Apple just dropped a killer open-source visualization tool for embeddings — Embedding Atlas — and it’s surprisingly powerful for anyone working with large text+metadata datasets. This reminds me of Nomic's Atlas, but I never got around to using it 😅 We’re talking real-time search, multi-million point rendering, and automatic clustering with labels. One of their showcase examples visualizes ~200K wine reviews using embeddings + metadata like price, country, and tasting notes. And it is lightning fast even on my browser! No separate code needed! It nails what most LLM devs need but often hack together: ✅ UMAP projections ✅ Faceted search across metadata (e.g. “country vs. price”) ✅ Hover + tooltip on raw points ✅ Interactive filters, histograms, and cluster overlays ✅ Cross-linked scatterplot + table views Under the hood: • Fast rendering using WebGPU (with WebGL fallback) • Embedding-based semantic similarity search • Kernel density contours for spotting clusters or outliers You just upload your .jsonl or .csv with text + vector + metadata. It handles the rest: clustering, labeling, UI layout, everything. This feels like the LLM-native version of Tableau — but optimized for text, chat and modern data needs If you’re building RAG evals, search tuning, clustering explainability, or even dataset audits — this could be your new favorite tool.

Really want to meet the people who convinced Infosys to invest in OpenAI in 2015. Damn, what a bet.

Are you depressed enough? If not, here is a prophecy from @karpathy

BLR is 3 emotions: Bellandur, Kadubessanali, Sarjapur and Whitefield: When will I get promoted? Kora, HSR, Indiranagar: When will my startup raise next round of funding? Jayanagar, Malleswaram, RR Nagar, JP Nagar: When can I raise the rent for my house in above areas?

PSA: If you're doing USD to INR conversion mentally, can just use 85 now.

Hiring is too broken. Govt should have a national employer portal tied to company GST etc. Every candidate you've ghosted, offered but later said no, orders placed to a vendor and then paid late, all POSH complaints, with action taken or NOT — gets tracked there.

Bengaluru rents be like: "This will be just 20K INR" You:

Parents for 20 years: Wake up early, take a walk in the morning sun Kid: No Huberman: direct sunlight within 30 minutes of waking up with forward ambulation really helps Kid: SIR YES SIR!

Do not spend a single rupee on them These are somewhere between a scam and incompetence I regularly get DMs from folks who did a bunch of these "ML" courses in the last hype-cycle with no tangible or learning outcomes I put these in the same category as Byju's for Adults

OpenAI's first India hire is a Policy hire — not sales or engineering I've so many feelings. None of them are good

Hey Cody, can you please put me in touch with your friend? I'd love to make a better offer to these engineers as a fellow country man

Folks who moved to UAE, what are your living expenses for the similar lifestyle as $120K in SF or 40-50 LPA in BLR?

Reminder that the best thing you can do for your career is get so good that the best people find you and want to work with you Everything else: offer shopping, leetcode grindmax is less efficient

This is absolutely amazing! This kills all startups wanting to hack with 4090s in their racks bought off Amazon or Ebay US GoI protects masses from engineers learning dangerous things like matrix multiplication Making sure that only Google and Amazon can "Make in India"

I hear IIM admissions has made a lot of people sad on the Internet It's a great time to remind yourself that comparably premier software engineers (top 0.01%): - make more $$, - work less hours, take less meetings, - 💕 their job/craft And probably do so for next 20 years

.@sama would ❤️ to host you in BLR! Creds: - Run BLR's largest Generative AI meetup - WA Community with 800+ people & 150 DAU - Trained LMs since 2019 (Hindi), maintainer of awesome-nlp (14K Github stars) Who can put me in touch w/ his planning team? RT for Karma please!

Might not look much, but this is what true Chads in India do — from his own Linkedin: The CEO of IDFC First Bank

There are so many amazing Udupi places but none which have TCP

Had a dear friend asking me if they should move to BLR from EU recently. Absolutely amazing how BLR itself has given a clear, decisive answer better than I ever could!

Extremely happy to share that @qdrant_engine — the fastest and most cost effective vector search has raised $28M Series A This raise will help us build our open source ecosystem of tools for efficient and fast search!

Is this a puzzle? Haven't you've given the answer yourself? Americans with money want to invest here for higher CAGR — while Indians don't have enough wealth for CAGR to matter? The Americans do, so it makes sense to go there and earn a honest, respectable living?

