76 (@amao5) profile picture
c.n.d.
Joined January 2010
76 retweeted
一个让你 Claude 的 Token 更耐用的经验:禁用 1M 上下文,尤其是在用 Fable 5.1 的时候很明显。 ~/.claude/settings.json "env": { "CLAUDE_CODE_DISABLE_1M_CONTEXT": "1" }
52
31
19
204
165,224
推荐一些我目前安装的Claude Skills: 1. Anthropic官方Skills github.com/anthropics/skills 2. Superpowers 1.6万Star的Skill精选,从脑暴、写需求文档、开发、测试全包含,口碑相当好。 github.com/obra/superpowers 3. Planning-with-files 参考Manus的Agent方法写的Skill。很适合多步骤任务,用这个Skill指导其他Skill工作也挺好。 github.com/OthmanAdi/plannin… 4. X-article-publisher-skill 王树义老师写的X文章发布Skill, 这个值得研究下,后续看看能不能实现其他平台的自动化 github.com/wshuyi/x-article-… 5. NotebookLM skill 自动上传PDF、Youtube链接到NotebookLM,很适合NotebookLM内容的自动化处理 github.com/PleasePrompto/not… 还有一些是自己平时总结的内容创作skills,后续继续分享具体使用场景案例!
70
465
21
2,381
622,305
Fuck Canada Post!!!
1
44
Why I don't use Scrum to manage my Remote Teams? TL;DR: It adds 8 hours of meetings per Sprint. That's 2 full days of lost productivity, per team member, per month! This is what I do instead:
15
56
5
690
154,900
🚨 BREAKING: GPT-4 image recognition already has a new competitor. Open-sourced and completely free to use. Introducing LLaVA: Large Language and Vision Assistant. I compared the viral parking space photo on GPT-4 Vision to LLaVa, and it worked flawlessly (see video).
58
570
33
2,922
681,698
The Outbox pattern is a game changer. 📦 Here's everything you need to know. The Outbox pattern is a technique for reliable messaging in a microservices system. Here's the gist of the Outbox pattern: - You don't publish events to a queue - Instead, you store them in the database - A worker service polls the database for new messages - The worker service publishes each message asynchronously Why would you want to do this? One thing is for sure with microservices - things will break. A downstream service could be down. The network isn't available. If your operations are coupled at runtime, then the whole operation fails. This also means your operation isn't atomic (transactional). The Outbox pattern helps you decouple your operation from its side effects. You know your transaction is atomic if you work with an SQL database. You can persist the messages to the Outbox table in a single transaction. And then, a background worker will process the messages later. Unfortunately, this comes at the cost of eventual consistency. Processing a message can fail, so you will need to retry it. On the consumer end, you still need to handle duplicates in case of retries. This messaging semantic is called "at-least-once" delivery. A few more benefits are improved reliability and performance. P.S. If you liked this, consider joining The .NET Weekly - my newsletter with 28,000+ engineers that teaches you how to improve at .NET and software architecture. Subscribe here → milanjovanovic.tech Did you use the Outbox pattern before?
30
79
2
381
63,032
This is a very interesting approach to improve the robustness of retrieval augmented generation on PDFs. Representing structured documents such as PDFs in plain text for doing RAG is not optimal. This messes up the information present in document structure such as tables, page info, etc. and results in poor answers for which knowledge about doc structure is required. PDFTriage uses OpenAI function calling over input query to extract required meta-data information to answer the query and trigger action based on extracted entities. It achieves this in 3 step method: 1️⃣ Generate doc metadata 2️⃣ LLM-based triage 3️⃣ Answer using the retrieved document arxiv.org/abs/2309.08872
3
21
2
149
17,579
Confused about TSConfig? Here's a cheat sheet that will get you 90% of the way there.
