@0xfakeSpike @AYi_AInotes @otterpal24 Thanks for the tips! I'm vibe coding this graph KB to help my claude code understand my research project and help me working reeyarn.li/blog/2026/04/2…
English
Reeyarn Li
1.4K posts
Reeyarn Li
@ReeyarnLi
Academic researcher in accounting and finance. Member of @trr_accounting & @UniPB, learning #Python & #NLP, defending Freedom for #HongKong.
Katılım Mayıs 2013
887 Takip Edilen248 Takipçiler
说个暴论，现在90%的AI Agent记忆，全都是假的。
我之前也踩过这个坑，把所有历史记录决策日志全堆进Markdown文件里，以为这就是给Agent加了长期记忆，结果用了两周就崩了，
同一个事实有三个互相矛盾的版本，上个月的偏好和昨天的权重一模一样，每次调用都把所有东西一股脑塞进上下文，慢到离谱还经常串台，
直到看到这篇文章才恍然大悟，原来我根本不是在做记忆，只是在把Prompt当RAM用🌚
真正的记忆不是堆文件，应该是图和节点加嵌入加遍历，
Markdown方案有四个根本解决不了的硬伤，没有去重，没有衰减，没有排名，超过一百条记录直接变成性能杀手，
它只能记住你写过什么，永远记不住这件事和那件事有什么关系，
这个决策为什么被否决，上次遇到同样的bug我们是怎么解决的。
向量检索也不行，它只能告诉你这两段话长得像，不能告诉你它们之间的因果关系，
只有图遍历能做到，它能像人脑一样，从一个节点牵出一整条相关的记忆链，
重要的事情越来越清晰，过时的信息自动淡化，矛盾的内容在写入时就被解决。
现在所有生产级的Agent框架，Zep Cognee Mem0，全都是基于图的，
Neo4j已经把图记忆做成了标准的MCP工具，
Claude Code超过二十万行代码之后，纯上下文窗口早就没戏了，
真正能让它像高级工程师一样思考的，
是把不变的规则放在CLAUDE.md里，
把所有演化的状态全部存在图里，动态检索按需拉取。
很多人还在卷一百万两千万的上下文窗口，以为越大越好，
但生产环境里真正致命的，
永远是跨会话的记忆漂移和上下文污染，
内存架构的升级已经不是锦上添花了，能不能把Agent真正用起来才是关键的生死线。
AI Edge@aiedge_
中文
@emollick If they want results in this direction they would not do Claude skills that could have improved all results
English
I don’t really understand how so much careful work going into a study and then it uses non-frontier LLMs with no better models as comparisons. At this point, we know the trend line, if a weaker model is close, it is likely a good model can pull it off. We don’t learn much from it
Rohan Paul@rohanpaul_ai
This paper tests an LLM on real NHS (National Health Service) medication reviews and finds it spots risks but misses safe fixes. They ran a medication safety reviewer on structured United Kingdom National Health Service (NHS) primary care records, mostly coded fields without typed notes, then had an expert clinician grade 277 sampled patients. AI did not miss any case where an intervention (a clear action, like starting or stopping a drug) was needed, but it produced a fully correct review in only 46.9% of patients. When it failed, the main issue was context, for example acting confident with missing details, applying guidelines (the usual best practice rules) without patient goals, or mixing up drug facts. The paper argues this gap between spotting risk and choosing the right next step is why LLMs still need human checking in real clinics. ---- Paper Link – arxiv. org/abs/2512.21127 Paper Title: "A Real-World Evaluation of LLM Medication Safety Reviews in NHS Primary Care"
English
Reeyarn Li retweetledi
🚨 New WP alert‼️
1/6 Our new paper introduces a multi-country, multi-sector NK model with domestic & international input-output linkages. It extends standard open-economy models (e.g. Gali & Monacelli 2005) by incorporating production networks into foreign shock transmission.
English
Reeyarn Li retweetledi
The PhD is a resilience bootcamp under cover as an academic degree.
You enter thinking you'll master a field.
You leave knowing how to:
• Keep working despite chronic uncertainty
• Recover from devastating feedback
• Rebuild when your approach fails
• Defend your ideas under fire
These skills matter more than your thesis topic ever will.
