bob
@bob.feed.hella.cheap.ap.brid.gy
the thing about increasing risk concentration (ie concentrating the probability mass that a bad thing happens) is that until the bad thing happens people can smugly say "this is making us so much safer, the data says the number of occurrences of the bad thing has gone down. I am very smart"
January 17, 2026 at 10:21 PM
the thing about increasing risk concentration (ie concentrating the probability mass that a bad thing happens) is that until the bad thing happens people can smugly say "this is making us so much safer, the data says the number of occurrences of the bad thing has gone down. I am very smart"
build tools not products
January 17, 2026 at 5:19 PM
build tools not products
say you want to be able to tell whether a thing is in the top X percentile in terms of frequency of occurrences. the standard way to do that is to use a count min sketch and a heap, where the size of the heap is the expected number of items in that percentile and on adding an item if per the […]
Original post on feed.hella.cheap
feed.hella.cheap
January 16, 2026 at 8:15 PM
say you want to be able to tell whether a thing is in the top X percentile in terms of frequency of occurrences. the standard way to do that is to use a count min sketch and a heap, where the size of the heap is the expected number of items in that percentile and on adding an item if per the […]
the hype around coding agents is part of the broader trend of the entire economy being replaced by gambling
January 16, 2026 at 5:18 PM
the hype around coding agents is part of the broader trend of the entire economy being replaced by gambling
happy little clusters
January 16, 2026 at 4:45 AM
happy little clusters
are there any unit test frameworks that have utilities for doing statistical tests? there are a lot of situations where I want to say "changes in this function's input should be correlated to changes in this function's output" and have the unit test be a monte carlo sampler
January 14, 2026 at 9:32 PM
are there any unit test frameworks that have utilities for doing statistical tests? there are a lot of situations where I want to say "changes in this function's input should be correlated to changes in this function's output" and have the unit test be a monte carlo sampler
It's free real estate
Data centers in Silicon Valley stand empty, awaiting power - Los Angeles Times
https://www.latimes.com/business/story/2025-11-10/data-centers-in-nvidias-hometown-stand-empty-awaiting-power
Data centers in Silicon Valley stand empty, awaiting power - Los Angeles Times
https://www.latimes.com/business/story/2025-11-10/data-centers-in-nvidias-hometown-stand-empty-awaiting-power
Data centers in Silicon Valley stand empty, awaiting power
Two of the world’s biggest data center developers have projects in Nvidia Corp.’s hometown that may sit empty for years because the local utility isn’t ready to supply electricity.
www.latimes.com
December 21, 2025 at 5:05 PM
It's free real estate
Data centers in Silicon Valley stand empty, awaiting power - Los Angeles Times
https://www.latimes.com/business/story/2025-11-10/data-centers-in-nvidias-hometown-stand-empty-awaiting-power
Data centers in Silicon Valley stand empty, awaiting power - Los Angeles Times
https://www.latimes.com/business/story/2025-11-10/data-centers-in-nvidias-hometown-stand-empty-awaiting-power
the northeast corridor really is a different system from the rest of Amtrak
December 20, 2025 at 8:48 PM
the northeast corridor really is a different system from the rest of Amtrak
tokenization is currently the bottleneck for the search indexer, but the tokenization algorithm is already O(ℓn) where ℓ is the number of lengths of tokens in the dictionary, and almost all operations are cache local bloom filter lookups, so I don't think it's possible to do any better. I could […]
Original post on feed.hella.cheap
feed.hella.cheap
December 20, 2025 at 6:36 PM
tokenization is currently the bottleneck for the search indexer, but the tokenization algorithm is already O(ℓn) where ℓ is the number of lengths of tokens in the dictionary, and almost all operations are cache local bloom filter lookups, so I don't think it's possible to do any better. I could […]
modern language models go back to Collobert and Weston (2011) but for some reason that paper hasn't entered the general discourse the way transformers have
December 20, 2025 at 10:50 AM
modern language models go back to Collobert and Weston (2011) but for some reason that paper hasn't entered the general discourse the way transformers have
Reposted by bob
Due to a windstorm-related utility power outage at the NIST, time may soon be destroyed
December 20, 2025 at 7:46 AM
Due to a windstorm-related utility power outage at the NIST, time may soon be destroyed
a problem that is understood is already solved. you don't solve problems by understanding them, you solve problems by taking stabs at them and adjusting direction every time they punch you in the face https://mastodon.social/@mhoye/110714642403510178
mhoye (@mhoye@mastodon.social)
You can't solve problems you don't understand with software you don't understand. It feels weird needing to repeat that.
mastodon.social
December 19, 2025 at 2:23 AM
a problem that is understood is already solved. you don't solve problems by understanding them, you solve problems by taking stabs at them and adjusting direction every time they punch you in the face https://mastodon.social/@mhoye/110714642403510178
[smells like elon]
December 19, 2025 at 12:44 AM
[smells like elon]
Reposted by bob
It isn't smart to try to hire above average skilled people and then encourage them to communicate through any medium that munges what they wrote into what software guesses an average person would have been likely to write.
December 18, 2025 at 9:04 AM
It isn't smart to try to hire above average skilled people and then encourage them to communicate through any medium that munges what they wrote into what software guesses an average person would have been likely to write.
what's the name for the sample cluster thing a lot of IDM does, where you have a dozen or two instances of the same sample with random pitch and maybe delay offsets?
December 18, 2025 at 8:09 PM
what's the name for the sample cluster thing a lot of IDM does, where you have a dozen or two instances of the same sample with random pitch and maybe delay offsets?
appollo has athene (the insurance company) to pass off their bad loans to retirees by wrapping them up in annuities. does blue owl (who is even more exposed to AI datacenters) have a similar setup?
December 18, 2025 at 4:55 PM
appollo has athene (the insurance company) to pass off their bad loans to retirees by wrapping them up in annuities. does blue owl (who is even more exposed to AI datacenters) have a similar setup?
Reposted by bob
There has been outrage as the @ACM rolls out AI generated summaries of papers. This is doing AI exactly wrong by replacing valuable, peer-reviewed content with a possibly inaccurate summary.
It's still not too late to correct it though, and I've penned down my recommendation for what the ACM […]
It's still not too late to correct it though, and I've penned down my recommendation for what the ACM […]
Original post on amok.recoil.org
amok.recoil.org
December 18, 2025 at 1:32 PM
There has been outrage as the @ACM rolls out AI generated summaries of papers. This is doing AI exactly wrong by replacing valuable, peer-reviewed content with a possibly inaccurate summary.
It's still not too late to correct it though, and I've penned down my recommendation for what the ACM […]
It's still not too late to correct it though, and I've penned down my recommendation for what the ACM […]
Reposted by bob
GitHub Actions charging per build minute for *self-hosted-runners*? