singhrac

9 ms·

91.
by singhrac 2y ago
Out of curiosity are you operating at a scale where "MinIO on a Hetzner node" isn't a viable replacement? I totally believe that's possible, just curious about the use case.
92.
by singhrac 2y ago
I find Copilot most useful as a kind of very good autocomplete. One way to think about it is: if I know exactly what the next line should be (e.g. I've done X transformation on dataset A, and I need to do the same transformation on dat
93.
by singhrac 2y ago
Ahhhh yes they did it. They delivered workspaces and editable dependencies, which is literally all I needed. I can't wait to set this up, I'm very confident they'll iron out any remaining bugs or gotchas (e.g. CUDA) quickly.
94.
by singhrac 2y ago
Wait, where's this $10/month plan? I don't see it. I see it at https://one.google.com/about , but I can't actually choose that. Edit: I found the answer on Reddit, as usual. You need to go to your plan se
95.
by singhrac 2y ago
I mostly agree, but partially because I can't tell (even though I'm paying $20/month) what tier of Gemini I am getting. I know most people won't care, but because they won't tell me, I will assume I'm getting G
96.
by singhrac 2y ago
At least when I was involved with math contests (early 2010s) it was more that there were a lot of ex-campers at MIT/Harvard, and that did a lot to convince me. Po Shen did a lot to recruit students directly to CMU, but I'm not su
97.
by singhrac 2y ago
We use DuckDB similarly but productionize by writing pyarrow code. All the modern tools (DuckDB, pyarrow, polars) are fast enough if you store your data well (parquet), though we work with not quite “big data” most of the time. It’s worth r
98.
by singhrac 2y ago
Maybe they need purchasing power parity adjustment for those outside the US? $3.5/user/month seems lowish compared to peers in the US, though I agree with what you’re saying about how to charge at 10+ users (that is a nice busines
99.
by singhrac 2y ago
I like it ok, but there is actually a lot of white space (I think current Linux desktops suffer from this quite a bit, though I haven't used one regularly in >8 years). For example, under "Fresh OS interface", the start me
100.
by singhrac 2y ago
I use Obsidian for the first two. Email was very useful but you need to essentially categorize the data yourself manually. I’m very surprised there isn’t a bot that will automatically attempt to categorize your emails for you via IMAP acces
101.
by singhrac 2y ago
This is really nice. I'll take a much closer look when I get time later. I'm very interested in how you think about incremental computation - I find cache invalidation to be incredibly annoying and time-consuming when running expe
102.
by singhrac 2y ago
Ah I didn’t catch that Quickwit had immutable records. That explains the focus on log usage. Thanks!
103.
by singhrac 2y ago
This is a bit misleading since Triton is a bit higher level than CUDA. But the idea is kind of right - there’s active development of AMD and Intel backends, and Pytorch is investing into Triton as well.
104.
by singhrac 2y ago
I’m not sure we agree on what zero sum here means, but one direct consequence of having a decent exit here is that the investors in Silo will get a capital return they can use to raise more funds. I don’t know what the founders of Silo will
105.
by singhrac 2y ago
Maybe I misunderstood both products but I think neither Quickwit or Turbopuffer is either of those things intrinsically (though log structured messages are a good fit for Quickfit). I think Quickwit is essentially Lucene/Elasticsearch
106.
by singhrac 2y ago
We use a self-hosted JupyterLab. I would love to have an external tool though our existing workflow isn’t too bad.
107.
by singhrac 2y ago
The problem is that we all have a lot of FUD (for good reasons). It's on AMD to solve that problem publically. They need to make it easier to understand what is supported so far and what's not. For example, for bitandbytes (a comm
108.
by singhrac 2y ago
This is also helpful: https://github.com/srush/Tensor-Puzzles
109.
by singhrac 2y ago
I'm afraid you've been reading too many sensationalist headlines. You are completely able to get a stable rate for power in Texas. No one is forced to choose a rate that can spike to $7/kWh. In fact power in Texas is cheaper
110.
by singhrac 2y ago
These are very fun. I would say that they are not entirely in order of difficulty so if you get stuck you should try some of the later ones (you can shim the ones you didn’t solve yet if you’d like). The visualizations are great. Incredibly
111.
by singhrac 2y ago
This is kind of headline bait. I work in energy forecasting so I look at these numbers all day. Facebook's 350k H100s might draw around 1kW each, let's say 2kW with the rest of the rack and cooling. That's 700MW for what is p
112.
by singhrac 2y ago
Yeah, and I think people forget all the time that inference (usually batch_size=1) is memory bandwidth bound, but training (usually batch_size=large) is usually compute bound. And people use enormous batch sizes for training. And while the
113.
by singhrac 2y ago
I've opened some issues over the years with Pandas on the transition to Arrow dtypes. They are generally very nice but even today I have many workaround patches in our codebase to avoid bugs (e.g. fillna and dropna working as expected)
114.
by singhrac 2y ago
The same is starting to be true for Texas as well (which is an example of great market design).
115.
by singhrac 2y ago
I haven't taken anything Gary says seriously since like 2017. I'd recommend others do the same, you'll save yourself a lot of time.
116.
by singhrac 2y ago
I traded crude oil professionally until early 2022 and I remember that was how we justified low domestic production in mid 2021 as well (i.e. "producers think these prices are temporary, so they aren't investing"). In retrosp
117.
by singhrac 2y ago
From what I understand FAFSA was delayed by 2 months this year (to December). But you’re right - if other colleges didn’t see a decline then it’s probably not an important factor. I guess it feels too tenuous a connection because I never th
118.
by singhrac 2y ago
This article is decent: https://modoenergy.com/research/ercot-battery-energy-storage... (no affiliation, looked like based on disclosure data). Look at the revenue: it's 6% (now 12%) from energy arb and 58% (now 3
119.
by singhrac 2y ago
Lol, this is probably because FAFSA was complicated this year. You'll find that applications to good schools are more affected by "can we afford this" than any perceived estimate of value (which is always high enough).
120.
by singhrac 2y ago
Curtailment is often a decision of the generator (I don’t know CAISO specifics but I would be surprised if it wasn’t ultimately a generator decision). If I asked you to generate power for -$15/MWh, you would refuse to do it because it

More ›