Reactions from ranked influencers
10 posts
Kimi K3 requires a lot of memory for all those 2.8 trillion parameters (Bloomberg)
Watch the latest WTF: https://www.youtube.com/watch?v=pSUyLfirP8Y
America just had its AI Sputnik moment. Kimi K3 dropped a bombshell: 2.8 trillion parameters, the largest open-weight model ever built headed. straight to number one. Possible built on chips two generations behind NVIDIA's best, under US export controls designed to make exactly this impossible -- K3 is #1 in front-end code and six other domains: marketing, design, data analytics, consumer products, simulations, and content creation. -- The model designed its own chip for its next generation and wrote its own inference kernels. Recursive -- 70% of elite AI researchers are not US citizens. The founder of Moonshot AI earned his PhD at Carnegie Mellon and started a Chinese AI startup one year into his US program. -- Bonsai 27B: a GPT-5 class model now runs entirely on a smartphone at 6 gigabytes.
In my long read on the Kimi K3 situation, I focus on the big picture for where we going next. 0. Chinese labs have 3 of top 8 most intelligent models, and have been climbing the ranking. 1. China’s recommits to open-source AI – showing a different (lower) read on near-term risks 2. Open models as the economic Achilles heel of frontier labs (as @deanwball said), and why this can be good for the ecosystem 3. China’s efficiency advantage is letting them make the best models for less, pointing to them staying competitive 4. A growing ecosystem of frontier, open models is coming soon 5. We are at the very beginning of a long story of frontier open-weight policy, which is extremely complex and something we are not handling well Kimi K3 is really the beginning, or the reset, of all these stories. I expect to continue to unpack them regularly.
Where Kimi K3 puts the balance of power — and bends the trajectory — of the AI ecosystem. https://www.interconnects.ai/p/kimi-k3-the-open-weights-escalation
This isn't rocket science. Frontier AI models are going to get much LARGER. Trillions upon trillions of parameters will require vastly more memory, flash storage, and AI GPU compute. Use your brain.
Kimi K3 requires a lot of memory for all those 2.8 trillion parameters (Bloomberg)