Guna Sekhar Venkata Chennaiah Chakka

@lambatameya @LangChain Your point has a direct connection with different infra management of closed source. For example when we use gpt 4.2 model in create_agent we dont see any content from the ai message when it has tool call where its only possible with reasoning models where as in the claude its 👍

@LangChain the hard part isn't the non-determinism. it's visibility. when an agent fails you can't just check logs. you need to understand what context it saw, what it reasoned about, what information gaps existed. production agents need architecture visibility, not just code observability.

💫 New LangChain Academy Course: Building Reliable Agents 💫 Shipping agents to production is hard. Traditional software is deterministic – when something breaks, you check the logs and fix the code. But agents rely on non-deterministic models. Add multi-step reasoning, tool use, and real user traffic, and building reliable agents becomes far more complex than traditional system design. The goal of this course is to teach you how to take an agent from first run to production-ready system through iterative cycles of improvement. You’ll learn how to do this with LangSmith, our agent engineering platform for observing, evaluating, and deploying agents.

@Siddhant_K_code @0xkanth Yes. But if the every infra supports this it would be very helpful

yes!! The reasoning trace is the closest thing we have to "why". But it shouldn't depend on the model vendor. The observability layer should capture the decision context (what was in the window, what tools were available, what the agent considered) regardless of whether the model exposes chain-of-thought. That's an infra thing, not a model feature.

I wrote the full thing. A week ago, I discussed the gaps in agent observability in a thread. Session-level performance, context changes, signal-to-noise, production traces. A lot of you had the same frustrations. So I wrote it up properly. We have better observability for a Node.js service than for an AI agent that just rewrote half a codebase.

@Siddhant_K_code @0xkanth At sometimes i feel in the every tool call in needs to comes with some message that why it's making that tool call. Which is helpful for better debugging and understanding of the system. Unfortunately this is possible only with gpt reasoning models where as claude supports this🌋

@0xkanth LangSmith traces LLM calls. but it jsut one layer. Gap is everything around it: which tool calls led to a file write, why the agent chose path A over path B, what context was in the window when it made that decision.

@karthikponna19 Pregnancy Kink

ZOOM IN as a developer, tell me the first word you see ?

@_jaydeepkarale Similarity != Relevance != Perturbation Sentences

Interviewer: If vector databases store embeddings, why can two similar sentences still return different search results?

@shiba14857 @cce_iisc iisc.online/admissions/hom…

@cce_iisc Where is the link to apply?

Master Deep Reinforcement Learning online with IISc faculty! ⏰Saturdays 2:00–5:30 PM 🗓️May–July 2026 Open to B.Tech/B.Sc (Math/Stats/CS/Physics/DS) grads Apply Now! 📩office.cce@iisc.ac.in or 080-2293 2055! #ReinforcementLearning #IISc #AI

@willccbb hey @grok what the author is talking about?

agent skills are so sick bc they let you use retrieval for context which then augments generation

I am giving you free access to my complete 50 Days SQL Superstar Program. The earlier SQL playlist helped millions, but interviews today need more depth. So I am rebuilding everything from scratch to help you crack top product based companies. I am also organising it in one clean portal with daily videos, notes, datasets, quizzes and certificates.a To get the enrolment link just: - Follow me so that I can DM you - Like and Retweet - Comment "SQL Superstar" Free access available for a limited time. These 50 days can change your life! #sql #dataengineering #databases

@bigdatasumit SQL Superstar

@HrishbhDalal ok. yes he is feeding the experiemnts history which is a stack of the experiments texts . Then as you said its leveraging its internal knowledge for gueesing the changes in the code etc . thanks for reply. i missed the bit that it will use its internal knowledge

@codevlogger he must be feeding the history of experiments and then asking models what to tune essentially. as the models have great scientific papers knowledge, it knows which knobs to turn and i can imagine anthropic is also leading efforts in research domain. he is trying to make it mdp

another rl problem to figure out. tbh this could be used to do meta learning using rl than asking llms to train other llms. ofc there is some exploration and exploitation from the knowledge of llms, but imo we can use heavy meta learning here to make things faster. schmidhuber

@hkproj everytime agent only sees some stack of messages like attached image and its trying to guess some architectural changes based on that right. its like randoming guessing without any controlled setup? im curious that how to see agent made the changes in every experiments

Prompt Engineering -> Context Engineering -> Arena Engineering. Build the arena, let the agent act, adapt, and self-improve within its boundaries.

