Alex Kim

Why economists should like booze A martini doesn’t just steady the nerves after a rollercoaster week

The Economist says Gen-Z’s aversion to alcohol caused: falling restaurant profits, lonelier people, low infertility, declines in innovation and productivity, etc.
Is this satire and I just didn’t get it?

economist.com/finance-and-...

it's more like semi-intentional creation of sludge: freakonomics.com/podcast/slud...
the more friction there's, the less spending the university incurs.

How would you comment on parts of technical writing that is clearly written by an LLM (e.g. it "delves", "embraces" and "revolutionizes")?

"too LLM-y"?

If you ask the Deepseek chat app to talk about Tiananmen Square, it'll refuse immediately.
If you trick it, it'll provide a complete response only to remove it completely when the generation is complete (see the very end of the video).

#deepseek #llm

I wrote a blog post on how to integrate two open-source projects SkyPilot and MLflow for LLM fine-tuning orchestration on Kubernetes

alex000kim.com/tech/2025-01...

depends on the quadrant :)

Engineers: reply in Slack threads
Everyone else: chaos directly in the channel

first transistor (1947) to first personal PCs (1970s)

My best decisions of 2024:
- finally went through with a tonsillectomy
- bought an ebike
- joined a brazilian jiu-jitsu gym
- read a few books on parenting and negotiations (surprisingly, a lot of overlap between the two genres 😀 )

Bsky vs that other platform

I am curious which part(s) keeps breaking for you?
We've had ours for 4 years, and so far so good. Folks on reddit say good things about Triflex HX1.
Also, I'd avoid the Shark brand.

I noticed that SLURM is a common choice for workload management for LLM training, so I wrote a blog post on how to get started with it:
alex000kim.com/tech/2024-11...

GitHub - KellerJordan/modded-nanogpt: NanoGPT (124M) quality in 7.8 8xH100-minutes NanoGPT (124M) quality in 7.8 8xH100-minutes. Contribute to KellerJordan/modded-nanogpt development by creating an account on GitHub.

I think the first thing I'll try is to see how much time it'll take to run this training github.com/KellerJordan... So far the record on 8xH100 is 5 mins.

Among my friends only people without an intense workout habit have experienced work-related burnouts.

So far yes, soon to have a whole cluster :)

What would you try with 8xH200 Nvidia GPUs?
#databs #machinelearning

Shoutout to the ones who say "I'll take care of it" when a messy problem is up for grabs and everyone else is dodging ownership. 🙌👏