46
429
16
3,022
250,696
⚡️Run LLMs locally with CTranslate2 We've added support for running local models with the blazingly fast CTranslate2. Thanks to GH eryk-dsai from @deepsense_ai for the feature and @HamelHusain for the great post that introduced us to the library! Docs: python.langchain.com/docs/in…
Replying to @HamelHusain
I updated this thanks to @abacaj CTranslate2 allows you to squeeze significantly more performance beyond vLLM
8
30
112
32,594
10 NFRs to consider in your architecture redhat.com/architect/nonfunc…
4
77
307
37,738
Prediction: more and more high-priced SaaS will be replaced by open-source alts, and their complexity will increase Already: ◆ Calendly → Cal․com ◆ Doppler → Infisical ◆ Docusign → Documenso ◆ DocSend → Papermark With each repo, the community learns faster than FAANGs
6
18
5
159
18,399
Improve your VS Code Explorer file tree structure. Add these settings in your configuration file for better visibility: ``` "workbench.tree.indent": 15, "workbench.tree.renderIndentGuides": "always", "workbench.colorCustomizations": { "tree.indentGuidesStroke": "#05ef3c" } ```
146
1,012
60
8,142
1,263,887
The proposal to add observables to the platform is progressing. Next week I'm traveling to Spain (on my own dime) to support this effort and answer questions at W3C TPAC. Getting this fundamental building block into the platform will be a big deal. If you read the explainer (linked in the linked issue), you'll see that a large number of very popular libraries and frameworks have organically hand-rolled their own observables. Shipping an observable with a basic set of methods to the browser will be good for people that love RxJS. But it will also be good for people that *hate* RxJS. Why? Because a minimal set of methods mean fewer people will reach for esoteric RxJS operators when they don't need them. IMO, in an ideal world, the observables built into the platform will replace the need for RxJS in 90% of cases where it's used. It'll also mean that libraries and frameworks don't have to roll their own, and instead use an observable that ships with the platform. github.com/WICG/proposals/is…
16
30
4
278
42,314
FWIW: EventTarget (which is on the browser and node) does NOT suffer from this issue. You can't try/catch the `dispatch`, and all handlers are still executed, with the error being thrown in its own call stack. You can handle it with `process.on('error')` or `window.onerror` if you'd like to do something with it. But realistically, you should be handling such errors in the handler itself.
Today I learned that node's EventEmitter suffers from an issue called "producer interference". That's where an error thrown by a synchronous consumer (whomever called `on`) prevents the producer (the `EventEmitter`) from notifying subsequent other consumers. This is something we had to change about Observable/Subject in RxJS 6 because of discussion about it in the original TC39 proposal and issues we saw it cause in production apps. Imagine that you hand this emitter instance to several components, and each of those components is managed by a different team. An error from one of those components will arbitrarily stop some of the other components from getting notified and therefor they'll seem broken, even if they didn't have a problem. "Spooky action at a distance" sort of errors. These can be very hard to suss out. (thanks @matteocollina and @dailydotdev for bringing this to my attention! very interesting!)
3
21
2
154
43,829
Gateway API represents the future of load balancing, routing, and service mesh APIs in Kubernetes, Kuma 2.3+, Linkerd 2.14+, and Istio 1.16+ are all fully-conformant implementations of Gateway API Gateway API v0.8.0: ServiceMesh support More details: kubernetes.io/blog/2023/08/2…
11
82
8,408
Threw up a little recipe on how to animate a number with Framer Motion: buildui.com/recipes/animated…
7
49
4
438
60,388
This is shocking. Falcon 180B has been released. Trained on 4x the compute of LLaMA 2 70B. It sits between GPT-3.5 and GPT-4 in terms of capabilities. We're now less than two months away from GPT-4-level open-source models. huggingface.co/blog/falcon-1…
36
87
23
567
130,136
Ever wondered how LLMs generate text❓🤔 Let's take the magic out of it and break things down to first principles! Today, I will explain the concept of conditional probability and its connection to LLMs. Let's go! 🚀
30
132
6
686
191,878