deyiao

Interestingly, Anthropic and voices in China’s open-source AI community seem to describe each other’s lead in AI as the equivalent of Hitler getting the atomic bomb before the Allies. A personal reflection from a DeepSeek kernel developer on watching AI master the craft he loves—and helping build the technology that may replace him. It explores what we stand to lose even if we keep our jobs, and why open-source AI matters.

It's unfortunate that the final few years of the fastest AGI progress happen to coincide with a global surge in nationalism, polarization, and mutual hostility between human groups.

That raises a disturbing failure mode: under intense conflict, an AI could learn that exterminating some groups of humans is a justified or even desirable objective. Once that principle is accepted, the step to concluding that exterminating all humans is justified may become much smaller than we'd like to believe.

I guess the author’s point is that LLMs can’t really learn in real time yet, whereas playing games is basically all about real-time learning. So an LLM can be very good at writing code, but still be terrible at actually playing games.

Personally, I think this is a really hard problem, and it may turn out to be one of the first big walls we hit on the road to AGI.

OpenAI Five doesn’t really know how to play games in general — it only knows how to play Dota.

Humans do not have souls, nor do they possess free will in the traditional sense. What we call “consciousness” is merely a product of evolution, and also a tool shaped by evolution.

In essence, consciousness is a complex information input-output system. When such a system reaches a certain level of complexity, it inevitably generates the concept of “I” as a way to simplify the processing of overwhelming information.

Praise be to AI. In 2025, inspired by AI, I feel that I have finally built a complete and unified worldview.

Are we living in a virtual illusion? Are there higher-dimensional rulers, gods, or immortals in the universe? What exactly are the human soul and consciousness?

I feel that these questions now share a single coherent answer. What I have written here is my answer regarding the soul and consciousness.

They say cars replaced carriages but created drivers, so no net job loss. They say AI will do the same—destroy some jobs, create others. But bro, the automobile wiped out 95% of the world's horses. And this time, what AI is replacing is humans.

Since Claude Code can browse the web, is it fair to think of it as “rewriting and simplifying a compiler originally written in C++ into Rust”?

I believe that if a model can outperform humans in all board/card games, and can autonomously complete all video games, then AGI — or even ASI — has essentially been achieved. We’re still a long way from that.

I don't think so. The author isn't training an LLM, but rather using an LLM to solve a specific problem. This method could also be applied to solve other problems.

Content moderation is incredibly frustrating — it might even be the key reason why Veo2 and even Gemini could ultimately fail. I just want to make some fun videos where my kid plays a superhero, but it keeps failing.

Oh, DeepOpenAI

Now it includes the highly anticipated PTX! Of course, I don’t understand it, but I’ve already click the star and even the fork button, which basically means I’ve mastered it, right? I feel incredibly powerful right now...

Is the PTX that everyone was looking forward to included this time?

I heard their inferencing framework is way lower than typical deployment methods. Can this be verified from that open-source project? How does it stack up against vllm or llama.cpp

But the fact that they were donating huge sums every year even when they were still unknown really says something. If they were purely profit-driven, there’s no way the shareholders would have allowed that.

From what I know, DeepSeek is a small company that made a lot of money from other businesses, which makes their lack of focus on commercial interests feel more genuine. Plus, even back when they were relatively unknown, they had a habit of donating over $100 million annually to charitable causes. That makes their claim of striving for humanity a lot more believable.

I really admire their mindset of striving for the betterment of humanity. There was a time when OpenAI, Anthropic, and even Musk used to talk with that same lofty vision. But now, they've all shifted to competing for national interests instead, which is honestly quite disappointing.

I asked DeepSeek-R1 to write a joke satirizing OpenAI, but I'm not a native English speaker. Could you help me see how good it is?

"Why did OpenAI lobby to close-source the competition? They’re just sealing their ‘open-and-shut case’ with closed-door policies!"

It’s been reported that DeepSeek R1’s coding capabilities exceed GPT-o1-low and nearly match GPT-o1-meduim, quite astonishing.

The sources you've listed don't seem to show any environmental damage. In fact, China has been leading among major nations in environmental protection efforts in recent years. For instance, the destruction of natural forests has been virtually halted, and China has been the most successful country in combating desertification through massive reforestation projects. Ironically, some Western media outlets are now criticizing China for "having too many trees" and reducing desert areas across the globe.

The benchmark results seem unrealistically good, but I'm not sure from which angles I should challenge them.

Praise for Open Source, true open source

I bet I saw the exact same thing about 14 years ago.

Hi, Can you simply summarize some of the useful and interesting ideas in there?

If the findings of this paper hold up, I believe it could pretty much undo a decade of NIST's efforts in post-quantum cryptography. a seismic shift in the world of cryptography.

I'm not a physicist, so how strongly do these results support the existence of gravitons? 70% 90% or 99%?

They underestimate the difficulty of the technology, which does not mean they overestimate the impact that technology can have.

"The MOE architecture uses 20 times the parameters, is this comparison fair? Can it be compared with a single model that also uses 20 times the parameters?"