Kai

🚀 my first World #1 achieved - My general AI / Agent Memory Algorithm. Weeks of no sleep, no food. -> 2024: The first idea 2024 → 2025: Develop a brand-new memory algorithm, inspired by my quantum physics experience + cognitive science. → 2026 First fully engineered -- showing breakthrough on retrieve efficiency -- accuracy / token cost. Need to catch up with the accuracy next and start to built to prototype of LLM wiki by @karpathy. Memory is the most critical foundational algorithm in AI infrastructure — and the strongest complement to large models. I believe that next big thing will be AI for science break through and memory will play a crucial role in it. I will publish the algorithm soon.

本周小试了下新的thesis，1周41%应该跑赢光，电，存配置了吧？ 无奈美股没啥钱，就换了两顿猪脚饭

@xingpt 注意力带宽不够，但是招人有一个很大的问题就是协作可能一样cost你的带宽。 你需要一个ppm orchestration， 我自己就在做，有用但是不够有用

想法非常多，想做的Project和项目也很多，要做的研究和内容也很多。 能力上我+Claude Code+GPT+DeeSeek+Hermes都能搞定，但问题是没时间搞； 这种情况怎么解决？要招人吗？

@qzhxjhj @LeePima 哈哈哈，很棒，@xhlin 就比较高深，给我们推荐situation awareness， 然后我们都只看了，然后没了😂

@Kai_Manifold 主要感谢Pima @LeePima ,我是从他推荐的股票里选了闪迪，其实是投SOL的思路，用Crypto的思路投美股的

self-reflection: 以前分析别人的案例，很喜欢总结成路径依赖，最终还是落到了自己头上，炒美股一周可以有30%的收益， 自己对AI和市场的理解还能帮上忙，偏偏要去crypto被画线割韭菜，当然依然相信BTC is the future

@qzhxjhj 厉害， $SNDK 是真nb，不但妖，还是最强叙事之一， 叠加出来的大怪兽

@Kai_Manifold 这次就买了一个 $SNDK ，幸运买到最妖的美股 Crypto我也是一样思路，找一个最妖的币

--Self-Reinforcement： One of the most important principles I’ve held since 2024 is this evolutionary paradigm shift: From “build a function to solve one specific problem” → to “systematically scale up specific human capabilities” as the real goal. Garry Tan’s Thin Harness + Fat Skills powerfully validates exactly this principle — proving I was on the right track all along. x.com/garrytan/statu…

We’re already seeing this evolution in the past: --v0: Basic chat capability (general language ability) --v1: Claude Code: Scaling coding-related thinking & reasoning as a reusable capability --v1.1: OpenClaw: Scaling long-term memory + proactive execution as durable cognitive capabilities.

This is the fundamental difference between the two revolutions: Internet era vs AI; and why so many AI startup trend disappeared so fast in the past 3 three year. Internet Era = create better programmable functions to solve a specfic problem / need. AI Future = turn human judgment, reasoning, memory, and initiative and so on into scalable, reusable cognitive capabilities.

One of the most fundamental conceptual understanding of 2026 by @garrytan : Thin Harness + Fat Skills. It is not just a technical focus shift in engineering, but actually a deeper paradigm shift — a principle I’ve held since 2024 — From “To build a function to solve one specific problem” to “Not solving defined problems, but systematically scaling up a specific human capability.” In a short -- Abilities >> Functions

@enzyme_dev 哈哈哈，是的，我觉得基本是时差导致它算不明白😂

@Kai_Manifold 他的服务器是不是和你有时差？😅

Deepseek V4 again made a temporal mistake. So 04/25/206 is Saturday not Sunday right? 😂

Do you remember when you joined X? I do! #MyXAnniversary

Some more result for more optimized parameters. In some categories, V4 is over Claude Sonnet, however show some instability and week on higher level understanding.

How Deepseek V4 is truly catching up with Claude on real complicated human comprehension tasks. This is one crucial step in our new algorithm: Manifold Cognitive — An Artificial Hippocampus for Agents, new and unknown engineering design (not the usual math and code). The figures tell everything 👇

claude code 和 codex 让我上瘾的程度，已经达到了能轻松对抗安眠药的程度， 有一种当年打魔兽世界，一定要把声望刷完， 每日做完，副本cd用完的感觉。