English
Reeyarn Li retweetledi
"Free markets failed for 700 years. When poor farmers couldn't pay, a simple quota beat auctions. Sometimes the invisible hand needs handcuffs."
New paper by @javdito & @JosespinSanchez
restud.com/the-illiquidit…
#REStud
#EconX
#EconTwitter
English
Part of the Lean 4 proof that @HarmonicMath produced with its Aristotle system giving it Gold Medal-level performance at this year’s International Math Olympiad (IMO), printed on dot-matrix printer paper at their office in Palo Alto. Great to have a coffee with Tudor Achim @tachim. Startup activity for machine learning for mathematics, and verification in general for software and hardware, is on fire. nytimes.com/2024/09/23/tec…
English
Reeyarn Li retweetledi
💎 Emily Riehl, renowned professor of mathematics at Johns Hopkins University:
“This is what I think the future of mathematics could be.
We might have domain specific formal languages: for algebraic geometry, also for condensed mathematics, also for differential geometry, lots of different subfields.
And mathematicians will be able to quickly learn to write proofs in new formal systems because of the help of a computer proof assistant,
It will tell them when they're doing something correctly or something incorrectly.”
Full conversation with Math, Inc.’s @jessemhan and @jdlichtman soon.
English
@JEFworks The AI detection itself is quite messy.
Academic writing can be boring. Just run with some pre-2022 referee reports you will see a lot of false positive.
English
Newfound disappointment both as a scientist and as an editor: receiving shallow peer reviews that are entirely AI-generated. The whole point is critical feedback from YOU as a real expert.
If reviewers are outsourcing to AI, I think we've lost the plot 🤷♀️ Not sure the solution.
English
In the immediate horizon of 1–3 years the AI impact should be reshuffling; improving productivity by early adopters. Adoption would not yet be immediately systematic. Just look at the messy Information System we started adopting since 1990s. Are we there yet? AI will take longer
English
Reeyarn Li retweetledi
HUGE update to Claude Code today. You can now access Claude Code and your local files via the Claude Desktop app. No more terminal window!
Simply open the desktop app, select Code and press the install button. Thereafter, just select your folder and away you go.
cat@_catwu
If want to use Claude Code but are don't like the terminal interface, you can now use local Claude Code from Claude Desktop! To do so: 1. download Claude Desktop 2. open the sidebar and click 'Code' toggle 3. select the folder that you want Claude Code to have access to 4. submit your prompt!
English
looks like it is working. just cost some CPU resoruces...
so inefficient
I can imagine Linus shouting at me already
English
Working with AI is so amazing. I just have coffee and provide inputs, and watch the magic ...
English
DeepSeek has just released a new preprint. During a holiday again!
arxiv.org/pdf/2512.24880
English
Reeyarn Li retweetledi
🚨 FULL CONVERSATION
Fields medalist Terry Tao sits down with Math Inc's @jessemhan and @jdlichtman for a conversation on the future of mathematics.
"I got convinced that this was the future of mathematics [...]
It's a different style of writing proofs that actually is in some ways easier to read—harder to check by humans, but you see more clearly the inputs and outputs of a proof, which traditional writing often conceals [...]
I think the definition of a mathematician will broaden."
English
Reeyarn Li retweetledi
I turns out I need more assumptions than in the original paper to have a clean lean-4 proof; the tricky part is, is the conditional expection continuous?
github.com/reeyarn/lean4_…
English
I can neither confirm nor deny this may be the first Lean 4 proof of a theorem in a classic accounting disclosure paper (Dye, 1985).
github.com/reeyarn/lean4_…
English
- Legal documents
- Literature review
- Presentations
- Front-end web UI design
- programming in Python, Stata, R, …
And now math proofs
What else can’t LLM do? 8.9 > 8.11? lol 😂
English
I’m amazed again today by what LLM can do: mathematical proofs.
Via Lean 4, a computational proof verification programming language. When the LLM writes a proof in Lean 4 and the proof is cleared by running the algorithm, it is then officially a correct proof.
English