HNHacker News
TopNewBestAskShowJobs

Jianghong94

73 karma · joined April 14, 2023

submissionscomments
Jianghong94··on Do We Need a 37-Cent Coin? (2009)
> Probability of a transaction resulting in value v is uniform from [0,99].

in reality, most of the transactions that use coins end up conforming to common existing coin combinations e.g. laundromats in US mostly price as multiples of quarters ($0.25)

Jianghong94··on Data Engineering Vault: A 1000 Node Second Brain for DE Knowledge
well people may need some time to relearn that forgetting is part of brain function for a reason
Jianghong94··on Data Engineering Vault: A 1000 Node Second Brain for DE Knowledge
This is sooo true it hurts.
Jianghong94··on g1: Using Llama-3.1 70B on Groq to create o1-like reasoning chains
yeah or a lot of people can just fake progress by attaching whatever viral tag onto their glue code. I mean to start with, unless you do a bit of fine-tuning + rlhf there's no way to do it o1-like.
Jianghong94··on Electric Clojure v3: Differential Dataflow for UI [video]
I work in a finance company and basically what our product is in terms of tech is a lot of client-server spreadsheets, with a huge $ amount per cell. So when I saw the demo and the use case I got extremely hooked! Unfortunately due to the hiring conditions and tech competence, there's almost 0 chance of adopting something like this, but still, I'm impressed.

I haven't got enough time to finish the video today, but I'd like to throw out some of our major technical domain problems and see how people familiar with electric would try to solve them:

1. How does it handle multiple different functionality servers? And potentially server-server networking? e.g. if I need to get some data from server A and some other data from server B, and combine them into the UI, is that doable in electric? Also what if I want to save the combined data onto the database on server C? Now for simplicity, let's just assume we're talking about less than 20 servers here and there's no problem with the port, so each one of them can keep 1 websocket per each other server open up. 2. How does it handle listening to a stream of events and triggering if the event meets a certain criteria? e.g. like if the price of GOOG is lower than a certain price then trigger a popup or something like that. This may not be the intended scope of electric Clojure but it is what we do day to day, so I'm curious. 3. How does it handle interop with existing Java apps? Is it just a normal java-clj interop?

Overall I really like the concept of being network transparent, or shall I say that devs would be better served by spending more time on business data and function, but not on coming up with derived request/response obj, API route, etc. Also, if this type of library ends up in some sort of more mainstream programming language, I would vouch for adoption in a heartbeat.

Jianghong94··on Show HN: Visualizing Chess Games
Really cool! Unfortunately, I don't play chess so can't comment on it. I used to watch StarCraft 2 (an RTS game) matches a lot and a recurring theme of viewership is that there's no specific display mode for the audience, so a lot of information is hidden and left to commentators. I believe future RTS games could develop a mode where the whole battlefield are split into grids and display a high-level unit movement/recon/territory capture etc akin to what you're doing here.
Jianghong94··on Ask HN: How to manage docs for LLM RAG app?
Thank you!
Jianghong94··on Ask HN: How to manage docs for LLM RAG app?
Yes. There're a couple of problems I encountered during my trial and ended up choose to build my own: 1. the version I was using requires the configuration of a 'reranking model', which AFAIK has to be from cohere. Funny enough, I can later opt to not use this reranking model, but I cannot get any functionality working if I don't preconfigure the cohere API key for this model. The problem with this is that, I need to request my company's infosec to allow API call onto cohere, for a functionality that I didn't intend to use.

And I think this results in my observation of the startup llm app world vs the heavily-regulated finance industry world: that the startup llm world is taking a default opt-in, default saas mode; however the regulated clients cannot benefit from these practices due to arcane regulation requirements. In this case, dify requires the setup of an API key for a service that I don't intend to use, and I have to go through the hurdles just to get the hello world going.

My analysis is that, there are 2 ways to make the heavily-regulated clients happy: 1. major cloud provider integrating everything into the myriad of its service catalogs and offer a full suite of services with the sales and compliance completed as a whole. 2. startup go down a curated, opinionated, battery-included oss approach with the least of sass fuss.

2. for some reason when I test using the simplest QA template, it seems dify cannot retrieve the document of any kind that I uploaded. I don't know why it happened but debugging the simplest offering of an app that boasts ease of use is not my thing.

I didn't go further from there. I wish my experience is just an extremely rare case since dify definitely seems to be the most polished all-in-one provider out there. If my experience turned out to be more positive I'll definitely stick with it, now I just feel picking a all-in-one is not the right choice.

Jianghong94··on Why Triplebyte Failed
> Are you referring specifically to salary negotiation here? I think this is one of the things that makes a trusted intermediary useful.

Yeah, I agree. I do feel like you're not going to take the venture-backed route this time? Cause in my opinion that's part of the reason why drive triplebyte went down this route. Based on your writing you probably concluded triplebyte model back in its heyday was going to be a decent business, just that it can't be a VC home-run.

