avatar
lidang 立党 （全网劝人卖房、劝人学CS、劝人买SP500和NASDAQ100第一人）
1,596,811 Twitter followers
Follow
co-founder @HedgehogLabHQ @Project_DARC
Posts
avatar
lidang 立党 （全网劝人卖房、劝人学CS、劝人买SP500和NASDAQ100第一人）
03-27
Zhang Xuefeng repeatedly emphasized that AI = mathematical modeling = mathematics. DeepSeek was developed by mathematics PhDs. Mathematics is a jack-of-all-trades; a mathematics PhD can easily switch between fintech, internet, AI, and semiconductors. There are two core concepts I've been discussing for nearly 10 years. One is that computation is computation, a model is a model, and mathematics is mathematics. Computer science, electrical engineering, aerospace, mechanical engineering, finance—a vast amount of theory is expressed using calculation formulas. Most people who don't understand mathematics or these fields would assume, "Mathematical formulas = mathematics." The correct understanding is that each industry has over a dozen courses, corresponding to over a dozen large and small subfields. All the theories and professional knowledge in these fields are expressed using formulas and symbols, using basic tools like matrices, calculus, and probability. However, the essence of these theories and professional knowledge is the knowledge itself. The fact that they are expressed with symbols and formulas does not mean "all professional knowledge = mathematics." Many vocational school graduates don't understand this principle and readily claim, "Symbols = mathematics," "Modeling = mathematics," "What looks like a mathematical formula = mathematics," "Learning mathematics allows you to solve all professional formulas and problems," "Studying mathematics to a doctorate allows you to solve all mathematical problems in finance, internet, AI, semiconductors, materials, and mechanical engineering." The correct understanding is that "Formulas, symbols, tools, theories, and models = quantitative symbolic expressions of knowledge in various subfields"—they have absolutely nothing to do with mathematical theory itself. Another key point is that mathematics is mathematics, and mathematics will always be mathematics. The field of mathematics has very clear and defined boundaries, whether at the undergraduate, master's, or doctoral level. The field of mathematics encompasses geometry, algebra, number theory, combinatorics, cryptography, analysis, topology, and so on. Remember, mathematics is simply mathematics. A mathematics major is not anything outside of mathematics. A mathematics major will not teach you how to write PyTorch, configure CUDA, build ResNet from scratch, or train and tune Transformers. Furthermore, it will not teach you classical machine learning, classical control theory, or classical aerospace fluid dynamics and finite element methods—because these are not mathematics at all, and do not belong to mathematics. And don't believe statements like "studying mathematics allows you to easily switch to any other major" or "mathematics is the mother of all majors."
avatar
lidang 立党 （全网劝人卖房、劝人学CS、劝人买SP500和NASDAQ100第一人）
03-23
Many people want to create their own AI agents or coding agents, hoping to build an all-powerful programming robot. I've repeatedly stated that from the initial SWE Agent, to the sophisticated context management of the cursor era, to the early Claude Code, to various fancy memory mechanisms, to plan mode, and then to a main agent controlling several sub-agents and backend tasks, the technology of the agent itself has undergone three or four industrial revolutions in just two or three years. You can understand that the design of a coding agent tool alone has gone through several iterations on par with horse-drawn carriages, trains, cars, airplanes, and rockets. Today, I must tell everyone that creating an initial SWE Agent is necessary because it has educational value, just like anyone could create an operating system or compiler 10 years ago—it's part of a hands-on lesson. However, if you want to catch up with tools like CodeX, Gemini CLI, or Claude Code, you need to step into the code of these projects and see how complex their designs are. Even back in the days of Rookie Coder, Cline, and Aider, products that were top-tier open-source stars in Silicon Valley a year ago are now generations behind Codex and Claude Code, completely outdated. Not to mention the few large domestic companies with their three haphazardly designed coding agents, which are completely different from Claude Code and Codex – they're products of a completely different era. Even a half-generation gap is like a steam train versus a rocket, and the gap is visibly widening in the short term. I must warn you, Claude Code and Codex could very well become the next Chrome garbage. While garbage, they will objectively become the industry's de facto standard. The end result will be that all coding agents on the market will be three or four generations behind Claude Code, making them all shrink back to selling cheap APIs and manually configuring APIs within Claude Code. Claude Code will become the king of closed-source, and Codex the king of open-source, with the two sharing the market. Others can no longer understand all the engineering details of Codex and Claude Code, just like how you wouldn't understand all of Chromium's open-source code. I just want to tell you that after three years of iteration, the complexity of coding agents is now vastly different. Even companies of Alibaba's, ByteDance's, and the LLM Six Little Tigers' caliber are likely to be left far behind by their Silicon Valley counterparts—a gap they simply cannot bridge.
CODEX
0%
avatar
lidang 立党 （全网劝人卖房、劝人学CS、劝人买SP500和NASDAQ100第一人）
03-21
Thread
The next-generation LLM benchmark should no longer be run on swe-bench lite/verified/pro/ultra. Instead, it should be a closed, separate, and randomly selected platform each month, finding 5000 issues on GitHub, running them with the same harness on all models, and having a referee agent write test cases to determine success. Since the issues are randomly selected, there's no need to worry about the number or percentage of resolved issues; just look at the relative ranking. This eliminates concerns about using swe-bench for fine-tuning or even cheating, and removes uncontrollable factors like vendor A being able to reproduce the problem while vendor B cannot, or vendor C cherry-picking the best result. Everyone only needs to look at the latest official leaderboard ranking for the month. This is far superior to the previous method of running two models simultaneously and having users manually vote on whether the left or right side was better. Note that this ranking's longitudinal comparison is meaningless. Perhaps model A, ranked first in March, solved 60% of the problems, while model B, ranked first in April, only solved 40%, because the issues in March and April are completely different, making longitudinal comparisons unnecessary. Just look at the relative rankings for this month. This is the real cyber cricket fighting.
LITE
0%
avatar
lidang 立党 （全网劝人卖房、劝人学CS、劝人买SP500和NASDAQ100第一人）
03-21
As I've mentioned before, goal-driven is the best solution. github.com/lidangzzz/goal-driv...… One subagent keeps running, while a master agent continuously verifies whether the work is complete and whether the criteria are 100% met. Once met, the entire process is stopped, allowing the master agent to repeatedly request the subagent to work overtime until all test cases are successfully run. This easily completes over 20 hours of work with precise results.
RainbowBird | 洛灵
@luoling8192
03-21
为什么大家的 Coding Agent 可以一直写啊，不需要 Human in the loop 的吗？ 为什么我交给它一个任务，如果碰到了困难就直接和我说“下一步继续”，就卡住了，我说“我去睡觉了”也不行😭
loading indicator
Loading..