basiccalendar74

Joined 10 years ago

Karma
93
Submissions
60
Member since
Jul 19
View on Hacker News

Recent submissions

Comment · 3 months ago
Weights will be fetched 4-bit from memory, but compute happens in FP8 (W4A8) or BF16 (W4A16). This is still better than fetching 8-bit weights from memory.

Both W4A8 and W4A16 schemes are supported by Hopper GPUs and commonly used to serve mxfp4 Kimi models on Hopper.

Comment · 3 months ago
can you show the math for your last statement?
Comment · 6 months ago
it is only "right" statistically as in conforming to the same distribution. but there is no guarantee of exact same output.
Comment · 1 year ago
A large cap stock increasing by 6x is not "unlikely to be achievable".

Apple went from 1T to 4T valuation in just 6 years. why did Apple not award 500 billion to Tim Cook for this achievement?

Comment · 1 year ago
by 'allocated', I mean allocated during the purchase decision. not that money is sent to respective company.
Comment · 1 year ago
Passive investing is not an issue, but the default bias towards large cap equities like SP500, Nasdaq100. Passive investing through total market ETFs (like VTI) maintains the status quo.

For example, if they are only two companies, say with 1T and 4T market cap. If one invests 5M into a total market ETF, 1M is allocated to company A and 4M to company B. But since company B is 4x bigger than company A, the upward price pressure is the same for both companies.

Comment · 1 year ago
Yes, the renewals are included
Comment · 1 year ago
curious how much of the growth is solely due to renewals since the cap on first time H1-Bs has stayed the same at 85k per year.

the green card queue is severely backlogged for India (and to some extent China, Mexico). this causes people who would usually get a green card after 3-5 years be on a continuous H1-B renewal cycle every 3 years.

Comment · 1 year ago
This is normal for most state border in India. They mostly inspect commercial trucks
Comment · 1 year ago
yes, it was in the 90s last August.
Comment · 1 year ago
Deepseek v3 needs 16 H100s or 8 H200 GPUs for inference.
Comment · 1 year ago
it is 23% of total energy. if you consider only acceleration energy, it will be much larger percentage, probably 80% (as in electric cars).
Comment · 1 year ago
you can add an '/s' at the end next time to avoid ambiguity. :)
Comment · 1 year ago
this is a well-studied tradeoff with immigration. high-skilled immigration helps US economy and innovation but low-skilled immigration depresses wages for citizens. you are making this distinction for a narrow case i.e. tech.
Comment · 1 year ago
yes

Live from the official Hacker News API

A modern reader for Hacker News · SvelteKit + Skeleton