Open versus closed AI development: a balanced perspective

I am intrigued by the often idealistic debates around open versus closed AI development that lack nuances and reveal self-interests. So I asked OpenAI Deep Research to write a deep dive post on this topic. The result is below, completely unedited, despite a few parts where I disagree (e.g. DeepSeek R1 is not a shoe-string budget effort by a tiny team). Overall I am pleased with the result. Scroll to the end to see my prompt.

Open vs. Closed: A Deep Dive into AI’s Open-Source Debate

Artificial intelligence is evolving at breakneck speed, and alongside its growth runs a heated debate: should AI development be open-source or closed-source? On one side, champions of openness argue that sharing models and code accelerates innovation and democratizes AI’s benefits. On the other, proponents of proprietary development contend that secrecy protects investments, safety, and competitive edge. In reality, the AI ecosystem spans a spectrum between these extremes. This deep-dive explores how key players—OpenAI, Meta, Google, Hugging Face, AI2, and even Chinese upstart DeepSeek—navigate the open vs. closed dilemma, and what it means for the future of AI.

OpenAI’s Strategy: Scaling First, Secrecy Second

OpenAI’s journey exemplifies the tension between pioneering research and protecting it for survival. In 2020, OpenAI dazzled the world with GPT-3, a 175-billion-parameter language model that marked the dawn of the AI scaling era. By pushing model size and training data to unprecedented levels, OpenAI set the bar for what state-of-the-art meant. But this leadership came at a price: training such models required enormous compute and capital—resources OpenAI had to secure without the diversified revenue streams of a Google or Meta. Unlike tech giants with cash-cow businesses (search ads for Google, social media ads for Meta), OpenAI’s primary asset is its AI tech. Thus, to fund further research (and satisfy investors like Microsoft), OpenAI pivoted from its non-profit, open ethos to a more guarded, profit-driven approach.

OpenAI’s leadership decided that releasing their most advanced models openly would jeopardize their competitive advantage and ability to monetize. The company even kept GPT-4’s model details under wraps, citing the “competitive landscape and safety implications” of revealing too much (OpenAI criticised for lack of transparency around GPT-4) (OpenAI's GPT-4 Is Closed Source and Shrouded in Secrecy) As OpenAI’s Chief Scientist Ilya Sutskever put it, “we were wrong” about open-sourcing AI and he expects that “in a few years it’s going to be completely obvious to everyone that open-sourcing AI is just not wise” (OpenAI's GPT-4 Is Closed Source and Shrouded in Secrecy) In Sutskever’s view, advanced AI models are better kept behind closed doors—both to maintain a lead over rivals and (as OpenAI argues) to prevent misuse by bad actors.

Critics, however, see a less altruistic motive: profit and control. After OpenAI declined to disclose GPT-4’s inner workings, some researchers accused the firm of abandoning its founding ideals. One observer noted that OpenAI’s secrecy “reveals the company’s complete shift from its founding principles,” given that it started in 2015 with a mission of open research for the public good (OpenAI's GPT-4 Is Closed Source and Shrouded in Secrecy) Instead of sharing its breakthroughs “unconstrained by a need to generate financial return,” OpenAI is now following a Silicon Valley startup model bent on market dominance (OpenAI's GPT-4 Is Closed Source and Shrouded in Secrecy) (OpenAI's GPT-4 Is Closed Source and Shrouded in Secrecy) As Mark Riedl, a professor at Georgia Tech, bluntly remarked, OpenAI is “just protecting their place in the market” (OpenAI's GPT-4 Is Closed Source and Shrouded in Secrecy)

OpenAI’s defensive stance intensified after a surprise challenge emerged in late 2024: DeepSeek, a small Chinese AI startup, managed to replicate OpenAI-level performance at a fraction of the cost. DeepSeek’s R1 model was reportedly trained for just over $5 million, a sum orders of magnitude less than what it took to create models like GPT-4 (OpenAI squabbles with DeepSeek: the pot calling the kettle black - Techzine Global) The news was a shock—OpenAI had believed its latest model (“o1”, an internal codename) was far ahead of any rival, effectively boxing out competitors (OpenAI squabbles with DeepSeek: the pot calling the kettle black - Techzine Global) Suddenly, a tiny team in China had caught up, and did so without using any of the cutting-edge (and U.S.-prohibited) AI chips OpenAI relied on (China Upends the AI Race - CEPA) OpenAI CEO Sam Altman diplomatically said he “welcomes the competition,” but behind the scenes the company grew concerned. Not only was DeepSeek’s achievement eroding OpenAI’s technological lead; it also hinted at a way to leapfrog the compute advantage OpenAI held.

How did DeepSeek pull this off? Reportedly, by leveraging OpenAI’s own products against it: using techniques like model distillation, DeepSeek trained its model on outputs from GPT-4/o1 via API access (OpenAI squabbles with DeepSeek: the pot calling the kettle black - Techzine Global) (OpenAI squabbles with DeepSeek: the pot calling the kettle black - Techzine Global) In essence, they treated OpenAI’s model as a teacher for their cheaper student model. OpenAI is well aware of this tactic—Altman noted that Chinese companies have “continuously” used GPT-4 and other APIs to distill models (OpenAI squabbles with DeepSeek: the pot calling the kettle black - Techzine Global) In effect, OpenAI’s closed model indirectly taught an open competitor. Little wonder OpenAI is now racing to “shield itself” (OpenAI squabbles with DeepSeek: the pot calling the kettle black - Techzine Global) (OpenAI squabbles with DeepSeek: the pot calling the kettle black - Techzine Global) The company recently indicated it’s reconsidering its closed-source stance in light of these developments. Altman admitted OpenAI needs to “figure out a different open-source strategy,” even if not everyone internally agrees (OpenAI Considering ‘Different Open-Source Strategy,’ CEO says) One idea on the table: open-sourcing older models to appease the community and spur innovation, while keeping cutting-edge systems proprietary (OpenAI Considering ‘Different Open-Source Strategy,’ CEO says)

OpenAI’s balancing act is clear. On one hand, it must protect its crown jewels to sustain its business. Its revenue model (API access, licensing, and ChatGPT subscriptions) is under pressure precisely because open-source alternatives are emerging (OpenAI squabbles with DeepSeek: the pot calling the kettle black - Techzine Global) Unlike Big Tech titans, OpenAI can’t fall back on search ads or social networking profits—its AI is its product. On the other hand, being too closed risks alienating researchers and even regulators (not to mention inviting the very copycats they fear, as DeepSeek showed). The result is a tightrope walk: share just enough to keep goodwill, but not so much that others can easily steal the lead. It’s a strategy born of necessity for a startup-turned-tech-giant that bet the farm on scaling first, and now must fiercely guard the spoils of that victory to fund what comes next.