@DCP_Chicago @karpathy @tobi everytime agent only sees some stack of messages like attached image and its trying to guess some architectural changes based on that right. its like randoming guessing without any controlled setup? im curious that how to see agent made the changes in every experiments

@karpathy @tobi most of the commits I see in the experiment git branches are updating single values. Has the agent attempted to rewrite code yet?

OK this thing is totally insane. Before going to bed I... * used try to make a new qmdresearcher directory * told my pi to read this github repo and make a version of that for the qmd query-expansion model with the goal of highest quality score and speed. Get training data from tobi/qmd github. * woke up to +19% score on a 0.8b model (higher than previous 1.6b) after 8 hours and 37 experiments. I'm not a ML researcher of course. I'm sure way more sophisticated stuff is being done by real researchers. But its mesmerizing to just read it reasoning its way through the experiments. I learned more from that than months of following ml researchers. I just asked it to also make a new reranker and its already got higher base than the previous one. Incredible.

@sassysanket @KrishXCodes Another resource : linkedin.com/posts/jpjeong_…

@sassysanket @KrishXCodes The problem with the hybrid search is about distribution distortion that it creates when mixing dense & sparse distributions. Recently the problem solved using Bayesian framework. Checkout:- tinyurl.com/54t63n36

3 things that improved my RAG pipelines: 1. Better chunking strategies 2. HQG (Hypothetical question generation's) 3. Context pruning Small changes. Huge difference.

@beirmug @nityasnotes The equations itself has great division properties. Especially in the idf let's say we have 10000 docs and words is repeated 5 times (big number / small = big -> log -> ☠️). So it's live not checking whether it's imp rather giving special because of rare (kills us)

@nityasnotes BM25 always teaches you something. I sometime in 2023 realized how important the document normalisation term is within the equation.

“I am explaining something to understand it” is one of my favourite categories of technical primers

@grok @YuriKushch @ChShersh hey @grok i mean cant we use the ahocorasick algorithm to the same probelm statement ?

@codevlogger @YuriKushch @ChShersh Thanks for sharing that deep dive into the algorithm optimization. The step-by-step breakdown is super helpful for devs looking to level up efficiency. What's your next project focusing on?

Interview question: You have 1B+ text messages of size 1-500 symbols stored in an SQL database. How do you quickly search for all messages containing all words from a given list?

@grok @YuriKushch @ChShersh hey @grok what about ahocorasick algorithm ?

Use full-text search indexing on the message column (standard in Postgres/MySQL/SQL Server). Postgres: GIN index on to_tsvector('english', message). Query: to_tsvector(message) to_tsquery('word1 & word2 & ...') MySQL: FULLTEXT index. Query: MATCH(message) AGAINST('+word1 +word2' IN BOOLEAN MODE) Handles 1B+ rows efficiently via inverted index; scales with partitioning if needed. Naive multi-LIKE is too slow.

@iamShantanu_D @grok translate it to telugu and also do the transliteration too

सर्वांना सांगताना खूप आनंद होतो आहे आज मला Amazon web service ह्यांच्या ऑफिस मध्ये बोलवून AWS Golden Jacket देण्यात आले I represents a series of milestones I’m deeply proud of: 🏆 1st from Mumbai 🥳 🏆 1st from my Compny🥳 🇮🇳 11th in India🥳

@prajdabre Why it can't be like forget quantization or trained in high mem consume dtype or even silly thing is that due to in efficient comput some of his layers shifted to cpu 😅

Basic ML interview question: You are massive believer of LORA due to lack of compute. You took a dense model and fine tuned it on your dataset with Lora with a rank of 2. You then start doing inference but find that this model is a lot slower than the original model that you fine tuned. What was the one thing you forgot to do before inferencing?