New Trends in AI Benchmarks and the Research Ecosystem
연구/벤치마크 | Sat Aug 01 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 4 sources
The DeepSeek V4 Flash benchmark, running the 2.78-trillion-parameter Kimi K3 on consumer hardware, and Chinese researchers expanding onto X are reshaping the AI research landscape.
Analysis
[Artificial Analysis] released the Intelligence Index v4.1 benchmark covering DeepSeek V4 Flash [5][3]
- Measures model intelligence across 9 evaluations
- Includes GDPval-AA v2
- Terminal-Bench v2.1
- SciCode
- Humanity's Last Exam
- GPQA Diamond
- AA-Omniscience Index assesses hallucination and knowledge reliability
- Openness Index rates model openness on a 0–100 scale
[WASTE (Weight-Aware Streaming Tensor Engine)] released an inference engine that runs the 2.78-trillion-parameter Kimi K3 on a consumer laptop
- Runs on a 64GB MacBook Pro using only 29GB of RAM
- Achieves speeds of 0.49–0.54 tok/s
- An embeddable engine written in C with no external runtime dependencies
- Leverages MoE characteristics to stream experts from disk []
[Chinese AI Researcher Community] expanded activity on X, with researchers from Moonshot, DeepSeek, and others joining the platform
- About 30 accounts related to Moonshot AI are active on X
- International branding push following the viral moment of DeepSeek R1
- Filling the public discourse gap left by reduced sharing from OpenAI and Anthropic researchers
- A more favorable environment for technical discussion compared to Chinese platforms []
[MIT Technology Review] published the AI Hype Index tracking the latest AI trends
- 1X's dexterous robot hand demo made waves
- Highlighted controversies including Grok's translation feature and Meta's smart glasses
- Raised concerns about surging Big Tech greenhouse gas emissions
- Noted how the AI boom has boosted bonuses for Korean semiconductor workers []