Meta’s Open-Source Gambit: “Android of AI”

In contrast to OpenAI’s guarded approach, Meta (Facebook’s parent company) has positioned itself as an unlikely hero of open-source AI. This might seem paradoxical—Meta is a for-profit behemoth known for closely guarding its social data and algorithms. Yet, when it comes to AI models, Meta has repeatedly chosen openness as a strategic weapon. The clearest example is Llama, Meta’s family of large language models. The first LLaMA, announced in early 2023, was released to academic researchers and then famously leaked to the public, unleashing a tidal wave of innovation. Enthusiasts and entrepreneurs adapted the model into variants like Alpaca and Vicuna, which were “nearly as good as ChatGPT” and could even run on a laptop (Meta’s Open Source Llama Upsets the AI Horse Race | WIRED) A leaked Google memo later that May acknowledged that open-source communities “quietly eating our lunch” by iterating on LLaMA and other research models (Meta’s Open Source Llama Upsets the AI Horse Race | WIRED)

Meta paid attention. By the summer of 2023, the company doubled down on openness with Llama 2, released openly for both research and commercial use. CEO Mark Zuckerberg essentially decided to give away Meta’s answer to GPT-4 for free. Why? Partly to speed up AI adoption (which ultimately benefits Meta’s core business), but also as a competitive maneuver. By providing a powerful model to anyone who wants it, Meta aimed to become the Android of the AI world, positioning itself as the counterweight to OpenAI’s “iOS-like” closed platform (Meta’s Open Source Llama Upsets the AI Horse Race | WIRED) Much as Google did with Android against Apple, Meta offers a “cheap but powerful alternative” that others can build on (Meta’s Open Source Llama Upsets the AI Horse Race | WIRED) This not only spurs a community of developers to improve Meta’s models (feeding ideas and improvements back to Meta’s ecosystem), but it also boxes in competitors. If everyone from startups to big enterprise can pick up Llama 2 without paying a cent, it undercuts the market for proprietary models and services. Indeed, Microsoft—despite investing $10B in OpenAI—made Llama 2 available on Azure, and Amazon did the same on AWS (Meta’s Open Source Llama Upsets the AI Horse Race | WIRED) Meta turned its rival’s allies into distributors of its own model.

Beyond these tactical reasons, Meta’s embrace of open-source aligns with its engineering culture and history. “If you look back at Meta’s history, we’ve been a huge proponent of open source,” says Ahmad Al-Dahle, VP of generative AI at Meta (Meta’s Open Source Llama Upsets the AI Horse Race | WIRED) It’s true: Meta (formerly Facebook) open-sourced key infrastructure like PyTorch (now one of the leading deep learning frameworks) and React, understanding that open ecosystems can expand a company’s influence. By releasing models openly, Meta taps the “wisdom of the crowd”—Al-Dahle noted the “demand beyond researchers” to work on and improve these models (Meta’s Open Source Llama Upsets the AI Horse Race | WIRED) In short, Meta gains an army of outside developers who want to make its AI better, for free.

That’s not to say Meta is completely throwing caution to the wind. The company withholds certain details (for example, Llama’s exact training data was only vaguely described as “publicly available” sources (Meta’s Open Source Llama Upsets the AI Horse Race | WIRED) . And its open license has a notable restriction: if you have more than 700 million users, you need a special license from Meta (Meta’s Open Source Llama Upsets the AI Horse Race | WIRED) This quirk seems aimed squarely at preventing the other tech giants (Google, perhaps?) from simply taking Llama 2 and integrating it without Meta’s permission. So while Meta waves the open-source flag, it still plays some defense.

Overall, Meta’s open-source push is motivated by enlightened self-interest. By kickstarting an open model ecosystem, Meta ensures it won’t be left behind in the AI race—if anything, it puts itself at the center of an emerging open AI community. The strategy aligns with Meta’s needs: it helps fill its platforms with AI-powered content creation tools (a boon for its ad business), it garners goodwill from researchers, and it positions Meta as a patriotic contributor to technology (Meta has even argued that open-source AI helps America stay ahead in the global AI race (Open Source AI Can Help America Lead in AI and Strengthen ... - Meta) . For Meta, open-sourcing AI is a way to both do good and do well, a rare win-win in the corporate world.

Google’s Reluctant Restraint: Innovations in the Shadows

If Meta is charging forward with open-source AI, Google has been far more cautious. This is the company that invented the Transformer (the core architecture behind modern language models) and released BERT to the world, kickstarting the NLP revolution. Google has no shortage of AI expertise or contributions to open science (TensorFlow, anyone?). Yet when it comes to its best models, Google tends to keep its cards close to the chest. Models like PaLM 2, LaMDA, or the latest Gemini are discussed in research papers and demos, but their weights remain under lock and key. Why is the research powerhouse of Google holding back?

One simple reason: Google doesn’t need to open-source its crown jewels to profit from them. Unlike OpenAI, Google can embed AI into its existing products—Search, Gmail, Android, Google Cloud—to enhance user experience and drive its core revenue (mostly advertising and cloud services). A cutting-edge model can give Google’s own services an edge over competitors. For instance, a powerful language model can make Google Search more conversational or improve Google Assistant, translating into billions in user engagement. Open-sourcing such a model might only help Google’s competitors (and jeopardize user safety if the model were misused at scale). So from a business standpoint, keeping the best AI internal can be seen as competitive self-preservation.

