In May 2023, a memo from an unnamed Google AI engineer leaked and quickly made the rounds of the tech newsophere, racking up thousands of comments on Hacker News. The memo was titled “ We Have No Moat…And Neither Does OpenAI”. The gist of the post was this — stop focusing on building the next giant model and start working with the open source community because they will beat us in the end.
“We’ve done a lot of looking over our shoulders at OpenAI. Who will cross the next milestone? What will the next move be? But the uncomfortable truth is we aren’t positioned to win this arms race and neither is OpenAI. While we’ve been squabbling, a third faction has been quietly eating our lunch. I’m talking, of course, about open source. Plainly put, they are lapping us. Things we consider “major open problems” are solved and in people’s hands today.“
This post was initially in response to the release of the first capable Llama models by Meta, which quickly yielded numerous offshoots tuned for myriad specific purposes. Back then, comparable open source models tended to lag closed source models on a variety of fronts, and required a number of months to reach parity. However, the handwriting was on the wall even then. Today, in the age of Instant AI models and DeepSeek, that slight time-to-market advantage has shrunk to 24 hours. Open source AI is not only winning but it’s moving so fast that Sam Altman has publicly admitted he may need to pursue a more rigorous open source strategy (My calendar is open for you, Sam!).
When I wrote last week about DeepSeek and why open source AI will win both in the short and long run, I predicted that the methods DeepSeek highlighted would lead to a wave of rapid innovation in AI models. Even then, companies were working on more reliable and commercially viable versions of DeepSeek’s R1 foundational model which sent stocks into a nosedive. Then, OpenAI released its own advanced research offering, Deep Research, a few days later. This combined an LLM (which can be selected from the current list of LLMs provided by OpenAI, 4o, o1, o3, etc.) and an “agentic framework” to instruct the LLM how to use tools like web searches or how to think through organizing its process into logical steps and to use tools like web search and organize its actions in steps.
It was designed to perform advanced reasoning, function like a research analyst, and deliver complex analyses on projects covering almost any topic. AI experts like Ethan Mollick raved about it, and many others online are singing its praises as the best general agentic tool they have seen.
A day later, HuggingFace released an open source version of Deep Research that, while not yet comparable, was pretty close. (HuggingFace has also made progress towards recreating R1’s missing pieces, such as dataset, and training code). Deep Research undoubtedly benefitted from the data, foundational model and training of DeepSeek and its R1 model in particular. Which is exactly why open source AI is so nimble and powerful. (Check out what Nathan Lambert of AI2 has to say about this). There are no moats by design, and it floats all boats by design.
Meanwhile, in the community, five other open source Deep Research competitors emerged (dzhng, assafelovic, nickscamara, jina-ai and mshumer — h/t to HuggingFace for finding them). By the time you read this, it’s entirely likely more AI developers have published newer and improved versions. The fast replication of capabilities surprised many, but not me. Open source innovation has moved significantly faster than closed source innovation for decades now. We’ve witnessed this at the Linux Foundation as open source software has eaten the software world. To be clear, the open source software didn’t do anything itself. It’s just source code. The world’s best developers, contributing and collaborating in an open, scientific peer review-esque model produced better software that many other developers chose to use because the open source software was just better than alternatives.
The bottom line of all this? Open source AI is accelerating. The ability of the community to quickly match any new development is clear validation of the velocity and power of open source AI, driven by the brains of a massive and growing community of talent all over the world. What’s more, keeping any AI innovation locked in a bottle is going to be incredibly challenging. Advances in model distillation—a process by which a larger, complex model is used to train a smaller, more efficient version—enable researchers to dissect publicly available AI systems more effectively. Through distillation, key training insights can be extracted, making it feasible to develop comparable or even more efficient models. This capability not only accelerates research and development but also broadens the ecosystem of innovation by allowing a wider community to experiment with and improve upon existing models. DeepSeek used distillation to enhance the reasoning capabilities of the Qwen and LLama series of models, effectively supercharging their performance on reasoning tasks.
Almost on cue, a paper dropped from Stanford researchers demonstrating a simple way to transform an open LLM into a reasoning model, using only 26 minutes of training time and costing $30 (obviously, the cost of the base model was a lot more). The secret trick? Feeding it 1,000 well-structured reasoning examples and extending its thinking process. Using this simple (and now publicly known) method, they were able to nearly match the performance of GPT o1 on match problems.
Because foundational AI capabilities are accelerating so quickly in open source, the value will move up the stack and flow to providers who can deliver the best security, ease of use, and developer and consumer experiences. This should be viewed as an amazing victory. Only a year ago, many in the technology realm were fearful that the world was bifurcating into GPU richies and GPU poors — those who had access to powerful AI and those who didn’t. We can already see that’s a moot point and that’s important. Democratic access to AI capabilities across countries, companies of all sizes and NGOs will be a fundamental building block of innovation for many years to come, across many fields. Open wins in technologies that drive a shared benefit. Always. Pull down that drawbridge because the tide is rising.