Great reminder for those thinking about staying back in India, you better: 1. make > 50 LPA, 2. work 40 or fewer hours, 3. learn more

It's pretty damning if we expect @paraschopra to be the flagbearer of LLMs from India after spending 10K crore on the AI in taxpayer money

Thank you @AnalyticsVidhya for recognising me as one of India's Top 5 GenAI Scientists! 🙏 Glad to be in great company with wayyyyy more persistent mofos like @bhutanisanyam1 and @ravithejads 😊

Most investors in India haven't studied a single balance sheet Many study unrelated things like Twitter, WA groups & YC "Trends" and thesis are learnt on the fly Investors resort to spray & pray or FOMO to make investing work Outcome is fraud & less than FD returns

Extra irony because Krutrim is made by a BITS Pilani + Georgia Tech alum

Can't believe there are people who love Pasoori without realising that the entire song is cursing that you don't find your love

Myth: India can do AI well because we've both the money and talent in software Likely: India has the money, but our talent pools are not deep enough For reference, China has TikTok, Baidu et al and 4 of the top 5 universities in Machine Learning globally are in China

There are only 8000 people who declare income of more than 5 cr

On working hard: Do it at the beginning of your career Hard work compounds like interest & the earlier you do it, the more time you have for the benefits to pay off It’s easier to work hard when you have fewer other responsibilities, which is frequently when you’re young - @sama

Indian devs getting angry about AI regulation is kinda funny to me 🤣 There are >50K downloads/mo across ALL Indian language models put together: OpenHaathi, Telugu, Kannada & Tamil Llama token expansion LoRAs No one is even using Indic AI 🤖 Bhailog thoda kuch ship karein?

Banger. The Employee Provident Fund of India has been hacked by our friend China — and of course EPFO never knew because of the "7 ft high walls" perhaps.

The primary utility of UPSC/JEE is illusion of meritocracy. It's a social contract b/w Govt and the people. It's not about the merit or excellent role-fit at all

The good thing about Devin is that it can work 168 hrs/week and not just 70 hrs/week like Infosys employees

The people dissing Deedy's main claims are clowns. Fight me. @SarvamAI did some amazing work. And yet Indian companies and devs don't care — that's an us problem, not on Sarvam.

1/ As vector databases continue raising funds, let's reevaluate the landscape. Scale is important, but FAISS did billion-vectors in 2017. What do we use them for? Private Search Engines They have vast applications across text, images and audio.

The problem with this is that I'll not be able to tell what's a hallucination from a human, and what's from a LLM

DSPy + Amazon Bedrock improved RAG accuracy from 30.3% to 68.1% — a 124% relative gain over baseline. AWS built a modular RAG pipeline with: •Claude 3 Sonnet via Bedrock for generation •Titan Embeddings v2 + ChromaDB for retrieval •DSPy for orchestration, prompt tuning, and eval They optimized prompts using 32 QA pairs and LLM-as-a-judge metrics (exact match, passage match, and semantic scoring). The DSPy BootstrapFewShot module compiled the best-performing prompt chain. Then they went a step further: Fine-tuned Llama 3 8B using LoRA + SFT, DPO, and ORPO. Despite being smaller than Claude 3, it outperformed both Claude Sonnet and Haiku in accuracy. 📊 Highlights: •Claude Sonnet baseline: 0.3031 •DSPy-optimized: 0.6375 (+110%) •ORPO-tuned Llama 3 8B: 0.6812 (+124%) All evaluation was done using LLM-as-a-judge in DSPy — with plug-and-play modules for generator, retriever, and evaluator. This shows how prompt optimization is necessary to do **before** you start thinking and talking about model fine-tuning (source link next)

This is what Urban Planning done well looks like: 1. 3 well maintained lakes 2. Indian Music Experience 3. Good hospitals, schools, food, coffee: Third Wave, Brahmin's, Dakshin Cafe All at walking distance Yes, this is Bengaluru

frankly the 80%ile engineering college graduate in India is worse at dev than gpt4o there is no reason for anyone with half a brain cell to hire people with so little training and skill & asking them to upskill is a societal failure — they literally spent 4 years in college!

my fucked up brain is thinking about a girl I knew almost 15 years ago who was a big Atif Aslam fan instead of fixing this FastAPI asyncio bug faaaak

Unpopular Opinion: The worst product that people love is not Linkedin, it's Notion

More evidence for my net short position on Indians in India. Lost generation. Beyond point of no return.

Why you should never use pgvector (e.g. @supabase Vector Store) for production: 😮 pgvector is 20x slower than a decent vector DB (e.g. @qdrant_engine) 🤯 And it's a full 18% worse in finding relevant docs for you And this can happen at as little as 10K documents when chunked!

The realisation that they are aging without growing 🤷‍♂️

That is okay: An entire generation is doing jobs that shouldn't have existed to begin with

Engineers working for Desi startups when they could've left for US via MS route

My agent framework is no longer maintained — PyPi project name is now with @dharmesh!

Date yourself by posting a video game that was released when you were 18 years old.

