iiinova — where your digital self keeps evolving

Gautham Pai @gauthampai.bsky.social · 04/01/2025

So, what happens when Google AI meets The Onion? 🤔

Gautham Pai @gauthampai.bsky.social · 04/01/2025

The applications decide what they trust by using the context of the statements; e.g. who said what and when and what credentials they had to say it.

000 0

Gautham Pai @gauthampai.bsky.social · 04/01/2025

Not everything found from the Web is true and the Semantic Web does not change that in any way. Truth - or more pragmatically, trustworthiness - is evaluated by each application that processes the information on the Web.

100 1

Gautham Pai @gauthampai.bsky.social · 04/01/2025

It's time we start building "trustworthiness" into the training system. I like the "Semantic Web's" approach:

100 1

Gautham Pai @gauthampai.bsky.social · 20/12/2024

That said LLMs will continue to need a second brain, just like how we use a note taking software. Not everything can happen in the LLM's memory.

100 1

Gautham Pai @gauthampai.bsky.social · 20/12/2024

Right now we use Knowledge Graphs in RAG which is external to the LLM. The LLM's brain doesn't have this structure. Intuitively, it seems that LLMs can do with better understanding of the world. After all, we don't see the world in terms of "alphabetical or language tokens".

100 1

Gautham Pai @gauthampai.bsky.social · 20/12/2024

The work happening with things like Large Concept Models, byte latent transformers and fine tuning is giving me an impression that in the near future we will see more of structured semantic information getting into the core of neural networks.

100 1

Gautham Pai @gauthampai.bsky.social · 18/12/2024

This screenshot is of Newsflash. Plus, I use Feeder on my phone. I follow some hashtags in credible sites to discover good feeds.

000 0

Gautham Pai @gauthampai.bsky.social · 13/12/2024

Can we bring #RSS back to the limelight and start displaying the icon prominently in our blogs, next to the other social icons?

112 2

Gautham Pai @gauthampai.bsky.social · 11/12/2024

Falling in love with Feed Readers again. I recovered one of my old OPMLs. 1. More than 50% of the feeds were no longer working. 2. Some were working but no longer maintained. Authors had moved on to new domains/hosts. I currently have 877 feeds in my reader.

100 1

Gautham Pai @gauthampai.bsky.social · 11/12/2024

Do you take notes with your laptop or the good old fashioned way using a pen and paper? Check this out: www.scientificamerican.com/article/… Although this is a very old article I think it's even more relevant today!

scientificamerican.com

A Learning Secret: Don't Take Notes with a Laptop | Scientific American

Students who used longhand remembered more and had a deeper understanding of the material

000 0

Gautham Pai @gauthampai.bsky.social · 10/12/2024

So excited to know that adding a /rss to the end of a Bluesky profile gives me its RSS feed! For Mastodon, it's simply suffixing with .rss! Here are my RSS feeds as an example: bsky.app/profile/gaut... mastodon.social/@gauthampai....

mastodon.social

000 0

Gautham Pai @gauthampai.bsky.social · 10/12/2024

Of late, I am spending more time reading feeds the old fashioned way using feed readers. I am finding it way more useful than sifting through tons of noise to find the gems in social media apps.

010 1

Gautham Pai @gauthampai.bsky.social · 07/12/2024

These are great! You should continue!

000 0

Gautham Pai @gauthampai.bsky.social · 03/12/2024

Oh alright. I get it now. You are talking about when we add tags to an article. It can definitely be an issue but for the tags that I have, I can align it better even if I make a mistake the first time.

000 0

Gautham Pai @gauthampai.bsky.social · 03/12/2024

Not sure what you mean by keeping track in our head. This page is auto generated from Frontmatter: jnaapti.com/articles/tags

jnaapti.com

All Articles | jnaapti

by Gautham Pai

110 1

Gautham Pai @gauthampai.bsky.social · 03/12/2024

1. Ctrl+Click works for me in Visual Studio Code in the editor. 2. Links/images are rendered in the Preview. When you click on Links in the Preview, it opens the corresponding linked page in the Preview!

000 0

Gautham Pai @gauthampai.bsky.social · 03/12/2024

Well, I am debating on behalf of the monolith'ers. 😂

010 1

Gautham Pai @gauthampai.bsky.social · 03/12/2024

I don't like someone debating X is good, Y is bad, without giving a clear answer to how all the benefits of Y can be achieved with X. The one thing I feel that may be a promising solution to supporting multiple frameworks is Polyglot frameworks that allow cross language communication.

100 1

Gautham Pai @gauthampai.bsky.social · 03/12/2024

Some people say Microservices is more to do with organizational culture, team sizes and the lines of code. I have used Microservices in my application when I was the only developer and the number of lines was less than 50k, because of precisely what you said.

000 0

Gautham Pai @gauthampai.bsky.social · 03/12/2024

However, I do acknowledge the issues with Microservices and I always ask engineers to evaluate the pros and cons before deciding the right architecture for their application.

000 0

Gautham Pai @gauthampai.bsky.social · 03/12/2024

One thing that I don't see any of these articles address is the ability to mix multiple languages and frameworks. Some parts of the application are better built with JS based stack, while something else is better built with Python. How can we address this in monolithic architecture?

100 1

Gautham Pai @gauthampai.bsky.social · 03/12/2024

I am in the same boat as you in the sense that I like the Microservices architecture. See what the article says here: Untangling these dependencies without breaking existing functionality takes time. This is precisely where problems start with the monolithic approach that can get worser over time.

