Pranav

Everyone’s talking about “AI agents,” but here’s how I actually built one that talks to any dataset, even the messy CSVs from data.gov.in. Here is whole story of how you can do it, and some usefull learning from it Currently, I have created a system that is submitted to Bharat Digital, which fetches or ingests the unstructured data from data.gov.in CSVs. The data is very unstructured, and it is loaded as a pandas DataFrame, cleaned, and then saved into a database. Now this database is linked with an SQL LLM agent that uses the Qwen2.5-Coder 7B model, which can perform expert-level queries on SQL tables, even across 10–100s of tables, and can handle complex questions, including table joins and other advanced operations. We can query it, and after that, the fetched data from the SQL LLM chain is used as context in the main conversation system, which uses the LLaMA 3.2 3B model as a chat interface for end users. So, it’s a flow where we can chat with any amount of data, even unstructured data. For example, if I want to ask, “Give me the state with the highest rainfall between 2000–2008, and also give the corresponding highest growth food or crop in that region,” it can generate a complex table-joining query, extract the data, and use that data as context for the main conversation. My learnings: - Qwen2.5-Coder 7B is the best medium-level reasoning model. - LLaMA 3.2 3B is a decent small-parameter model for conversation. - Don’t use a vector database for complex unstructured dataframes that have thousands or millions of rows — it won’t work perfectly. - The 7B and 3B models are both workable and give decent results. Here is video of demonstration : youtube.com/watch?v=HYfXqB…

@sanskar0627 Yeah and i think programming in go feels more pleasantas compare to rust and python

Major companies like Reddit moved their comments backend from Python to Go, cutting critical latency in half! This highlights a common choice for dev teams: - Python: Great for building fast (quick prototypes, rich libraries) - Go: Great for running fast (low latency, handles many users at once)

@ThePrimeagen Go and C syntax kinda look same, i guess cuz Go creator was also created C. But i love to code in Go, it super simple and more power on your hand

i do find it a little offensive that C has a more robust type system than Go

@craigzLiszt and those who became unemployed after doing vibe code

there are 11 types of engineers: 1. those that vibe code 2. those that will vibe code 3. those that will be unemployed

@krishdotdev due to its speed

Why does UDP even exist ?

@Yuchenj_UW i think most of official & business tasks is already possible in GPT

I knew Sam’s endgame for ChatGPT was an everything app. Social is just phase one. Expect ads, a productivity suite to rival Google’s docs/slides/photos, payments, and full-stack communication ( text, voice, and video).

@_trish_xD it does not to be popular it is already popular, cuz most of default option for developer in SQL database is postgresql oracle or SQL variant i.e SQLite

SQLite is a database that doesn’t want to be famous. It just wants to be useful.

@HeySkidee yeah

@0xprv thinking that we are in a simulation does nothing. it doesn't improve anything. the concept of network of super intelligence is also just a hypothesis and he just assumed that if we create a super intelligence, it will join it if not made humble lol

I don’t know why, but every third person in the AI space seems to start talking in these out-of-the-box philosophical terms, ideas that are completely unrealistic. They know it’s not possible, but they say it anyway just to look intellectual in front of their audience. Like seriously, brother, just sit back, relax, do your work, and enjoy your life.

@i2cjak Bro you forgot rust

mental illnesses: char short unsigned long long long mental healthnesses: uint8_t int16_t uint16_t uint32_t uint64_t

@Yuchenj_UW or is it just matter of luck, like chain of success and same thing happening for gpt, chain of losses

GPT-5: lost 71% in a week. Qwen 3 Max: gained 70% in a week. How is Qwen 3 so good at trading??

@bod_zsn well bro you cover 90% audience of twitter ..

I don’t know why, but every third person in the AI space seems to start talking in these out-of-the-box philosophical terms, ideas that are completely unrealistic. They know it’s not possible, but they say it anyway just to look intellectual in front of their audience. Like seriously, brother, just sit back, relax, do your work, and enjoy your life.

Nick Bostrom says if we create ASI, we might be introducing it to a universe already full of gods. If superintelligence already exists across galaxies, quantum branches, or simulations, we'd be birthing a mind to join that cosmic host. Perhaps that's our only way into the larger order of intelligence.

That is very disturbing data. In just a few years, AI generated content has reached nearly 90%, while human-created content has dropped to around 10%. So what is the purpose of the internet then? Machines are now operating the internet, posting content, analyzing algorithms, and even deciding our choices, what we buy on Amazon, what we see on Facebook, and more. Where is our liberty in all this? I believe there’s an urgent need for a multinational government-led body or a global tech council to regulate this situation. Otherwise, we’re heading toward serious trouble.

To say this is a problem is a gigantic understatement. We live in the golden age of AI slop. Hopefully we will soon gain the right control the algorithms that control our online life in order to filter out AI slop to a large degree.

@cloneofsimo tsmc : ????

ai company: mine gold nvidia: sell shovels google: mine gold, sell shovel, distribute gold, export gold, make shovel, rent shovel, make jewlery, sell jewlery, distribute the jewlery, other jewleries, make cars that ships the jewlery, make chips that make the cars, make recsys to sell the jewlery, make the filesystem to make the recsys, make golden quantum computer, make gold videos, images, audio, ......

I use cProfile for profiling plain Python code. It basically tells you how many times each function is called in your program, and how much time each function takes to execute. Most of my Python performance bottlenecks are easily solved using this tool. I highly recommend it. Here’s an example of its output:

i think none of company working on AGI seriously they saying about AGI for taking funding and marketing stunt to make there's new products on hype everytime, i think in future the company who will make somewhere near AGI intelligence would be a underdog, company from nowhere that emerges

Reminder that AGI is not coming anytime soon. You can tell by observing that OpenAI is trying to become a traditional big tech company like meta with a bunch of slop products and copium revenue streams They made a lot of money from the AI hype and now they're trying to figure out how to keep it.

Like this masterpiece

@slimjimmy Sometime over complexity kills

less than half a person per service

Excessive Supply of low quality content can degrade the demand of content, like if you ask a person what the like watching original contents by human or AI skip generated content, so most of them like Human generated original content, bcz human like human connection it AI connection, idk why these big tech bros promoting these narrative of kind of saying that there AI product can replace XYZ proffesion in future, everyone can do everyone, society don't run on that way.

Andreessen said AI is gonna let creative people make insane movies without ever touching a studio budget We’re about to see a whole new kind of filmmaking, raw, personal, and from voices that never had access before