SDE-3 or higher at Amazon/Google make 80+ LPA cash, before taxes and excluding stocks I think that's a useful benchmark for the ambitious

Hey @HanchungLee saw your draft lecture on Evaluating LLM System Would love to invite you to a course I'm putting together on @MavenHQ to share that with Senior Engineers and CTOs Almost everyone I know would benefit from a double click on this slide alone! leehanchung.github.io/blogs/…

Useful if you're booking full body health checkups for a loved one in BLR: Worst: Thyrocare, stabby technicians who can't find a vein, multiple doctors don't trust the reports, worst service Middling: Lal Path Labs (expensive), Apollo Best: Orange Health

There are at least 5 companies which pay $120K or 1 crore INR per annum in Software Engineer III or SDE-3 in India With the highest cash being Udaan — $97K

Enterprise adoption of "vanilla" LLM use cases like Text2SQL requires good AI engineering. See this Prompt Journey diagram from @SwiggyTech — constrained environment with stable schemas, defined relationships, column names, and past queries — and still so much complexity!

Unpopular Opinion: BLR or even India doesn't have enough nerd density for SF Commons to happen or work long term Look at the graveyard of hackathon and tech community organisers nerds: curious about tech for the sake of tech, not just paisa and solving a problem/craft

Almost as hard as getting my first job for me tbh, easier than college elimination exams though: Lost 16.9 kg in 366 days

That is amazing no? Would indicate that the abstractions are robust and you can do pretty decent engineering without building expertise in the underlying concepts. Because if that's the bar, I don't think anyone of us can make a chip from sand? And not many can make a new DB?

Google Photos resurfaced one of my few good memories from college — Everyone except me in that photo now lives & works in North America and has a family there

Typed out Career notes for a friend, sharing here: The universe rewards a specific wish, and punishes a vague desire. You can improve your odds of success of by beginning with something which appears narrow/small & growing Big and successful today? Has too much entry barriers

I've known this for many years, yet I'm still surprised by how true it is: the universe rewards specific desire and punishes vague wishes.

Time for BLR startups to accept leave applications with "lack of motivation due to cuddle weather"

Since too many people have asked me India's AI startups that I keep an eye on, here is the list in no particular order: Dashtoon — PocketFM for Comics Segmind — Model compression, distillation E2E — GPU Cloud for India Sarvam — Indic LLMs+RAG PortkeyAI — LLM Gateway Add more?

Investors & Journalists often ask if there are teams in India making LLMs. That's missing the point tbh. The more worrying fact is that we're not even making world-class tools to interact with LLMs: Orchestration: Langchain Store: Qdrant UX: Gradio Modelstore: Huggingface

Announcing the GenerativeAI Makers meetup: ⏰ 24th June, Saturday 📍Bellandur, Bengaluru 🎙️@zerodha CTO Kailash Nadh @amodm, via Flipkart, Udaan Moderated by Game Dev @charu_tak What questions do you have for them? Please RT for Karma hasgeek.com/generativeAI/jun…

low key Varun Mohan did more harm than Soham to the employer-employee social contract

.@C_P_Gurnani sir since you accepted the challenge, this aligns with your goals: Requesting you to build and enable: - Compute cluster 1/5 of this size - In the next 3-6 months and get it up - Cheaper than Cloud/Serverless Very doable: - Just 500 H100 - 1-2 Tbps network - Doesn't need much: Just money and execution skill What you get: - Bragging rights for generations — India's largest supercomputer, the only one dedicated to AI, 400x the size of IISc - You still have margins left because electricity and salaries are both cheaper in India than us-east-1

ROFL. In 3 hours, you're still coming from the airport. Not in BLR yet.

This looks very optimistic. I think the number is closer to 5 than 2000? There's Talukdar and Pruthi at IISc, Dr. Pratyush, Dr. Mitesh via AI4Bharat — that's it, right? Did I miss someone? (Tempted to include Dr. Jawahar fwiw)

I hate when people ask questions on twitter just to get engagement. What are some of the worst examples you’ve seen?

Navarasa: India's best OSS Indic model got covered at Google I/O! 🔥 This comes from dear friends and Indie hackers: @ravithejads and @ramsri_goutham They both not only used PEFT for Gemma but also curated a lot of the datasets for this!

This definition of agents from @langchain's @hwchase17 continues to be among the simplest and most complete Whether it's a binary if-else or a multiple choice switch-case, the key distinction lies in who dictates the flow: the developer upfront or the LLM in real-time.

Terrible decade to be a junior dev?

