Kimi CEO: AI R & D is entering the age of "AI lead research" and the industry standard "has been challenged"

March 26th news, yesterdayDark Side of the Moon CEO Yang Chik LunSPEAKER AT THE PLENARY SESSION OF THE CHINA-KWAN VILLAGE FORUM ON THE THEME " OPEN SOURCE AI: ACCELERATING THE EXPLORATION OF SMART CAPS " , SYSTEMATIC PRESENTATION Kimi Recent developments in large model scale efficiency and bottom-up structural innovation。

Kimi CEO: AI R & D is entering the age of "AI lead research" and the industry standard "has been challenged"

YANG SHIH-SIK SUMMED UP THE ESSENCE OF THE BIG MODEL AS "TO TURN MORE ENERGY INTO INTELLIGENCE", EMPHASIZING THAT SCALEING IS THE CORE FOUNDATION OF AI DEVELOPMENT OVER THE PAST FEW YEARS。

Kimi, he said, is scaling up efficiency around three dimensions:

TOKEN EFFICIENCY: EXTRACT MORE INTELLIGENCE FROM THE SAME DATA THROUGH A BETTER NETWORK STRUCTURE AND OPTIMIZER

Long context capability: Kimi self-researched Kimi Linear architecture can obtain lower loss functions in longer context to support more complex tasks

Agent Swarm: By moving hundreds of Agents in parallel, mission execution times no longer increase with the complexity index, thus making it possible to "unattainly impossible to achieve " 。

Yang Shizuku also cited slides from the Young Wei Da GTC 2026 Huang In-hoon keynote address that the Kimi series open source model is becoming the global baseline for chip manufacturers and research institutions, and that "if you want to release a new chip, you have to evaluate your performance enhancement through Kimi or other open source models."。

In the judgement of the R & D paradigm, Yang Shizuku describes a clear course of evolution:

THREE YEARS AGO, FOCUSING ON NATURAL INTERNET DATA WITH A SMALL NUMBER OF MANUAL LABELS, LAST YEAR MOVED TO A LARGE-SCALE INTENSIVE LEARNING, AND THIS YEAR WILL BEGIN WITH THE "AI LEAD RESEARCH" PHASE, WHICH WILL PROVIDE AN AUTONOMOUS SYNTHESIS OF MISSIONS AND TRAINING ENVIRONMENTS, DESIGN INCENTIVE FUNCTIONS, AND EXPLORE NEW NETWORK STRUCTURES。

In his view, technologies that have been considered industry standards for the past decade (e.g. Adam Optimizer, Standard Attention) are now in a position to be challenged。

Many things that were previously considered standards can now be challenged。

statement:The content of the source of public various media platforms, if the inclusion of the content violates your rights and interests, please contact the mailbox, this site will be the first time to deal with.
Information

An exposed apple will push Sri independent App against the ChatGPT

2026-3-26 11:31:22

Information

Volcanic engine creation

2026-3-26 11:33:40

Search