110 1

Gautham Pai @gauthampai.bsky.social · 03/12/2024

The computational needs per user is a small slice of what the infra provides and someone is helping us with this slicing (eg: Cloudflare). And this usually comes at an additional cost plus a vendor lock in. PS: I love the idea, but I do have some concerns.

000 0

Gautham Pai @gauthampai.bsky.social · 03/12/2024

A database per user is getting normalized with people rethinking how SQLite fits here. When it comes to computer and real time connection, depending on how the underlying infra is managed, we need to see if this results in a bigger bill for every user. 😄

100 1

Gautham Pai @gauthampai.bsky.social · 03/12/2024

In one of my current projects, I use front matter to store tags. During the build process, I create a map of tag to articles and use it to generate pages statically. I find it easier to manage. There are some old sites that I manage that still run on WordPress where data is in the database.

120 2

Gautham Pai @gauthampai.bsky.social · 02/12/2024

A few usecases I can think of are in extending the LLM's world view with additional information (which we want to connect with its existing world view), teaching it new languages, styles of replying or even censoring it.

000 0

Gautham Pai @gauthampai.bsky.social · 02/12/2024

In the context of a CMS, if we intend to build pages with client side rendering, I agree with you. It's easier to store them as tables and query them.

100 1

Gautham Pai @gauthampai.bsky.social · 02/12/2024

Under classic Unix, as many programs as possible are written as simple filters, which take a simple text stream on input and process it into another simple text stream on output.

000 0

Gautham Pai @gauthampai.bsky.social · 02/12/2024

git diff is just one of them, but even things like grep, sed, head, tail, etc would be unusable. From TAOUP: Unix tradition strongly encourages writing programs that read and write simple, textual, stream-oriented, device-independent formats.

100 1

Gautham Pai @gauthampai.bsky.social · 02/12/2024

I like when you throw new ideas and make us question long term assumptions. 😄 But, what would be the benefit of storing config data in SQLite compared to YAML? Would this not break the Rule of Composition that many Unix tools are based on?

200 2

Gautham Pai @gauthampai.bsky.social · 02/12/2024

These docs don't look out of the ordinary. There are logical sections, so it shouldn't be hard to chunk them. Could some very docs have swayed the total size? You can arrange the docs by size and see the top 5%. But these should be fine!

000 0

Gautham Pai @gauthampai.bsky.social · 02/12/2024

Also, if data is converted to triples, the use of Graph RAG techniques can help us run Graph queries to query and filter what is needed and only inject that into the context. Thoughts?

100 1

Gautham Pai @gauthampai.bsky.social · 02/12/2024

Don't you think that context size and the use of tools may play a role here before we start talking about fine tuning? Some of the reasoning/aggregation can be done outside the context and then the results can be injected into the context.

200 2

Gautham Pai @gauthampai.bsky.social · 02/12/2024

Is there a logical way to divide each document to smaller chunks? If not, I would agree with @angilly.bsky.social that fine tuning may be needed. But you can still start with RAG and assess the results first anyway.

110 1

Gautham Pai @gauthampai.bsky.social · 02/12/2024

Agreed! What is your current understanding atm?

000 0

Gautham Pai @gauthampai.bsky.social · 02/12/2024

Can you also provide the count? Because it does seem like 32MB is a lot for 7k files.

100 1

Gautham Pai @gauthampai.bsky.social · 02/12/2024

Hey, that worked! Thanks for helping out. 👍

010 1

Gautham Pai @gauthampai.bsky.social · 02/12/2024

PS: I had to run my answer via an LLM to fit it into 300 characters. Sorry if it sounds robotic but the thoughts are 100% mine! 😐

010 1

Gautham Pai @gauthampai.bsky.social · 02/12/2024

Further, document chunking strategies impact retrieval & performance. Semantic chunking (breaking notes at logical points) ensures context is preserved. Overlapping context (adding adjacent sentences) helps too. Start simple and refine your chunking as you better understand retrieval and query flow.

110 1

Gautham Pai @gauthampai.bsky.social · 02/12/2024

Fine-tuning your LLM isn’t always necessary. If relationships like "Python Library for RAG" are well-captured in note chunks (via links, tags, or structure), pretrained models can retrieve effectively. Fine-tuning helps only if your model doesn't understand domain-specific terms or relations.

100 1

Gautham Pai @gauthampai.bsky.social · 02/12/2024

The design of your chunks will depend on the structure of your notes, how these relationships are captured and the type of queries that you would want the LLM to answer.

100 1

Gautham Pai @gauthampai.bsky.social · 02/12/2024

For queries like "Python library for RAG", how is the connection between "Python Library," "RAG," and "LangChain" concepts captured in your notes? Is it part of the hierarchy, tags (like Zettelkasten), or backlinks?

110 1

Gautham Pai @gauthampai.bsky.social · 02/12/2024

What are "related notes"? For a query like "What is LangChain?" related notes can be chunks containing answers, retrieved with basic keyword-based methods. But what about queries like, "Which Python library can I use for RAG use cases?"? Here, explicit relationships in the notes are crucial.

100 1

Gautham Pai @gauthampai.bsky.social · 02/12/2024

I'm passionate about this problem because it combines two of my interests: #LLMs and #KnowledgeGraphs Let’s analyze your requirement further: How can related notes or ideas from a knowledge vault be retrieved effectively for LLM queries? It starts with how your notes are structured and connected.

120 2