Just helped an e-Commerce CTO plan to scale vector similarity from 1M to 100M vectors. Check out our roadmap: 🔵 Blue: Vector DB e.g. @qdrant_engine's capabilities 🔴 Red: What I'll build for them

I've averaged about 20 LPA before taxes over the last few years. But I've also done years where I made 33 LPA I've taken salary cuts almost every year I've also done a few odd jobs which I don't talk about on Twitter which have made me a few tens of lakhs in the last 1-2 years

why 50+ LPA? the simpler solution is to get a 1cr+ job out straight out of college — can get to 10 cr in 10 years for sure that way

This is a *Qualification* for an AI Engineer — so yeah, don't bet your career on becoming an AI Engineer overnight

Giving the credibility of mainstream press like Moneycontrol to scams is a reputation risk for you and them Please consider speaking to ML/GenAI practitioners in the industry Chandra Here is a list of folks who I know to be good & their Linkedin docs.google.com/spreadsheets…

Ask and you shall be given

"Delved" into a 30 yo today Folks who're >40 years old: What would you do differently if you were 30 years old today? Can skip investing and "go to gym" advice 🙈

It's better to lend money and lose friends, then to not lend money and keep the same friends.

my controversial opinion continues to be that Indians should take the summer off and not winter because that's the more brutal season across most of India

My controversial opinions continues to remain that Ola/Uber have done more for the country than ~most philanthropic initiatives

Had a blast speaking at AI4Bharat / IIT Madras! Staying true to my brand, I plugged two amazing solutions from friends: farmer.chat by Dev (BLR) and KissanAI by Dr. Pratik Desai (SF) Thank you Dr. Pratyush, Dr. Vivek for giving me the opportunity! 🙏

After years of cribbing that I'm not a good enough developer, I've done the most developer thing: Made a Python library for devs: FastEmbed — simple, fast, efficient for text embedding⚡️ And it'll be maintained by the amazing friends at @qdrant_engine 🫶

1. Microsoft owns about 51% of OpenAI. Jasper is a MS Word plugin. 2. Almost all modern production grade DL GPUs are AMD or NVIDIA 3. Even today, the largest "utility" ML deployments: Search, Maps, Ads are from Google/Alphabet

Only people new to NLP/search believe that ChatGPT can replace Google search

If you or anyone you know is able to get an alcohol license in BLR. Please ping me, I'll invest! It has a very high entry barrier and I like those businesses

Kailash Nadh blogs about the impact of GPT4 on organisational efficiency nadh.in/blog/this-time-it-fe…

I am perennially grateful to Twitter and everyone here: 2018, Work: Got into (L)LMs via @jeremyphoward's AWD-LSTM, who I found here. Have been doing NLP x Perf since! 2021, Family: Fought someone here on therapy & it's efficacy, met her to apologize over coffee and we're married now 2022, Friends: Met @dementorSam via Smit, who pinged me on Twitter. Taks & him are truly an inspiration 2023, Career: @andre_z and @qdrant_engine stumbled onto me via X threads, built FastEmbed & grew it to 1M installs, used by NVIDIA's Nemo there — Another NLP x Perf 2022-23, Living: Was quite depressed, @adityarao310 pointed me to therapy & >30% odds, saved my life 2024, $: I got my first LLM services client via @waitin4agi_

I'd retweet this, but worried it might be like oxygen during COVID years — half the Indians who need it, won't get it

The amazing #GenerativeAlHack in Numbers: 🧑‍🤝‍🧑 90+ hackers on Day 1, 150+ folks at Demo day ⚡️ 43 submissions, 32 in-person, 11 remote from 6 cities 🙏 13 sponsors from 3 countries, 2 COSS: Weaviate, AppSmith 🤗 20% raised from communities 💰 5 Lakh INR in Prizes

Sequoia India is now @peakxvpartners Even VCs can't escape @peakbengaluru's magnetic pull

How can you always keep your answers fresh and automatically updated? Tutorial Colab Notebook here w/ Detailed Diagrams: bit.ly/updatedQAColab Built with (@CohereAI + @OpenAI + @qdrant_engine) using @llama_index

Put together a list of friends I endorse professionally as well nirantk.com/endorsed/

Every time I get a "Let's get on a call?" I send this: Hey, I'm an engineer by trade and 30 minute calls will cost me 3 hours of focus and energy. I'd love to make that, but I'll have to decline for now.

Best earbuds below 10K INR? - Audio quality for work calls - Music: White noise, ambient and bass not important - Preferably something I can wear while walking Please RT for Karma too? And like the ones in replies which you endorse!

Instead of writing a generic GPT4 hype thread, let me use this moment to just think out loud about what this changes across tech, society and business This is random musing, please do disagree with me politely!

Folks who can finetune a LLM (vision or text), and have some business relationships — can make $10K for about 120-200 hrs/mo of work I cannot in good conscience refer the best engineers to a founder offering to pay $4K/mo 🙏

10 years out: The wealth gap between working at Google and get rich and another who is broke because of working in startups 10 years out will only get larger

From @nvidia to @Cloudflare CEO come to give guest lectures at CS153 Infra at Scale @Stanford I can't think of a single Indian university which has an "applied" course of comparable scale, in any CS sub-discipline