Yeah. Open source is to gain market share, customer goodwill and visibility. No need to keep some mythical good model as a secret. Like Anthropic keeps Mythos secret is because it is simply too inefficient and it just does not have the compute to ship it. That makes it useless.
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Agree. The last thing you want an adversary to know or suspect is the peak of your capabilities. Then they start to develop countermeasures or ways to thwart it. They also can anticipate what you are capable of. At the same time, releasing near peer models for public use provides a source of revenue and things the public and companies can develop solutions around (not to mention it keeps revenue away from US firms). The US and other countries all hide their best capabilities and technologies from public view as much as they can.
But give
Why not? In fact highly probably they do. These are models are trained so they can have something similar to K3 or better but specifically trained for cybersecurity or other military reasons is not out the question.
Qwen big models has always be closed sourced, they only open sourced the smaller models and they are used a lot in the research community and that is how Qwen got their fame. What Qwen is losing right now is the diffusion race, the only reason why they are still relevant is because two models they open sourced a few months ago. They are getting beaten by almost every AI Chinese lab. Lets hope they turn things around because their models are important for the research community even if they keep their big model closed.
Dude there is zero law prohibiting Chinese companies from open sourcing their advanced models publicly. This are private companies not SOE. So they are run like any other private companies globally. Stop believing China is run by some authoritarian government who controls every single details of what’s private companies do outside of the normal laws/rules. The ones you mentioned were prohibited by the government, just like US government prohibits some sensitive technology from being sold .You might as well believe the US government also has some shadow advanced models they don’t allow their private companies from releasing ( well at least they did stop Anthropic from releasing mythos at some point, so they might have others as well?).
Believing there is some shadow models the Chinese government collects and prohibits from its private AI companies from being open sourced is a fallacy. Just like it’s not the Chinese government who told Alibaba to make her latest Ada aced models closed sourced. It was a business decision by its board of directors/CEOs . If tomorrow Kimi or GLM decide to close sourced their latest models it will be their business decisions not the CCPs. So we should stop acting like private Chinese companies are run by the CCP , that’s western propaganda.
People need to realize China is not AI pilled to the extant US is. They are not a declining power. US is basically clinging onto AI as their savior that will allow them keep their dominance in global tech space. So, a lot of US fixation on AI is overhyping it and then believing their own hype.
China doesn't treat AI as a special magical technology. They treat it as any other software tech and China does not consider software to be strategic. They consider hardware to be strategic.
So, they r more considering closing the AI chip gap and chip gap in general to be the main strategic bottleneck they face.
So, no, I don't think there is some secret AI model that China is keeping for themselves. The AI companies are competing desperately to keep their relevance and Chinese govt is just watching from the sidelines to see how it plays out.
so this is actually pretty good job of a model's SWE capabilities, which again translates well to cyber security and attacks and such. I tend to think that Kimi k3 is a pretty good jump over even GLM-5.2 here because just the size of the base model allows it to do so. 5.2 was quite impressive in bringing that much power out of a base model of 750m parameters, but I'd imagine the so called 5.5 will be significantly larger.
China has never been one to be laissez faire on security, so if it has security implications, China will consider it strategic.
Granted Dario probably overhyped the security aspect, but this is where another Chinese trait comes in: China has never been one to be reactive when it comes to security.
I don’t understand why it’s such a stretch that China keeps models secret when we know that the U.S. does the same thing.
What kind of security are you talking about?
Political speech security? Yeah, they do care a lot about AI not saying stuff about Chinese politics. And they do control for that heavily.
But cyber security? Thats BS. If AI can find find vulnerabilities in code, it can also fix it. Why would you want to limit power of AI if it can help you find bugs?
Again, as I said, US is overhyping AI as some magical tech and trying portray them being on top once again.
China doesn't believe in that and is fine to diffuse AI technology
Not necessary, in fact is likely that could add additional vulnerabilities that to the shear code size humans will overlook them.
it could require human supervision to make sure is patched properly. These models are NOT conscious, they don't have nuance. They just mimic human language and they do it really well
Last edited: