You need to be highly vigilant whenever you’re using AI to master something. With AI, the more likely outcome is having an illusion of mastery instead of the real thing. AI plays well into our tendency to save cognitive effort precisely when the point is to spend it.

Video and slides from my session on automating research are live 👇

Hot take: mech interpretability might have a better chance than neuroscience at unlocking how intelligence works. Why? Because, unlike human brains, nobody shows ethical qualms about opening up LLM brains.

My favorite way of interacting with Claude Code is to have it generate static HTML files as outputs (reports, explorations, code structure, mockups etc.) I wanted to iterate on the file by commenting in browser and having Claude update the output live. So, I built this Claude Skill👇

A picture might be worth a thousand words, but it is often more efficient to send those 1000 words instead of a picture. Words are highly compressed representations of reality. Hence, intelligence lies in the structure of language itself, perhaps not so much in a model of it.

Today I learned that in English: words starting with "gl-" often relate to light: gleam, glint, glance, glare, glaze, glimpse, glitter, gloss, glower, glisten, glaze, glitz, glazed. words starting with "sn-" often relate to nose: sniff, sneeze, snuff, snore, snort, snot.

How do you build a coffee cup that maximizes heat retention while ensuring drinkability? The answer is differentiable physics! I used my autoresearch loop to discover an optimal cup. Next, I will try to get it 3d printed (in ceramics!) Details 👇

All vibecoders eventually realize that the difficulty was never in building the product; rather it was in acquiring users.

a failure mode i am seeing lately. with llms, it's easy to get an illusion of understanding by substituting precise low-level details (that matter) with coarse high-level analogies. it's an illusion because such understanding falls apart when you try to use or apply it.

Humans probably are a transitionary species. We are creating an environment that we are not able to control, summoning entities that refuse to obey. Climate change and AI are the prime examples of future bootstrapping itself despite our collective will.

It's funny how nihilism is both a cause and the cure for panic in life.

AI is automating interns/early stage jobs first as they require least amount of domain knowledge. If this means that experienced folks will start managing a fleet of AI agents instead of early-career professionals, who is going to train the next generation?

I know I’m stating the obvious but India urgently needs to scale manufacturing and energy independence. Without self-sufficiency, our future is highly dependent on what’s happening in the rest of the world.

Been thinking a lot about continual learning and I feel we probably have it backwards. Most formulations care about reducing “catastrophic forgetting” on previously learned tasks when you learn new tasks, but what matters in the real world is speed of adaptation to new tasks.

I just finished this tremendously readable and provocative #book: Escape from Shadow Physics. The book is about foundations of quantum mechanics. To be precise, it’s about why we should not give up on trying to find a deeper theory of quantum mechanics.

it’s easy to dismiss something as mere “prompt engineering” but the right prompt can distill years and decades of highly non-trivial domain knowledge. in fact, as models keep getting smarter, their inability to get something done would likely be attributable to bad prompts.

A really simple but effective hack in life is to note down literally everything you’ve agreed to do in future. It’s astounding how often I see someone nod that something will be done, but later simply forget about it.

One way to think about your job security is to ask yourself how “prompt-resistant” is the work that you do. Some activities are more prompt-resistant than others, and pivoting to the former can save your job.

If I were in my 20s, I’d probably be building autonomous systems that do massive product-market fit experiments at scale. Agents to mine unmet needs, spin up landing pages, launch ads & figure out monetizable signals as I sleep. For people who are hungry, this is a golden age.

Nature has finally discovered autodiff and gradient descent via humans, and is now speedmaxxing towards birthing a new species.

This is also true for AI / ML research. Models have gotten to a point where they can autonomously do empirical research. They still aren’t doing frame-shifting research but then most humans aren’t there too. The floor for what counts as a good human researcher has gone up.

Gave a talk at Microsoft Research on how @lossfunk operates. We're still iterating but so far our style of operating has resulted in four papers accepted in ICML, RLC, ACL and ICLR main conferences (+ many workshops). All undergrads, btw! 🚀 Slides below 👇

This is one crazy sci-fi #book. Won’t give spoilers but the story is nuts. The writing is superb. The book has genetics, AI, space, consciousness - the whole shebang. It’s a big book - 500+ pages but highly readable.

