone let you run and tweak the result, but without the data, you can’t audit legality/bias or faithfully reproduce training.

Companies avoid sharing training data because it is expensive to collect, and tied to copyright and privacy rules. Even if the data were public, only groups with huge budgets could retrain these large models, so true openness isn’t very realistic.

This has led to complaints of open-washing, where models are called open even though they have rules attached.

Safety fears and pushback

Critics of open-sourcing also argue that allowing access to powerful AI models could be dangerous. Some, like Vinod Khosla, say it creates national-security risks by giving advanced U.S. technology to rivals like China. Others worry that open access lowers the barrier for harmful uses, especially for less-skilled actors such as terrorists or lone individuals.

Even governments have talked about the risks of model misuse for nefarious purposes like developing biological weapons and compromising cybersecurity.

Regardless, open-source models continue to develop at a rapid pace and are likely to catch up with closed-source models.

And with that, regulations are expected to grow more complicated in a bid to address the use of both open and closed source AI models, as well as the use of platforms that use a hybrid approach.

The unending debate

In the meantime, just as historians are unable to come to a consensus on whether the story of humanity is best told through rulers or through cultures, the AI world is unlikely to settle its own dispute over openness.

Both lenses reveal something important, and both leave something out. Open and closed models will continue to shape one another, just as politics and culture always have. Adding to the complexity is the question of what open-source AI actually means, and you have an entanglement that may become part of the future and process of AI and its history.

Roko Pro Tip

Bite-Sized Brains

• AI can replace 11.7% of jobs: A new study says current AI could already do nearly 12% of U.S. jobs, but cost still holds it back.

• U.S. and EU still clash on AI: Promises of alignment aside, the West’s two powers remain divided on AI rules.

• Character.ai launches fan stories: Teens can now co-write AI-powered stories with their favorite characters.

Monday Poll

Meme of the Day

Rate This Edition