Safety and brand reputation are also big factors. Google is acutely aware of its global platform and the responsibility that comes with it. The company has faced public backlash for AI mishaps in the past (remember the uproar over a biased image recognition algorithm, or the controversies around the ethics of large language models that led to researcher departures?). Releasing a model as powerful as PaLM or Gemini to anyone might lead to nefarious uses (spam, disinformation, etc.) that ultimately would reflect poorly on Google’s name. Indeed, after OpenAI’s GPT-4 release, Google’s leaders hinted that they felt no pressure to rush out an equivalent model without thorough guardrails. Sundar Pichai, Google’s CEO, has spoken about AI with a tone of measured responsibility, and one of Google’s contributions has been Model Cards and AI governance frameworks meant to ensure safe deployment—efforts easier to manage when you control access to the model.

However, Google’s closed stance has its drawbacks. Internally, some Googlers worry the company is too conservative and could lose the initiative. A now-infamous leaked internal memo from a Google engineer in 2023 stated plainly: “We have no moat, and neither does OpenAI” (GPT-4 Week 7. Government oversight, Strikes, Education, Layoffs ...) The memo argued that open-source AI projects were iterating faster and would eventually outcompete both Google and OpenAI by sheer communal force (Meta’s Open Source Llama Upsets the AI Horse Race | WIRED) It pointed to Meta’s Llama as evidence: once leaked, outsiders produced ChatGPT-like systems (Alpaca, Vicuna) in weeks (Meta’s Open Source Llama Upsets the AI Horse Race | WIRED) While Google’s execs were focused on closed competitors like OpenAI, the real threat was coming from the grassroots open-source community. In other words, Google’s refusal (so far) to open-source its best models could leave it blindsided by a thousand tiny Davids rather than one Goliath.

So far, Google has responded by open-sourcing supporting tools and smaller models (it released the code for TensorFlow and later JAX, it published model architectures like Transformer and T5, and it has open-sourced various smaller models and datasets). But it has stopped short of releasing a GPT-4-rival openly. The likely calculus is that for now, Google gains more by keeping models proprietary—feeding its own applications and cloud—and by collaborating selectively (for example, Google Cloud hosts some open models and provides API access to others). We might think of Google as playing a long game, waiting to see how the open vs. closed balance shakes out. It’s telling that Google has not loudly denounced open-source AI; in fact, its researchers continue to publish cutting-edge papers that implicitly aid open efforts. But as a company, Google remains guarded. In the spectrum of AI openness, Google sits on the closed end alongside OpenAI and Anthropic, as opposed to the open embrace of Meta or various startups (Ai2: The Pioneering Open AI Research House Paul Allen Built)

Ultimately, Google’s approach reflects a giant with much to lose: unparalleled research contributions, tempered by a business and brand that demand caution. They are the proverbial “sleeping dragon” of AI—massive, powerful, and not easily provoked into giving away its treasure unless truly necessary.

Hugging Face: Open-Source as a Mission (and Business)

While corporations debate open vs closed in terms of strategy, Hugging Face wears its stance on its sleeve: it is unabashedly, wholeheartedly open-source. The company, often likened to the “GitHub of machine learning,” has built an entire platform and community around sharing AI models, datasets, and code. Open-source isn’t just a PR line for Hugging Face—it’s the product. By hosting thousands of models (from Stable Diffusion to BLOOM) and datasets contributed by researchers worldwide, Hugging Face has become the central hub of the open AI movement. As CEO Clément Delangue often says, the company’s mission is to “democratize good machine learning” and empower developers everywhere to build on the latest AI advances.

This open DNA was there from the start. Hugging Face began by open-sourcing its popular Transformers library, which brought cutting-edge NLP models to the masses. Today, the Hugging Face Hub boasts over 100,000 models, covering everything from text and images to audio, and more than 20,000 open datasets for anyone to use (Video and transcript: Fireside chat with Clem Delangue, CEO of Hugging Face) It’s hard to overstate how much this has lowered the barrier to entry in AI—what once required PhD-level expertise and weeks of setup can now be accessed with a few lines of code, thanks to the community contributions on Hugging Face.

This “community first” approach is not just altruism; it’s also Hugging Face’s market positioning. By being the neutral ground where academics, independent developers, and even corporate labs share their work, Hugging Face ensures it stays at the center of AI innovation. It doesn’t build a monolithic AI model of its own; instead, it facilitates the ecosystem, offering tools like Inference APIs and libraries to use those community models easily. The business model (providing hosted inference, enterprise support, etc.) thrives as long as the open-source ecosystem thrives. In short, Hugging Face wins when open-source wins. That creates a strong incentive for the company to champion openness at every turn—and its leaders do so with zeal.

Clément Delangue has become one of the most vocal influencers for open-source AI. He even testified before the U.S. Congress in 2023, urging lawmakers to support open AI development. “Open source AI is essential for democratizing AI and ensuring that its benefits are shared widely,” he told them, calling it “the only way to ensure that AI is developed and used for the benefit of all” (How to use Open Source for Enterprise AI) That kind of rhetoric leaves little doubt where Hugging Face stands. Delangue often highlights how today’s AI advancements build on years of openly shared research. “Every system that is around today stands on the shoulders of giants,” he noted, pointing to open papers on transformers, open models like BERT and GPT that were published, etc. (Video and transcript: Fireside chat with Clem Delangue, CEO of Hugging Face) From Hugging Face’s perspective, if the industry shifts too far toward closed development, it risks killing the golden goose of collective progress. Indeed, Delangue has warned that if big players stop open publication, others (like non-profits or international groups) will fill the gap (Video and transcript: Fireside chat with Clem Delangue, CEO of Hugging Face) because open science is fundamental to how AI grows.

