Hieu Pham
@hyhieu226
@xai | ex: @augmentcode, @GoogleBrain, @LTIatCMU, @Stanford, ACM ICPC, IMO🥈
You rich kid. I saved my childhood money for the ATI Radeon X800 to experience FarCry. Crytek was a dream come true. I had no idea I would be writing kernels for those beasts today.
When I was a kid, I saved up for a year to buy a watercooled radeon r9 290. I overclocked it to within an inch of its life---when I played games, it felt like someone had turned on a space heater. Not much has changed, except now I work with 10^9 times the flops.
While a notable milestone, this is already borderline trivial for AI
Human-level math competitions will soon be trivial for AI, just as human-level chess/Go/etc became trivial for AI. In 1997, a mainframe was needed to beat Kasparov at chess. Now, your phone can beat Magnus Carlsen.
Yes, you can just paste thousands of lines of your source code to @grok.
Test-time compute (self-improve based on a defined target score). arxiv.org/abs/1611.09940 We did it before it was cool (ICLR reject, workshop accept, 2.2K citations).
After using uv for a while, I am convinced that installing Python by default on an operating system is such an evil blasphemous act.
TIL apt install -y has never meant to be --yes. Also IIRC, in the Borg era, Googlers usually do "yes | borgcfg xyzt" to auto say yes. Everything should have been yolo.
TIL the Gemini CLI has a rickroll embedded in its help documentation. 🤣
One useful skill in the era of AI is to be able to quickly read through AI answers to find what you need, or to tell that the answer does not have what you need. This is skill is useful not only for humans but also for AIs themselves.
When you get deep into technical work, it's easy to forget the big picture, and hence diverge from the first principle. Always stay alert. If you see an intermediate requirement, always ask if it's absolutely needed. If not, it's dumb. Takes practice to train a mind that way.
In case you don't know, when I was a PhD student, we had Transformer and Adam but not yet trillion text tokens from all languages. Then, the first omni-lingual parallel dataset was The Bible. Which makes total sense. It was translated into virtually all languages.
you really can just train an LLM in a church
Can't believe I argue with a neighbor on this question. If you uniformly randomly pick a real number between 0 and 1 inclusively, what is the probability that the number is rational?
Nothing boosts the software development productivity as much as a functional IDE and a smooth terminal. AI can be in the top 5, but nowhere near those two.