The Push for America to Open Source AI
13:32 Share

Thank you.

Hello, friends. It has now been a few weeks since the Black Swan event that was the launch of DeepSeek. DeepSeek, of course, is a Chinese company spun out of a hedge fund, no less, whose AI models have recently totally challenged expectations and thoughts around just how far ahead the U.S. actually is.

Now, we've had a lot of chance to talk about different aspects of the DeepSeek story. How part of the reason that their app has been so popular is that whereas OpenAI was giving people the subpar models in their free chat GPT app, DeepSeek was actually giving a reasoning model right there.

There were also some UI innovations based on the way that it exposed its thought during its reasoning. And of course, the biggest part of the debate has been around distillation techniques and how they were able to get this much performance with so little money. And or on the flip side, disbelief that that actually happened. But the part of the conversation that I want to come back to, which I think is potentially the most significant when it comes to shifting the industry, is the idea of the implications for how the United States in specific thinks about open source versus closed source AI.

Now, this has been an interesting and ongoing debate. OpenAI was, of course, called OpenAI, but over the years started to shift its policy. It stopped sharing its research in full, and obviously none of the big models have been open source for some time. A lot of the justification for this was about safety and about it being dangerous to share things like model weights with the wider world because of all the bad people out there who might use them for nefarious purposes. There

There, of course, have been also tons of counterarguments, alongside a lot of assessment of the pragmatics of trying to keep things closed source, and more than a fair bit of skepticism that safety concerns are actually the reason as opposed to competitiveness concerns. I think complicating this fact is that one of the loudest people who has been contra to OpenAI's strategy is Elon Musk, who obviously has a very big axe to grind over there, and who, as Sam Altman has pointed out, hasn't been open sourcing the main Grok models either.

In any case, it feels very much like DeepSeek has shifted the nature of the conversation. We're going to read a quick piece, or rather turn it over to AI to read a quick piece, by former Google CEO Eric Schmidt called Will China's Open Source AI End U.S. Supremacy in the Field? With the advent of DeepSeek, the balance of power between the two nations appears to be shifting. So I'm going to throw it over to an 11 Labs version of myself, and then we will come back and keep discussing. It has become almost a cliche to say that the artificial intelligence landscape is changing fast.

But in recent days, even those on the cutting edge of AI research were taken by surprise by a Chinese company. Last week, the AI company DeepSeek released its R1 reasoning model, which is on a par with OpenAI's O1 and much better than the chat GPT models across a variety of logic tasks, including math and coding.

The cost of running it is also much lower, only about 2% of what OpenAI charges. And on Monday, DeepSeek released Janus Pro, a model small enough to run on your laptop that can generate synthetic images, which it claims outperform OpenAI's DALI 3. DeepSeek's speed of AI innovation is taking the world by storm. What's even more remarkable is that DeepSeek's entire collection of models is open source, which in this case means they have open weights that anyone can reproduce and build on top of.

It's a peculiar moment when a Chinese company becomes the de facto open source leader, while most major American firms, with the exception of Meta, continue to keep their methodologies tightly under wraps. In fact, this is a growing trend for Chinese AI companies, from startups such as Minimax to tech giants such as Alibaba, that are giving developers worldwide free access to their AI models.

Until now, closed-source models such as OpenAI's O3 and Anthropic's Claude 3 Opus were considered the industry standards with the most advanced capabilities. And they were built in the United States. Open-source and Chinese models were thought to be months behind. But DeepSeek's R1 and Janus Pro show just how quickly the tides of technological supremacy can turn. The introduction of these models has roiled stock markets and caused U.S. tech stocks to plunge.

The balance of power now appears to be shifting along two key axes, one between the United States and China, and another between closed and open-source models. Defenders of closed-source models are betting that they can preserve their capability gap by protecting their model weights and training methodologies. Open-source advocates, on the other hand, argue that transparency, allowing others to build on their work, can enable these systems to rapidly catch up with larger closed models.

If the open-source thesis is correct, this would turn the AI ecosystem on its head. Open-source models are generally cheaper to use, so when two equally capable models are available, one open, one closed, the open-source model is likely to gain wider adoption, giving it a strategic advantage. The United States already has the best closed models in the world. To remain competitive, we must also support the development of a vibrant open-source ecosystem.

The race between open and closed-source AI, as well as between the United States and China, does not yet have a clear winner. But there is clearly mounting pressure on America's big tech players if DeepSeat can compete with them using far fewer resources.

Export controls were aimed at choking off China's access to the most advanced computer chips, impeding its ability to keep pace. But in fact, the relative dearth of high-performing chips in China might have pushed the nation's companies and researchers to be more efficient and led them to uncover new methodologies that significantly reduced training costs. For example, DeepSeq demonstrated that large-model training could be made more efficient by bypassing the traditional supervised fine-tuning stage.

They even created R10, a model that omits this step in AI training, to challenge the research community's assumptions about fine-tuning's indispensability. DeepSeek's success has also called into question the importance of pre-training, which involves training ever-larger models that predict the next word based on vast amounts of text. This process requires enormous upfront investment in graphics processing units, GPUs, and data. So much data that OpenAI co-founder Ilya Sutskever recently noted we might soon exhaust all the data available on the internet.

But there is another emerging way to improve models' performance. Introduced with OpenAI's O1 model in December, this approach enables models to perform reasoning through self-reflection, similar to how humans reason, using intermediate steps and self-correction to reach a final answer. The training recipe for this approach had previously been closely guarded by OpenAI. DeepSeek blew the lid off that by publishing a paper detailing how it works, allowing others to implement the process.

DeepSeq even demonstrated that you can do this much more cost-effectively by taking a publicly available base model, such as Meta's LLAMA3, and teaching it to reason through reinforcement learning, a trial-and-error process with human-devised feedback and rewards.