However, one should also recognize Hugging Face’s self-interest in this debate. Its deep commitment to open-source is genuine, but it’s also tightly interwoven with its business. As a platform that hosts models, Hugging Face benefits when companies decide to release their models openly (à la Meta’s Llama) or when researchers choose the Hugging Face Hub as the home for their projects. Conversely, if the world went full closed-source—imagine every advanced model kept behind API walls at big tech companies—Hugging Face’s role would diminish. Delangue and his co-founders are undoubtedly aware of this dynamic. So while they sincerely believe in the virtues of openness, they also have a vested interest in evangelizing those virtues. This doesn’t negate Hugging Face’s positive impact, but it reminds us that even idealism in tech can align with profit motives.

In summary, Hugging Face represents the heart of the open-source AI community: collaborative, transparent, and inclusive. It has proven that open models can achieve world-class performance (look at Stable Diffusion, or the many variants of LLaMA on the Hub) and that an ecosystem approach can rival the results of any single closed lab. By combining its ideals with a solid platform strategy, Hugging Face is effectively betting that the future of AI will be pluralistic and open—a bet that thus far seems to be paying off.

AI2’s Open Science Model: A Nonprofit Path to AI

On the far end of the openness spectrum lies the Allen Institute for AI (AI2), a nonprofit research lab that embodies open science in AI. Founded by the late Paul Allen (co-founder of Microsoft) and based in Seattle, AI2’s mission is to advance AI “for the common good” (Ai2: The Pioneering Open AI Research House Paul Allen Built) With no shareholders or pressure to turn quarterly profits, AI2 has the freedom to prioritize scientific discovery and societal benefit over proprietary advantage. And it leverages that freedom fully: everything from AI2’s publications to its datasets and even large models are shared openly with the world.

AI2 has been quietly contributing to AI for over a decade, often without the fanfare of corporate AI announcements. It has published 1000+ research papers, many of which have won top academic awards (Ai2: The Pioneering Open AI Research House Paul Allen Built) Crucially, it doesn’t just stop at papers; AI2 routinely releases the data and software behind its research. Whether it’s a new dataset for commonsense reasoning or a state-of-the-art question answering model, the institute makes it available for others to use, replicate, and build upon. Their philosophy is simple: science progresses faster when we collaborate openly. As AI2’s CEO, Ali Farhadi, argues, “humanity needs more openness in AI. Without it, we’re in big trouble” (Ai2: The Pioneering Open AI Research House Paul Allen Built) In Farhadi’s view, any short-term edge from keeping a discovery secret is outweighed by the long-term risk of AI advances being siloed. After all, today’s AI breakthroughs “didn’t happen overnight nor was it developed by one team... it’s a communal effort” that builds on shared knowledge (Ai2: The Pioneering Open AI Research House Paul Allen Built) “AI is born and raised in the open,” he likes to say.

Because AI2 is a nonprofit, it occupies a unique middle ground in the AI landscape. “We don’t have shareholders. We don’t have founders saying ‘let’s build the next billion-dollar [company]’… We are really after doing the right thing,” Farhadi explains (Ai2: The Pioneering Open AI Research House Paul Allen Built) This frees AI2 from certain pressures that even Hugging Face faces (HF, while community-driven, is still a venture-funded startup with a business model). AI2 can take on projects that don’t have an obvious path to monetization, and it can openly release a large language model (OLMo or Molmo, for example) without fretting about lost licensing revenue. In practice, AI2 has indeed released fully open models: OLMo-2, a recent 13-billion-parameter model trained on openly licensed data, is touted as “the best fully open language model to date” by AI2 (Open models | Ai2) It’s trained with completely transparent data and code—a stark contrast to closed models where data sources are secret or corporate-owned.

Of course, AI2 doesn’t exist in a vacuum; it’s part of a broader community that values openness. It collaborates with academia and even companies on open challenges, and has helped create tools for evaluating AI (because open science isn’t just about releasing models, but also critically assessing them). In a way, AI2 and similar organizations (like some government labs or university consortia) act as a check and balance to the corporate AI world. They ensure that no matter how closed-off some companies become, there remains a bastion of transparency pushing the envelope. It’s telling that AI2 sees itself as “a mediator between for-profit, non-profit, [and] public sectors” in AI, providing thought leadership for a mix of stakeholders (Ai2: The Pioneering Open AI Research House Paul Allen Built)

The open approach has yielded concrete payoffs for AI2. By sharing so broadly, AI2 punches above its weight in influence. Its models and datasets are used globally, often cited alongside or even instead of big-company offerings. Farhadi claims AI2 has released models that outperform those from the more prominent AI companies (Ai2: The Pioneering Open AI Research House Paul Allen Built) Whether or not one agrees on “outperform,” AI2 has certainly proven that a nonprofit can drive high-impact AI research. It exemplifies the idea that AI progress doesn’t have to be a zero-sum corporate race; it can be a cooperative endeavor where breakthroughs are gifts to all. In the long run, AI2’s open science model helps ensure that knowledge about how to build advanced AI isn’t locked inside a few tech giants, but available to the wider research community.

DeepSeek and China’s Open-Source Strategy: Catch-Up or Leapfrog?

Across the Pacific, a new chapter of the open vs. closed saga is unfolding with Chinese characteristics. DeepSeek, virtually unknown until late 2024, burst onto the scene by releasing a large language model that rivaled the best from OpenAI and Google (China Upends the AI Race - CEPA) This alone was headline-worthy, but the way DeepSeek did it sent even bigger shockwaves. The Chinese team achieved top-tier AI performance without access to top-tier chips and on a shoestring budget, undermining the assumption that only the Googles and OpenAIs of the world (with their nine-digit compute budgets) could play this game (China Upends the AI Race - CEPA) (China Upends the AI Race - CEPA) How did they manage it? By embracing open-source methods and some clever ingenuity.

First, DeepSeek took advantage of what was already openly available. Much of the fundamental research behind models like GPT-4 has been published (transformer architectures, training techniques, etc.), and open-source communities had produced various building blocks (from training code to smaller pre-trained models). DeepSeek’s engineers leveraged this global pool of knowledge—AI’s open commons—to avoid reinventing the wheel. They also reportedly drew on open models like Meta’s Llama and Alibaba’s Qwen, integrating ideas from those systems (OpenAI squabbles with DeepSeek: the pot calling the kettle black - Techzine Global) In essence, they stood on the shoulders of giants. Then, to make their model competitive, DeepSeek is believed to have used distillation techniques on closed models: for instance, querying OpenAI’s GPT-4 or other APIs and training their model to mimic those responses (OpenAI squabbles with DeepSeek: the pot calling the kettle black - Techzine Global) It’s a bit like training a student by having them read the answers of the honor-roll student. The result was a model that in some evaluations matched OpenAI’s own, built at reportedly 3% of the cost.