Jianghong94··on Why Triplebyte Failed
I can't help but want to draw the parallel between triplebyte and dating app: both are marketplace for 2 parties, or say networks; both networks are inherently two-tier, one premium tier and one, let's say subpar tier.

For the premium tier, the problem is less so the efficiency of the matching process because participants have powerful signals, I would say the problem is the efficiency of negotiating a fair term because both parties are not incentivized to go to a public auction, hence no market price. Hence the hiring process may be less so investigating potential candidates and more so "let' just wait two weeks and see who's the highest bidder".

For the subpar tier, the problem is less so the efficiency of the matching process (this still is part of the problem), but more so a cost-effect analysis: is it worth throwing X amount of resources to filter a huge pool of unknown quality candidates?

Obviously, one more problem is that there's no easy way to use whatever incentive to quickly bump up the supply.

So intrinsically, triplebyte/dating app tried to use tech to alleviate the pain point, but it is a part of the bigger, convoluted mating dance, and improving a non-critical part of the whole pipeline doesn't yield a so-called step function gain overall. That, plus venture backing, forces triplebyte to go down the rabbit hole and try to acquire as much of the lower subpar tier network as possible, yet this dilutes their attention and human touch on the premium tier, and goes down a death spiral.

Sorry I do not intend to bash the author or the founders for picking this problem/user need to solve, it just occurs to me that this heterogeneous network with no easy way to increase supply/lower demand, is really hard for a marketplace company and sometimes simply throwing technology onto the white-hot competing segment may not be the groundbreaking solution. I wish the author all the best with their next endeavor

Jianghong94··on Why Triplebyte Failed
Maybe we can add: 1. going remote and support remote; 2. set up offshore offices.
Jianghong94··on Show HN: Cognita – open-source RAG framework for modular applications
Congrats on the launch!

I find it relevant to what I want to do next and put in some time to understand the application vs other stuff e.g. Langchain. And if my understanding is correct, what this tries to do is:

For a lot of typical web services, there're non-realtime batch-processing data processors, e.g. search engine's crawler and indexer, or database's OLAP system, Hadoop, spark, etc. Once their processing is done, they will output data in arelevant, easy-to-use form for real-time web services to consume, e.g. search engine's index, or a list of e-commerce's best selling items.

If we extend such analogy to today's LLM RAG application and compare it with an out-of-the-box Langchain or LlamaIndex implementation, we'll realize everything is in one process altogether. Of course, for demo purpose, they have to.

Cognita tries to fit in by splitting the process into real-time and not real-time parts, on top of existing LangChain and LlamaIndex, and comes with an API endpoint for each part and a web UI for user querying.

For my use case, I'm looking into setting up a very basic RAG-based internal doc QA app, to see if this helps with some of our notoriously bad wikis. So I'm likely going to use this UI and just shovel whatever simple LangChain or LlamaIndex implementation into it. I'm not that interested in the modular design. Honestly, I could see a couple of different ways each market segment approaches such a problem: for demo/mainly static document/low stake application, the need to periodically refresh vector-db is non-existent; for companies with enough engineering expertise, they'll likely put the data processing part into existing data processing framework; for the rest segment, they probably can also get away with putting the whole offline data processing into a very long python script, setup cron and call it a day.

---

I haven't look into RAG in a year or so, but my overall sensation is this: 1. the RAG layer (on top of vector-db) isn't technically difficult, vs say OS development, database development, etc, after all, text manipulation has been around since 60s. 2, since the whole LLM generation is very sensitive to prompt, an early, too rigid abstraction likely do more harm than good.

Jianghong94··on AI for AWS Documentation
A couple of things come to mind: 1. embedding methods: there're a couple of ways to do that, the most used one is OpenAi's text-davinci-002, although in my use case(short sentence description of API) it didn't work pretty well; 2. how you truncate documents into pieces: for this langchain has some implementation and helpful pointers.

I think you have to do lots of experiment on this till you find your best information retrieval strategy

Jianghong94··on AI for AWS Documentation
AWS has some great documentations
Jianghong94··on Severance payments at Wikimedia Foundation
I mean, can't we form a web3 dao to perform acquisition of WMF (or a hostile-takeover?), if it works out donors get to vote next leadership, etc etc, drive cost low and make sure things are in good hands, if it doesn't work out then just return the cash back, plain and simple.
Jianghong94··on Building a Better Book Club: A Strategy for Efficiently Ingesting Nonfiction
What's the difference between this methodology and others already proposed out there other than this incorporates LLM summarization? See, the slowest part of learning is when your brain deeply engrain the concepts into your subconsciousness, and there're few ways to do this other than repetition and practicing. Read 10 best books again and again is way better than skim-read them
← PreviousPage 2 of 2