Dhruv Agarwal

Grad school applications are stressful! I shared *practical* tips from my experience at Microsoft Research last year and my slides were greatly appreciated. To encourage inclusivity in academia, I am making my slides public: agdhruv.github.io/files/practica… #AcademicChatter @ThePhDPlace

@SmithaMilli will share once ready!

@agdhruv ahah I'll be curious to see your experiments!

Community Alignment v1.1 is out! ✅ +41k comparisons => 233k comparisons total ✅ Improved demographic representativeness => All 5 countries have subsets balanced on age, gender, ethnicity ✅ More natural lang explanations => in 44% of comparisons, users explain their choice huggingface.co/datasets/faceb…

@adityaag Hard relate. Writing code was always my creative expression and it's been disorienting to see it go away. Earlier the joy was in *how* you build stuff, now it's in *what* you build. Wrote more about this feeling as an ode to coding before AI: @agdhruv/an-ode-to-coding-before-ai-3acc35cf95da" target="_blank" rel="nofollow noopener">medium.com/@agdhruv/an-od…

It's a weird time. I am filled with wonder and also a profound sadness. I spent a lot of time over the weekend writing code with Claude. And it was very clear that we will never ever write code by hand again. It doesn't make any sense to do so. Something I was very good at is now free and abundant. I am happy...but disoriented. At the same time, something I spent my early career building (social networks) was being created by lobster-agents. It's all a bit silly...but if you zoom out, it's kind of indistinguishable from humans on the larger internet. So both the form and function of my early career are now produced by AI. I am happy but also sad and confused. If anything, this whole period is showing me what it is like to be human again.

@TrungTPhan Also check out AI-generated commentary: x.com/agdhruv/status…

The Australian Open (AO) has a very underrated use of AI. It doesn’t have full broadcast rights for all matches, so the AO YouTube livestream uses AI to help render Nintendo Wii Tennis cartoon avatars that mimics the action on a 2-minute delay. As a result, this animated clip of Daniel Medvedev smashing his tennis racket on the net from last year’s tourney remains among the best AI-related video outputs to date.

The NeurIPS hallucinated citations debacle is a good example of Goodhart's law: When a measure (NeurIPS acceptance) becomes a target (prestige), it ceases to be a good measure (people will do anything to achieve it)

AGI will come when models stop using .get() to index into Python dictionaries

@actuallysoham Haha yeah I built the precursor to messcat (which I gather was on WhatsApp instead of Messenger). Not sure if it's still around!

@agdhruv I had no idea you had made what we used to call MessCat haha 🤣 I remember maintaining it for a while and I wonder if it’s still in use?

An Ode to Coding Before AI Like most others, I learned to code in 2014, long before AI coding agents, autocomplete, and such. My friend and I would exchange HTML/CSS/JS tutorials on 1TB hard drives. I’d flip between VLC, Sublime Text, and Chrome on my family’s old HP laptop, trying to center a div or fix a PHP bug I barely understood. We built websites at high school tech competitions without any internet access. In college, I built a lowkey viral Facebook Messenger chatbot that shared our campus food menu and bus timings. It crashed when someone sent it a cat GIF haha. I "vibe-coded" so many other projects (the vibe came from my inexperience) and encountered creative bugs on the way. Every bug was painful, but every bug taught me something: OS quirks, file systems, logs, APIs, threading, the GIL, tensor dimensions. These days, I use Cursor, Claude Code, and all the AI tools. They’re amazing! But I do feel a weird nostalgia. When AI writes so much of the code, you enjoy the product more than the process. As an engineer, I was wired to enjoy the process, but that's changing. AI coding makes me productive, but a part of me misses when coding was slow, messy, and deeply personal...

I also wrote a longer version of this essay in case you're interested :) @agdhruv/an-ode-to-coding-before-ai-3acc35cf95da" target="_blank" rel="nofollow noopener">medium.com/@agdhruv/an-od…

Excited to be at the Alignment Workshop at @NeurIPSConf 2025 to talk about AI safety and alignment. Unexpectedly some of the conversations were about pre-training, not just post-training! Looks like Gemini 3 Pro is making pre-training cool again!!!

AI commentary unlocks new possibilities - Personalized commentary: beginner-friendly, numbers-focused, drama-heavy, biased to your team...your choice! - Interactive: e.g. discuss strategy with the commentator as the game unfolds - Democratizing sports: commentary can drive viewership for smaller/local games Built this with @justachetan and Sushrut

I've been thinking about AI applications beyond chatbots and coding agents. Here's one we came up with: AI-generated sports commentary! Check it out:

@shaily99 Thanks for the thread...exciting times ahead!

Happily surprised to see OpenAI curating cultural benchmarks, especially focused on India. BUT, cultural knowledge != culturally aligned generations. My work for 2+ years focuses on cultural competence in generative tasks, like creative writing. Sharing some papers in LONG 🧵

@vinodgansan @perplexity_ai @OpenAI My cognitive investment at this point is for Chrome, not Chromium. Investment = muscle memory, knowing exactly what it where, down to ~pixel level. If I switch to Comet or Atlas, I lose that flow, and I gain very little (and Chrome will cover up even those small gains soon)

@agdhruv @perplexity_ai @OpenAI If it’s base + delta, where base being chromium, what cognitive investment are you exactly leaving behind. Resistance to delta I can understand, but the base part I’m unclear.

My hot take on AI browsers: if it's Chromium-based (@perplexity_ai @OpenAI), I'll try it and move back to Chrome. Too much cognitive investment to leave behind (and Chrome will ship the same features anyway). The only way to dethrone Chrome is to rethink browsing – from scratch!

@anupamsobti @perplexity_ai @OpenAI Doesn't it slow you down for less than proportionate productivity gains? My browsing on Chrome is muscle memory. Minor design changes inhibit that flow without much gains

@agdhruv @perplexity_ai @OpenAI I actually like to leave browsers behind, just to start from scratch. Sure, you have a few extensions to set up, but that’s usually all I need to do. Liking my move to atlas so far. Good suggestions and lower barrier to ask GPT.

@cleoabram Great question @cleoabram Our paper earlier this year tested this question exactly! arxiv.org/abs/2409.11360

In prep for my interview with Sam Altman, I reached out NVIDIA CEO Jensen Huang for a question. Here's what he asked: "Fact is what is. Truth is what it means. Facts are objective. Truths are personal – i.e., depends on perspective, culture, values, beliefs, context. One AI can learn and know the facts. How does one AI know the truth for everyone, in every country, and every background?" Here's the start of Sam's answer (more in the full interview):

If you code in python and aren't using uv yet, this is your sign to jump ship! It's so fast and fun...I look forward to installing dependencies now lol @astral_sh

This @acm_chi paper finds that AI assistance homogenized writing in a controlled experiment with American and Indian participants (n=118). With a classifier, it was harder to predict writer origin without AI (83.5% vs. 90.6%). dl.acm.org/doi/abs/10.114…

@ishaan_jaff @yash_347 litellm's focus is models, not tasks. eg for a simple translation task, one could abstract out the message list setup & make it ready to go. Also it has too many features & the docs are a bit distracting. Wanted to make something beginner-friendly. litellm is great overall tho!

@agdhruv @yash_347 Curious what were the challenges setting up litellm @agdhruv ?

🚀 Introducing MiniAI - AI made ridiculously simple! Sick of OpenAI boilerplate? Tired of writing complex chains to do simple AI tasks? MiniAI is a minimal, flexible solution for byte-sized AI tasks. pip install miniai Examples ⬇️