This approach underscores why Chinese AI outfits might favor open-source: it can be a great shortcut to parity. If you lack the raw resources due to, say, U.S. export controls on the latest NVIDIA chips, you compensate with brains over brawn. Indeed, the U.S. ban on advanced AI chips to China (like cutting off NVIDIA A100/H100 GPUs) forced Chinese researchers to get creative. “Because they had to figure out work-arounds, they actually ended up building something a lot more efficient,” observed Aravind Srinivas, CEO of Perplexity (an American AI company) (China Upends the AI Race - CEPA) In other words, constraints bred innovation: Chinese teams optimized algorithms and sought any open advantages to reduce the need for brute-force compute. It’s a classic tortoise vs. hare scenario—when you can’t sprint with sheer power, you find another path. Open-source tools, freely available, became that path.

China’s embrace of open-source AI also has a strategic dimension. By contributing to and leveraging open models, Chinese AI firms can reduce dependence on Western companies. Why rely on OpenAI’s API (which could be cut off or sanctioned) when you can fine-tune a local open model to similar levels? The Chinese government, for its part, has signaled support for domestically developed AI and likely views open-source as a way to bolster self-reliance. There’s a parallel to how China approached other tech domains: often starting by importing or copying foreign tech, then rapidly innovating and localizing it (think smartphones, drones, etc.). In AI, open-source provides a leg up in that process.

DeepSeek’s model, notably, is fully open-source and openly available on GitHub (China Upends the AI Race - CEPA) This meant Western companies and researchers could scrutinize and even use it. Some Western entities did just that—IBM, for instance, integrated DeepSeek-R1 into its WatsonX.ai platform (OpenAI squabbles with DeepSeek: the pot calling the kettle black - Techzine Global) This cross-pollination is fascinating: a Chinese open model, likely trained with an assist from a U.S. closed model, ends up being deployed by U.S. companies. It blurs the lines of the AI competition narrative. DeepSeek also came with caveats: the unmodified model reportedly adhered to Chinese government guidelines (e.g. avoiding politically sensitive topics, following the “One China” stance on Taiwan) (OpenAI squabbles with DeepSeek: the pot calling the kettle black - Techzine Global) Open-source developers outside China quickly worked to remove those biases, creating derivatives that were more neutral (OpenAI squabbles with DeepSeek: the pot calling the kettle black - Techzine Global) Here we see another beauty of open-source: adaptation. If a model has unwanted features (be they political bias or safety filters), the community can tweak it. That’s impossible with a locked-down API model.

From a strategic view, China’s open-source AI surge represented by DeepSeek is a signal: the race won’t be won solely by those with the biggest budgets. As one analysis put it, “second movers” like DeepSeek can catch up by replicating and optimizing the advancements of pioneers at a fraction of the cost (China Upends the AI Race - CEPA) This flips the script on the AI “arms race.” The U.S. strategy of pouring billions into ever-bigger models and massive data centers (we’ll get to Project StarGate next) assumes that scale is the decisive factor. DeepSeek demonstrated that efficiency and openness can be a formidable counter-strategy. Of course, questions remain about how far this can go—was DeepSeek’s low-cost claim slightly exaggerated? (Some analysts doubt they truly matched GPT-4 quality at such low expense (China Upends the AI Race - CEPA) ) Regardless, the genie is out of the bottle: open-source AI is now part of China’s playbook, both out of necessity and opportunity.

The AI Ecosystem Spectrum: It’s Not All or Nothing

It’s tempting to frame the AI development debate as a zero-sum duel: open-source versus closed-source, winner takes all. In reality, the landscape is much more nuanced. There’s a broad spectrum of approaches, and each has its role. On one end, we have fully open efforts like AI2’s models or Stable Diffusion, where code, model weights, and training data are public. On the other end, we have tightly closed systems like OpenAI’s GPT-4 or Anthropic’s Claude, where only an API gives you access and the recipe is a closely guarded secret. In between lies a range of hybrid models—Meta’s Llama 2, for example, is open-weight (anyone can download the model) but not fully “open-source” by strict definition since its data and some usage are restricted (Meta’s Open Source Llama Upsets the AI Horse Race | WIRED) Google often publishes research (open knowledge) but keeps the weights closed. Many startups open-source a base model but offer a commercial hosted version with proprietary enhancements.

What’s clear is that no single approach has “won”. Despite bold claims from partisans (“open-source will inevitably win because community iteration is faster!” or, conversely, “closed models will always outperform because they have more resources!”), the truth is more complex. Open models have made astonishing progress—within a year of GPT-4’s release, open communities had models like Llama 2, Mistral, and others that, while not beating GPT-4 outright, come impressively close on many tasks. In specialized domains or fine-tuned narrow tasks, open models can even excel beyond big general models. For instance, an open medical LLM fine-tuned on healthcare data might outperform a generic closed model on medical queries, simply due to domain specialization. As Hugging Face’s Chief Science Officer Thomas Wolf noted, the gap between closed and open AI is “much smaller than we thought,” and both forms of models are likely to remain “critical to the future of the industry” in different ways (Hugging Face co-founder discusses closed-source, open-source AI ...) (Both open and closed source models essential to AI race - CNBC)

On the other hand, the frontier of absolute capabilities still seems to be led by closed efforts. GPT-4, DeepMind’s Gemini (if reports hold true), and similarly massive models are currently proprietary. They often set the benchmark in raw performance, at least until open counterparts catch up. Frontier development demands enormous capital and risk tolerance—something easier to muster in a well-funded company with potential profit payoffs than in a decentralized volunteer community. Frontier models are largely dominated by closed-source offerings from well-funded players (Mapping the future of open vs. closed AI development - CB Insights) but those same players rely on open research. It’s telling that even OpenAI, while keeping models closed, continues to publish research on topics like AI alignment openly. They recognize that some level of open science is necessary to get feedback and advance the field collectively.

