my take is to skip learning without a goal. top-down >> bottoms up. because you should know why you're learning something. read a research paper's abstract, then learn what every concept means while you're reading it through and understand why they're getting their result

my parents gave me enough privilege to take bets where I can do cool work even tho the immediate payoff is lower than what I could have gotten with some boring work

now that Lotus Biscoff is available for ₹30 a pack in India, I wonder if the demand for Biscoff-flavoured Cafe items (cheesecakes, ice-creams, shakes etc) will go down

buying gadgets on quick ecommerce feels weirdly bratty behaviour to me, even when it's cheaper than amazon/flipkart

I've been in a systems research lab for 2½ years now and I wish I had this sooner. the framework for cs systems research 1. choose a general process (LLM inference, distributed serverless functions, etc) 2. pick a hardware (edge CPUs, server blades, etc) 3. read papers and run

I've been in a systems research lab for 2½ years now and I wish I had this sooner. the framework for cs systems research 1. choose a general process (LLM inference, distributed serverless functions, etc) 2. pick a hardware (edge CPUs, server blades, etc) 3. read papers and run

a lot of what we think of what consciousness is comes from language

Here Comes Another Bubble (AI Edition)

@Grummz @Wikipedia it's crazy we're championing killing of the single most valuable resource on the Internet that humans laboriously crafted for an AI algorithm that has never seen the outside, real world

quantized versions of an LLM can beat the original non-quantized one? does that make MT-Bench a bad benchmark?

I have been drinking medium-light roast arabica pour-over for a while and maybe I'm brewing something wrong but robusta espresso does have a kick that I miss sometimes

“A truly blue ocean because who has had the audacity before this (and the outside chance of it actually being true) to say it’s now possible to cheat on everything or carry around a digital person” “we’re in a world where the product improves on its own”

this can’t be real

Great communicators are the riskiest people to hire. They could be great, but they also could be great salespeople and lack all of the other substance you need in an A+ hire. I've botched several hires because of this. Now, whenever possible, I require candidates to write

since switching to Arc from Chrome, one thing I really miss is Google Password Manager being built right in. Apple's Password Manager just takes more clicks to do the same thing. And doesn't sync with my Android. :/

Raise $10M seed. Put in a 4.5% yield checking account. 3 days later tell investors “we’re at a $450K run rate.” They give you another $50M. 3 days later tell investors, “you’re not going to believe this but we’re at a $2.7M run rate” Repeat until you IPO.

college clubs doing events/workshops for free is nice and needed for a good top of the funnel but at some point there needs to be an incentive for more people to work in or start more clubs, and existing clubs to survive and do smth grander.

YC has been a magnet for posers almost from the beginning. If you lack substance, you need brand, and YC supplies instant brand. So separating the earnest founders from the posers has always been a big component of the application process.

Anthropic should really focus on marketing Claude as the best webdev coder. With better LLMs, someone will choose a model over another only bc they BELIEVE its better in their task, even if the other model maybe marginally better. Thats been the case with Google vs Bing

OpenAI/ChatGPT vs Anthropic/Claude could be something like Microsoft/Windows vs Apple/macOS.