Over time, the models seem to spontaneously learn how to reason, backtrack when they hit dead ends, and explore novel approaches. This method eliminates the need to expensively pre-train a new base model, and its implications for AI innovation are profound. Traditionally, even the top-funded university labs have struggled to contribute to AI research due to computing and data limitations. With DeepSeq's breakthrough, the moat surrounding large, well-funded companies might be shrinking.

It is unlikely that American frontier model companies will change their business models anytime soon, nor is it immediately clear that they should. Open and closed competition will most likely find a natural equilibrium, with a range of different offerings and price points for different users. But DeepSeek's release marks a turning point.

The path forward for American innovation involves not just ramping up open-source development, but also encouraging the sharing of training methodologies and increasing investment in AI research and development. Exemplified by the White House's recent announcement of the Stargate Project, which aims to spend $500 billion on AI infrastructure over the next four years, America's competitive edge has long relied on open science and collaboration across industry, academia, and government.

we should embrace the possibility that open science might once again fuel American dynamism in the age of AI. Today's episode is brought to you by Vanta. Trust isn't just earned, it's demanded. Whether you're a startup founder navigating your first audit or a seasoned security professional scaling your GRC program, proving your commitment to security has never been more critical or more complex. That's where Vanta comes in.

Businesses use Vanta to establish trust by automating compliance needs across over 35 frameworks like SOC 2 and ISO 27001. Centralized security workflows complete questionnaires up to 5x faster and proactively manage vendor risk. Vanta can help you start or scale up your security program by connecting you with auditors and experts to conduct your audit and set up your security program quickly. Plus, with automation and AI throughout the platform, Vanta gives you time back so you can focus on building your company.

Join over 9,000 global companies like Atlassian, Quora, and Factory who use Vantage to manage risk and prove security in real time.

If there is one thing that's clear about AI in 2025, it's that the agents are coming. Vertical agents by industry, horizontal agent platforms, agent-based platforms.

agents per function. If you are running a large enterprise, you will be experimenting with agents next year. And given how new this is, all of us are going to be back in pilot mode.

That's why Superintelligent is offering a new product for the beginning of this year. It's an agent readiness and opportunity audit. Over the course of a couple quick weeks, we dig in with your team to understand what type of agents make sense for you to test, what type of infrastructure support you need to be ready, and to ultimately come away with a set of actionable recommendations that get you prepared to figure out how agents can transform your business.

If you are interested in the agent readiness and opportunity audit, reach out directly to me, nlw at bsuper.ai. Put the word agent in the subject line so I know what you're talking about. And let's have you be a leader in the most dynamic part of the AI market. Hello, AI Daily Brief listeners. Taking a quick break to share some very interesting findings from KPMG's latest AI Quarterly Pulse Survey.

Did you know that 67% of business leaders expect AI to fundamentally transform their businesses within the next two years? And yet, it's not all smooth sailing. The biggest challenges that they face include things like data quality, risk management, and employee adoption. KPMG is at the forefront of helping organizations navigate these hurdles. They're not just talking about AI, they're leading the charge with practical solutions and real-world applications.

For instance, over half of the organizations surveyed are exploring AI agents to handle tasks like administrative duties and call center operations. So if you're looking to stay ahead in the AI game, keep an eye on KPMG. They're not just a part of the conversation, they're helping shape it. Learn more about how KPMG is driving AI innovation at kpmg.com slash US. All right, back to real NLW here. In an almost throwaway line, Schmidt gets at, I think, what has captured so many people's attention with this whole story.

That line is, it's a peculiar moment when a Chinese company becomes the de facto open source leader, while most major American firms, with the exception of Meta, continue to keep their methodologies tightly under wraps.

This is a strange turn of events. It doesn't seem like what should be. And of course, DeepSeek is not free from influence of China in the way you'd expect. It will not engage with certain politically sensitive questions, which is why, of course, many people who have chosen to engage with DeepSeek have done so in versions that are powered by the API that can get around some of those restrictions.

When it comes to the big labs, some have doubled down on their arguments against open source. Czarnik shared an interview with Anthropic CEO Dario Amadei, who said that AI safety evaluations conducted on DeepSeek showed it was the worst performing model they'd ever tested at generating potentially dangerous information. They said it had absolutely no blocks whatsoever against generating this information. Now, Amadei pointed out that he doesn't think that these models are actually dangerous in any way, but that we're on these exponential curves, and so that is a safety consideration.

Then again, some people rejected that position entirely, with Marc Andreessen writing, "'Fear-mongering for regulatory capture and to kneecap open-source AI. The existential threat of open-source AI is the big AI cartel.'" He actually even then quote-tweeted himself, saying, "'The reality is the secrets are out. Everyone knows how to code a transformer, how to RLHF, how to use reinforcement learning for reasoning. There will be thousands of open-source implementations in addition to DeepSeek and LLAMA. There is no putting this back in the box.'"

And whether it's just an acceptance of inevitable reality, a change of opinion on the safety, or competitive pressure, it's not just Andreessen saying this. Sam Altman in a Reddit AMA said that he thought that OpenAI might have been on the wrong side of history with this and needs a new open source strategy. And then just yesterday, news started to come out that suggested that Baidu had announced that it would be open sourcing its future Ernie models. After, as Interconnected Capital's Kevin Hsu put it, being one of the staunchest closed source model makers.

So it definitely feels to me like there are shifting sands here. I think the next big question is going to be seeing how it all plays out in the policy sphere, as that could really shape this discourse as well. For now, though, that is going to do it for today's AI Daily Brief, Long Reads edition. Till next time, peace.