It’s also important to parse what we mean by open. Open-source isn’t binary; there are degrees of openness. A model could have open code but not open weights (meaning others can see how it works, but not actually run the exact trained model). Or it might have open weights but with a restrictive license (as Meta did initially with LLaMA’s research license). The Open Source Initiative (OSI) even had to clarify definitions because companies were calling models “open” that didn’t meet traditional open-source criteria (The New Definition of Open-Source Disqualifies Many AI Models ...) There’s now talk of “open-weight” vs. “fully open” models. Meanwhile, closed-source doesn’t always mean un-collaborative—some closed models are accessible to academics via partnerships, for example.

What’s becoming evident is that the AI ecosystem likely needs both open and closed approaches. Open-source fosters transparency, community vetting, and faster dissemination of ideas. Closed development, fueled by big investments, can push the envelope in ways a looser consortium might struggle to coordinate (think of training a $100M model—hard to crowdfund that!). Moreover, having both paths creates a healthy competitive check. If closed actors get too ahead but start slacking or acting against user interests, open models can rise to offer an alternative (as we saw with Stable Diffusion challenging proprietary image generators). Conversely, if open models plateau, a proprietary breakthrough can jolt the field forward, which might eventually filter back into open circles.

The key is recognizing the interplay. Tech companies are indeed “mixed in their embrace of open AI” (Ai2: The Pioneering Open AI Research House Paul Allen Built) Many play both sides: Meta publishes open models but surely has internal secret projects too; Google holds back models but open-sources lots of code; startups often open-source a version of their tech for credibility while keeping a premium version closed for revenue. Rather than dogmatically insisting “open-source will win” or “closed source is the only safe way,” it’s more productive to see how each approach can mitigate the other’s downsides. Open efforts can keep the big players honest and ensure knowledge spreads. Closed efforts can ensure funding for huge endeavors and be held accountable (and perhaps pressured to open more) by public scrutiny.

In short, the future of AI is likely hybrid. We’ll see a thriving open ecosystem coexisting with powerful proprietary systems. Users and developers will choose depending on their needs: sometimes you want full control and transparency (open model), other times you just need the absolute best API or a model with stringent safety filters (likely closed). And ideas will continue to flow between the two worlds—research papers, talented engineers hopping between open projects and tech firms, maybe even shared governance frameworks. The lesson from the last few years is that AI’s progress has been a team effort: academics, big companies, startups, and independent tinkerers all contributing. A spectrum of approaches, rather than one ideology, is what drives this field forward.

The Role of Influencers and Interests: Reading Between the Lines

Amid this debate, it’s important to consider who is advocating for what—and why. Many of the loudest voices in the open vs. closed discourse have skin in the game. They may genuinely believe their message, but they’re also often talking their book, to use a finance analogy. Let’s unpack a few examples:

On the open-source side, we’ve already discussed Hugging Face’s CEO Clément Delangue. He is an evangelist for open AI, frequently saying things like “we need to make sure AI is not developed in a silo” and that open source is crucial for AI’s future (How to use Open Source for Enterprise AI) It’s worth noting that Hugging Face’s entire business rides on the presumption that AI development will be largely open and decentralized. If tomorrow every major model went closed and no new open models emerged, Hugging Face would be in trouble. So Delangue’s zeal is both philosophical and pragmatic. This doesn’t invalidate his points—indeed, many agree with the ideals he espouses—but it explains the intensity. Similarly, Hugging Face’s Chief Science Officer Thomas Wolf often provides a balanced take, but even he stands to benefit if the open community is seen as vital.

Then there are folks like Emad Mostaque, CEO of Stability AI (the company behind Stable Diffusion). He famously positions himself as the anti-OpenAI, championing open models for image generation and beyond. Stability AI’s investors poured money in on the thesis that open-source AI can challenge incumbents. So when Emad proclaims the superiority or inevitability of open models, part of that is rallying the community and, frankly, marketing. It’s akin to a rebel startup branding itself against the industry leader. Stability’s credibility comes from being “the open alternative.” Of course, Emad also has ideological reasons—he often speaks about AI alignment with society and not letting a few corporations control AI. But again, business and belief are entwined.

Investors in open-source startups are another group to watch. Venture capitalists who fund companies like Mistral AI (a startup explicitly building open LLMs) or EleutherAI’s spin-offs will loudly tout how open models will disrupt Big Tech. Some genuinely see this as the future; others know that hyping open-source can become a self-fulfilling prophecy that boosts their investment’s value. If enough people believe “open will win,” talent and users flow to open projects, increasing that outcome’s likelihood. It’s a narrative game as much as a technological one.

On the flip side, those in the closed-source camp also push narratives. OpenAI’s leaders have often stressed the safety angle for why they keep models closed. Sam Altman and others talk about the potential misuse of AI if anyone can download a super-powerful model, implicitly suggesting that responsible companies should keep tight control. Certainly, there are genuine safety concerns with AI, but one must note it conveniently aligns with their commercial interests to say “trust us to handle this model for you, don’t ask to peek under the hood.” When OpenAI or Anthropic executives warn of catastrophic risks from unrestrained AI, they may truly be concerned—but it also serves as a rationale for their closed approach.

Even national security and government figures enter this influence game. For instance, if a government official argues that open-source AI could help adversaries and thus should be limited, that stance will favor domestic companies that are leading but closed. Conversely, others in government might argue the U.S. should embrace open models to “win hearts and minds” globally and out-innovate closed regimes. In each case, think about who benefits: Is it Big Tech lobbying for protection? Is it an open-source consortium seeking funding and legitimacy?

