Tech Blog
Event Recap (1) | Stronger Models “Cheat” More: Yiming Du on Data and RL for Coding Agents / 活动回顾(一)|越强的模型越会“作弊”:杜一鸣谈代码智能体的数据与强化学习
On the afternoon of October 17, NICE held an in-person event in Shenzhen, “Agentic Data: From Computer Use to Robot Use.” Part 1 covers Yiming Du's talk on LegoX: why training coding agents needs real, verifiable tasks, four lessons from...
Event Recap, Part 3 | Looped LM Panel, Second Half: Loops and CoT, Expressiveness, MoE, Extrapolation and Speedups / 活动回顾(三)|Looped LM 圆桌(下):Loop 与 CoT、表达力、MoE、外推与加速
The third and final part of the recap of NICE's event "Looped Language Model: Rethinking Model Depth and Parameters": whether looping along depth is the same thing as CoT, whether expressiveness proofs still mean anything, how loops should be combined...
Event Recap, Part 2 | Looped LM Panel, First Half: Bottlenecks, Fair Comparisons and Fixed Points / 活动回顾(二)|Looped LM 圆桌(上):瓶颈在哪、怎么比才公平、要不要学不动点
Part 2 of the recap of NICE's event "Looped Language Model: Rethinking Model Depth and Parameters," covering the first half of the panel: where the bottlenecks in LLM architecture are, what looping actually costs, what makes a fair comparison, and...
Event Recap, Part 1 | Looping Saves Parameters, Not KV Cache: Chunyuan Deng on LT2 and Historical KV Replay / 活动回顾(一)|Loop 省参数,却不省 KV cache:邓淳远谈 LT2 与 Historical KV Replay
NICE hosted an online talk and panel, "Looped Language Model: Rethinking Model Depth and Parameters." Part 1 covers Chunyuan Deng's talk: a loop saves parameters, not KV cache — so LT2 moves looping onto linear attention. Plus Historical KV Replay,...
Some Thoughts After Hearing the Story Behind Prove2Me / 听完Prove2Me背后的故事后的一些想法
Dozens of Claude agents wrote about 13 million lines of Lean in 11 days to produce the first computer-checked proof of Fermat's Last Theorem, on a platform called Prove2Me. Notes from a NICE livestream with Shuze Chen — how the...
AI Is Improving AI Faster Than We Can Check
AI is already helping to build the next generation of AI, with people still in the loop. What limits its speed is verification: agents produce work far faster than people and tests can check it. Five reasons verification is the...
Where AI Self-Improvement (RSI) Stands: 8 Takeaways / AI 自我改进(RSI)走到哪一步了:8 条 take away
On October 3, NICE hosted an online workshop on the mechanisms, evidence and limits of AI self-improvement, with four researchers who work on RSI. The main approaches, the current bottlenecks, how to evaluate it, and what it means for academia...
MatrAIx: The Hardest Part of AI Evaluation / MatrAIx:AI评测中最难的部分
Behind 8.3 billion simulated users, a startup is targeting the part of AI evaluation that benchmarks never measured. An interview with Xiaomin Li and Yuexing Hao, the two founders of MatrAIx: how the personas get built and turned into agents,...
英伟达为什么要保卫开放权重,以及它将如何重塑AI算力市场 / Why NVIDIA Is Defending Open-Weight Models—and How They Could Reshape the AI Compute Market
Open weights are not just a model-release choice; they reshape who buys compute, who controls inference demand, and how much bargaining power enterprises have against closed APIs.
The Evolution of Agents: From Context Engineering to Long-running Harnesses / Agent 从 Context Engineering 到 Long-running Harness 的演变过程
From LLM + tool use to context engineering, and then to long-running agent harnesses: agent capability is becoming a system property composed of the model, harness, context, tools, evals, sandbox, and state management.