Yann LeCun的动态

Framing DeepSeek’s success as just a “win for open source” is dangerously naive. Yes, they leveraged open research, just like every major AI player does. But let’s not ignore the bigger picture: DeepSeek isn’t just another open-source model—it’s a strategic play by China to break the West’s AI monopoly while proving that cutting-edge AI can be built at a fraction of the cost. The West has been comfortable with the idea that AI leadership is theirs to lose. DeepSeek is a wake-up call: innovation isn’t just happening in Silicon Valley anymore, and China isn’t just "building on top"—they are redefining efficiency, cost, and scale. If this was just about open-source collaboration, why did the markets react like they did? Why is Nvidia bleeding? The real story isn’t open source vs. proprietary—it’s about who moves faster, smarter, and with a long-term vision. And right now, the West is still debating while China is delivering.

With respect, both are true. In the US, companies like Meta invest heavily into AI and open source, and kudos for that, but in the US, the government has done nothing during this AI renaissance. Playing spectator while adversaries surpass the US in AI is a clear failure. In China, the CCP has invested heavily in R&D, as they should, and in this way supports “open source”, but closes the door onto how it was achieved or any third party participation in the process. The models are full of blatant CCP propaganda. If you don’t belive me, just query their models which have built in denial of the Tiananmen Sq massacre (model v3), or tell you that the answer causes harm (model r1). Or ask either model what is the greatest nation in the World. In my opinion, DeepSeek’s $5 mil scratch training methodology, if true, is one of thr most impressive achievements. It will save billions in energy, hardware cost, and time in advancing AI. The dataset, is also certainly most novel, and how they collected it, that should make quite the story - which leads me to conclude, I am skeptical, not without knowing how it was achieved. Is any of what we are doing good for society? We should all question this every day, and strive to keep AI net positive.

This benchmarking comparison has nothing to do with US vs China AI-superpowers. Open source is open source - the entire planet has access to it, including big tech teams from Meta, Google, OpenAI etc... This has only proven that some talented and smart individuals exist and are getting results with what they had available. The beauty of having great performance from DeepSeek is having an optimization mindset; seeking performance with little resources (Energy being on this list of resources that we need to optimize here!!!). Time for BigTech to start consuming less energy while getting good performance, since smaller teams have proven that to be possible. PS: "They came up with new ideas and built them on top of other people's work." This is literally what research is supposed to be, building on top what is existing, to get better results. At some point, on va pas réinventer la roue :)

DeepSeek claims they managed to train their DeepSeek-V3 model with an expenditure of ~ $5.6M in compute costs. If that's entirely true, then why on Friday the Bank Of China (BOC) announced a ~ USD 137B during a 5 year period to invest in AI infrastructure? My math isn't mathing here. You think the US didn't knew about this before hand ? You think ScotiaBank, SoftBank, or companies and foreign entities didn't knew about this before committing to a massive investment in the US?

The important lesson is economic. They exploited 100s of billions of dollars of investment in the U.S. and built a competitive model with a few million dollars and ingenuity rather than brute force and scale. Just imagine what could be done if more AI research were focused on efficiency, accuracy and safety rather than scale and theft to protect the interests of the Big Tech oligopoly.

It's also altogether unsurprising given Google's been saying "we have no moat" for nearly a couple of years at this point. As I've written about here at Nscale (https://www.nscale.com/blog/why-open-source-matters), open source is the great equaliser but it does also mean everyone can benefit from the research and methods employed to further accelerate AI research. Being able to do more with less is nearly always a good thing, the corollary?being you can still do more with more.

Sorry, I don't buy the "yeah some quants hacked it together as a side project on some spare compute from mining rigs". They used a frontier model to skip to the front of the open source line and they certainly had a bit more compute than they're claiming.

Indeed, a choice also made by Mistral, the French AI Company. And, if I am not mistaken, Deepseek and Mistral also elected for the same technology infrastructure - When Open AI mobilize a “complete and big Brain” - a single large model for each request (requiring more power, hence higher cost), Mistral and Deepseek rely on a “multiple experts" technique which is more energy and cost efficient as it mobilizes only what is necessary without losing accuracy.

But how do you explain the difference in energy consumption, orders of magnitude less than OpenAI? As a big data engineer I'll tell you what explains it: optimizing the processing jobs. I've seen it at MN: amateurs using the wrong storage, not actually distributing the data, don't understand ser/de bottlenecks... And they "throw more money at the problem" (which doesn't help) instead of learning how distributed systems work and utilizing them correctly. This industry is full of imposters who literally have no idea what they're doing, wasting everybody's time and money ?? https://www.dhirubhai.net/pulse/spark-performance-issues-check-your-serde-nira-amit-wwdfe

If llama/DeepSeek are open source, then where is the training data? If you can't reproduce something and can't explore exactly what made it, the 'model source' is not open.