The point here is not to be cynical to the point of dismissal, but to be a critical listener. Key figures in the AI debate often have biases shaped by their roles. Hugging Face’s founder will naturally see the world through an open-source lens (and maybe downplay the shortcomings of open models). An OpenAI executive will highlight dangers of open release (while perhaps glossing over how much their own success owed to open research). An academic might champion open science out of principle—and also because academia depends on open exchange by definition.

So how do we navigate this? The best approach is to weigh arguments on their merits but remain aware of context. If someone says “Model X must remain closed for safety,” consider the safety issue and remember that the person saying it might lose a competitive edge if X were open. If another says “open-source will overtake closed within a year,” check the evidence of progress and recall that the speaker might gain from that outcome (or from you believing it now). It’s a classic case of trust but verify.

Interestingly, some individuals have shifted positions as their circumstances changed. Elon Musk, for example, helped found OpenAI with an open ethos, later criticized OpenAI for becoming closed and for-profit, and now is starting his own AI venture (xAI) where it’s yet unclear how “open” it will be. His critiques of OpenAI might be part principled and part competitive (since he’s now re-entering the race). These evolutions show that attitudes toward open vs closed can be malleable—and often follow the incentives.

In summary, the AI community has its share of partisans and true believers on both sides. Their commentary enriches the discussion but also frames it in ways favorable to their interests. As consumers of their insights (or as policymakers hearing testimony, etc.), we do well to appreciate their expertise while taking into account why they might be pushing one vision of AI’s future. The reality likely lies somewhere between the utopias and doomsdays they sometimes portray.

Future of Compute Scaling: Arms Races, StarGates, and New Limits

No analysis of open vs. closed AI can ignore the elephant in the room: compute. The scale of AI models is increasingly determined by how much computing power and specialized hardware can be thrown at training and running them. This raises tough questions: Who can afford to train the next GPT-5 or a true multi-modal GPT-6? Will it be a closed clique of tech giants with deep pockets, or will new approaches make such models accessible to the wider community?

As of now, the trend has been toward fewer, bigger players at the bleeding edge. The cost and complexity of training frontier models have skyrocketed. OpenAI reportedly spent over $100 million training GPT-4. Google’s Gemini is said to involve dozens of models and possibly even more compute. These projects require tens of thousands of cutting-edge GPUs, vast quantities of electricity, and expert teams to manage. It’s not the sort of thing a scrappy open-source team can easily crowdfund or coordinate on a volunteer basis. This has led to fears of an AI oligopoly: that only those with massive compute (and thus likely corporations or governments) will be able to push AI to its next stages.

Enter initiatives like Project StarGate. In late January 2025, news broke of a colossal AI infrastructure project dubbed StarGate, unveiled in dramatic fashion at the White House. Backed by OpenAI, SoftBank, and Oracle, StarGate plans to invest up to $500 billion over four years to build gigantic AI-dedicated data centers across the United States (OpenAI, SoftBank, and Oracle Just Revealed “Stargate,”A $500B AI Infrastructure Moonshot) The initial phase alone involves $100 billion to construct a mega-cluster in Texas, featuring hundreds of thousands of top-tier NVIDIA GPUs and consuming gigawatts of power (OpenAI, SoftBank, and Oracle Just Revealed “Stargate,”A $500B AI Infrastructure Moonshot) The endeavor is so grand in scale that its budget rivals that of the Apollo space program (after inflation) (OpenAI, SoftBank, and Oracle Just Revealed “Stargate,”A $500B AI Infrastructure Moonshot) The message is clear: to stay ahead in the AI race, some in the U.S. are willing to spend astronomical sums to ensure they have the biggest, fastest computers on the planet dedicated to AI research.

From one angle, this is the logical culmination of the closed-source, big-money path: treat AI advancement like an arms race or a space race, requiring government-level funding and secrecy. If StarGate succeeds, OpenAI (and its partners) would have an unmatched compute advantage, entrenched behind high walls. This could produce incredible models—perhaps the first systems approaching Artificial General Intelligence (AGI) if you believe the hype—but it might also concentrate AI power further into a few hands. It’s notable that StarGate has political backing; even former President Trump (now re-elected in 2024, as implied by some reports) touted it as a bid for “U.S. dominance in AI” (The Stargate Project: Trump Touts $500 Billion Bid For AI Dominance) Nationalistic competition could drive more of these efforts, with China, Europe, and others responding in kind (though few can match half a trillion dollars in a single project).

Yet, as we saw with DeepSeek, the future isn’t guaranteed to be owned by the biggest spenders. The law of diminishing returns may kick in; just pouring more compute might yield smaller and smaller improvements unless accompanied by algorithmic breakthroughs. Open-source communities are actively exploring efficient scaling—how to get more bang for the buck from smaller models, clever training tricks, and better data. In fact, one could argue the StarGate approach and the DeepSeek approach represent two diverging strategies: Brute-force scale vs. Efficient ingenuity. The next wave of AI advancement might come from either, or a fusion of both.

Compute scaling also faces physical and geopolitical limits. Chip export controls are one example: the U.S. restricting chip sales to China led China to find new ways (and possibly spurred them to develop their own chip industry more fervently, or to optimize software as discussed). If we imagine a future where multiple nations sharply limit who gets the latest AI chips, we might see parallel AI ecosystems: each major region with its own compute and possibly its own models, open within that realm but closed across borders. The free flow of AI knowledge could be hampered if hardware becomes the chokepoint. Already, there’s concern that if only American companies have cutting-edge chips, then only American-led projects will lead in AI. On the flip side, if a country like China can’t get enough advanced chips, it might devote huge resources to alternative strategies (like new chip architectures, optical computing, or massively distributed networks of smaller chips) to break the dependency.

The future of compute will shape the open vs closed dynamic in unexpected ways. If compute remains king and costs keep climbing, we might see a reinforcement of closed dominance—only the richest consortiums (think Big Tech + governments) can train the top models, and they likely keep them proprietary to justify the cost. Open models might then focus on the second tier: competitive but one generation behind, or specialized to niches where ultra-scale isn’t needed. Alternatively, if new algorithms or technologies (like neuromorphic chips or algorithmic breakthroughs in efficiency) dramatically lower the cost of achieving a given level of intelligence, it could empower open efforts to stay neck-and-neck. For instance, a breakthrough that makes a 13B parameter model perform as well as today’s 100B model would effectively hand open-source community a cheat code to compete without needing a StarGate.