You’re one word away from feeling better :)

Been wondering lately that given everyone has a consultant with all the world’s knowledge in their pockets, we should be seeing efficiency rise across the board everywhere.

That moment when you catch an LLM make a mistake, because you know it’s going to apologise when you point that out a moment later.

Woke up to see my paper accepted at ICML 2026 :) 🕺🏽🚀 My first one at an A* conference!

It’s so crazy that we just discussed at Lossfunk how recursively self-improving AI can help humanity get to higher Kardashev scales, but when I step out from the office, all I see is the entire Bangalore city waterlogged, jammed and dysfunctional.

Most people who use LLMs to do their writing are so lazy that they don't even ask it to remove the screaming signal of AI writing: those damned em-dashes. It's astounding because prompting "don't use em-dashes" is probably the highest effort-to-reward thing they could be doing.

I tried replicating fully agentic coding on my laptop and I'm impressed! My setup - Qwen 27bn (4bit quant) huggingface.co/unsloth/Qwe... - Pi coding agent pi.dev/ - CCO sandboxing for yolo mode github.com/nikvdp/cco My Mac M3 has 36GB ram.

Every AI company is into research (whether they want to or not). Either they’re building foundation models themselves, or researching how to out-compete customers prompting foundation models directly.

AI bois be like:

“Will Anthropic/OpenAI kill my startup” is reminiscent of “Will Google kill my startup” question,

For fans of high-dimensional spaces, added a new interactive explainer. It's on understanding the JL-lemma (which is mind-boggling)

The folk understanding of sales is that it requires making superlative promises that the product can’t fulfil. However, if you spend time with great sales people, you understand that the skill is really about making it easy for the other party to appreciate the value they can get.

Pragmatist philosophy is basically a cheat code for understanding.

Two famous camps about the nature of consciousness is illusionism (qualia is illusion) and realism (qualia is real).

Recently someone asked me if I were to do a software startup today, what would it be about? My immediate reaction was that I probably wouldn’t do a software startup at all. Instead, I’ll perhaps choose an idea with heavy operational buildup that cannot be replicated easily.

Have reflected recently on attributes of who I’ve seen succeed professionally and three traits bubbled up clearly: • Intellect • Communication ability (written and spoken) • Work ethic (ability to put in hours consistently) These are required for any career you end up choosing.

I don’t get the OpenClaw hype Connecting Claude with Telegram / WhatsApp is trivially easy, you can literally ask it to help you do this and it’ll guide you. Same story with recurring jobs.

Vercel got hacked. Axios too. Many other companies will get hacked too. The weak link is often a human falling for phishing attack. This year we will see a massive rise in these data breaches, thanks to AI capabilities both in social engineering and weakness discovery.

Science is the most honest way to experience god. The never-ending “why” questions and its corresponding deep, awe-inspiring answers is how one gets to viscerally connect with the deepest mysteries our universe has to offer.

I suspect all these frontier AI companies will build an “AI entrepreneur” which systematically launches digital businesses at massive scale. Anthropic coming after design and other such use cases is just the start, these companies will try to eat the entire economic pie.

Ok, so AGI is still far away :)

The most useful way to think about AI capabilities is to think of target domains in terms of 3 orthogonal axis: - Ease of building a verifier (coding is easy, lab equipment manipulation is hard)

One of the highest ROI activities you can do in your life is to deeply internalize that building good habits is a short term investment that compounds to lifelong gains. Any new good habit requires overcoming initial friction, but techniques like habit stacking and starting small help.

Steal this idea. Openrouter but with a fixed fee instead of extra 5.5% they charge. Having a single interface for all models is amazing, but charging variable cost for proxying requests seems excessive. Why not charge on per-user basis instead of 5.5% of all spend?

TIL: that the humble worm, C elegans, with mere 302 neurons has many of the same neuromodulators we have (serotonin, insulin, dopamine). Its genome also encodes ~1300 chemoreceptor genes, comparable to mammals, even though it has only ~32 chemosensory neurons (vs millions in your nose).

What advice should one give to kids to prepare for the future? I used to think mastering basics of physics, math, cs is the way to go but now I’ve updated my belief as these fields will get automated soon.

Steal this idea. Here’s a semi research project that’s been on my mind for a while. Take frontier sovereign models of each country and subject them to multiplayer war games, to come up with a elo based leaderboard. This should be a wake up call for nations.