Reactions from ranked influencers
3 posts
Investments in high-quality RL environments are a great leveler because they can mitigate or remove the american tech giants' massive data advantage As in all things RL, this environment creation isn't the most scalable but as Delip points out if you have a huge workforce...
Some folks are dismissive about this lead being “just in the frontend coding arena” or that the win claims are propaganda/overreach. I wish they realized K3’s frontend coding capabilities are the result of investments in high-quality, frontend-specific RL environments and trajectories. From here to AI dominance in science, cybersecurity, healthcare, or any other domain is scaling — environments, talent, and compute. If there is one thing we can be assured about China is they can scale with efficiency. They are amazing hill-climbers at people, org, and national levels. As a proxy metric, around 13 million students take gaokao *every year*. They will happily build environments when asked by the right folks. Their national gen AI adoption rate was around 42% at 600M users. That’s a lot of trajectories to mine. With state level investments (see my tweet about Xi’s speech yesterday), exceptional talent, and grindmax culture, we are looking at at most a 6 month timeline for all domain parity with the best of west, if not dominance.
Some folks are dismissive about this lead being “just in the frontend coding arena” or that the win claims are propaganda/overreach. I wish they realized K3’s frontend coding capabilities are the result of investments in high-quality, frontend-specific RL environments and trajectories. From here to AI dominance in science, cybersecurity, healthcare, or any other domain is scaling — environments, talent, and compute. If there is one thing we can be assured about China is they can scale with efficiency. They are amazing hill-climbers at people, org, and national levels. As a proxy metric, around 13 million students take gaokao *every year*. They will happily build environments when asked by the right folks. Their national gen AI adoption rate was around 42% at 600M users. That’s a lot of trajectories to mine. With state level investments (see my tweet about Xi’s speech yesterday), exceptional talent, and grindmax culture, we are looking at at most a 6 month timeline for all domain parity with the best of west, if not dominance.
CHINA’S KIMI K3 HAS SURPASSED ALL AMERICAN MODELS IN FRONT-END CODING WHILE BEING SMALLER THAN MOST CLOSED-SOURCE FRONTIER MODELS. Great work by the @Kimi_Moonshot team.
@deliprao I’d say much of environment and synth data gen is being automated and humans are no longer the direct authors. This is fundamental to scaling. Even with China’s higher capacity (and ignoring language issue), it is no longer possible to scale env and synth data only with humans.