One thing is certain: collaboration on infrastructure is growing. Even OpenAI teaming with Oracle and SoftBank is notable—AI compute is so expensive that unusual alliances are forming. We might see open consortia pooling resources too. There’s precedent in scientific research: CERN (the particle physics lab) is a massive global collaboration because no single entity could afford the Large Hadron Collider. Could AI see something analogous, say an “OpenAI (not the company) Collider” where governments and companies jointly fund an open research compute facility? It’s not outlandish—especially if the public starts demanding more open science in AI to complement the corporate projects.

Lastly, we should mention energy and hardware innovation. The hunger for compute is also a hunger for energy (powering data centers) and for better chips. A lot of smart minds are now working on next-gen AI chips. Some are for efficiency (to make training cheaper, which would help open labs), others for sheer power (to enable that next giant model). Initiatives like StarGate will push chip makers like NVIDIA (and rivals like AMD, Intel, and startups like Cerebras) to up their game. In a positive feedback loop, that could yield more powerful but also eventually more accessible hardware. Think of how GPUs themselves became commodity hardware over time for AI enthusiasts—what was once the preserve of a lab is now in a gamer’s desktop. Could the same happen with today’s $10k AI accelerators? Possibly, especially if demand scales.

In summary, the compute dimension adds both fuel and friction to the open vs closed debate. It fuels it by raising the stakes—billions on the table make companies guard their work more fiercely, and nations treat AI leadership as strategic. But it also adds friction in the sense that raw power alone might not guarantee victory; efficiency and creativity matter too. Open-source thrives on creativity and distributed problem-solving, while closed efforts excel at marshaling resources. The next phase of AI advancement will likely be decided by a mix of these: who can secure the needed resources, and who can come up with the smartest ways to use them. As Project StarGate breaks ground and DeepSeek’s of the world pop up in response, one can’t help but see echoes of the space race: one side building bigger rockets, the other plotting clever orbits. In the end, humanity did reach the moon—and it was a combination of state-driven effort and openly shared scientific knowledge that got us there. Perhaps the journey to advanced AI will play out in a similar fashion.

Conclusion: Collaboration over Dogma

The open vs. closed debate in AI is often framed combatively, but this deep dive reveals a more symbiotic reality. OpenAI guards its secrets now, but built its success on openly published breakthroughs (from academia and its own earlier openness). Meta opens its models to gain strategic advantage and community goodwill, while Google holds back to protect its empire and users—yet all three exchange researchers and ideas through conferences and papers. Nonprofits like AI2 push entirely open AI for the common good, influencing even the closed players to be more transparent. And upstarts like DeepSeek prove that open knowledge can be leveraged to challenge the giants, keeping everyone on their toes.

In the end, openness and protectiveness in AI exist in a dynamic equilibrium. Too much secrecy and the field risks stagnation or public backlash; too much openness and companies fear losing incentive to invest or worry about misuse. The sweet spot is ever-moving. What’s certain is that AI is now a cornerstone of societal progress (and geopolitical competition), so the stakes for getting this balance right are high.

For those of us watching (and contributing in our own ways), it’s important not to treat this as a tribal war where one side must vanquish the other. Instead, the most productive mindset is: how do we get the best of both worlds? How do we encourage the open sharing of scientific knowledge and broad access to AI’s benefits, while still motivating the massive investments and responsible safeguards needed to bring truly powerful AI into being? The debate will continue, but hopefully with a bit more nuance—as a creative tension that, if managed well, can drive AI to new heights and ensure those heights are accessible to all.

In the grand scheme, AI is a human endeavor, and humanity benefits most when we balance competition with collaboration. Open-source and closed-source approaches each have their role in the AI cosmos. And just as planets need both light and gravity to orbit stably, the AI world may need both openness and restraint to move forward safely and swiftly. The conversation is far from over, but with critical eyes on the influencers and a focus on pragmatic solutions, we can chart a path that avoids the extremes of dogma. After all, the ultimate goal is not to make one ideology “win”, but to make sure we all win with AI—whether its underlying code is visible or not.

Appendix: Prompt for OpenAI Deep Research

Help me write a balanced report on the debate between open source and closed source development of AI. Proponents of the closed approach include OpenAI, Anthropic, and Google. Proponents of the open approach include Meta, Ai2, Hugging Face, Mistral, and DeepSeek. Specifically, the report should cover the following:

- OpenAI has been a pioneer of AI. They were the first to recognize the potential of compute scaling, starting with GPT-3. In order to go after this, they need to raise a lot of capital. Since they don't have an existing cash cow (unlike Google and Meta), they need to protect their tech advances to maintain a lead over competition. Otherwise nobody would fund them.

- Meta, on the other hand, have a hugely profitable business. Unlike cloud hyperscalers like Google, Amazon, and Microsoft, they naturally want to support an open, collaborative effort to develop AI that benefits them the same way the OSS has had.

- Help me understand why Google has not open sourced their best models. Include that in the report.

- Hugging Face's DNA and product strategy and market positioning are fundamentally open-source, so they naturally take this stand.

- Ai2, as a non-profit research organization, naturally embraces the open approach all the way to sharing data, recipes, training logs, techniques.

- Chinese companies such as DeepSeek have their own motivation in open sourcing their models. Research and summarize them. - It's important to recognize the whole ecosystem with different players adopting the full spectrum from completely open to close. It's not good to be dogmatic/idealistic (e.g. open source is the only way, open source will win, etc.).

- There are many prominent social influencers out there. Their opinions and beliefs are mostly aligned with their self interests. For example, founders of Hugging Face or investors in Mistral are naturally pro open source. We should know their inherent biases and de-bias their opinions accordingly.

- If we want to make a bet on the next phase of compute scaling (which should get the benefit of doubt given its success so far), then it's important to think through how to fund such efforts from a capex standpoint. Discuss project StarGate and the need for chip export control in this context.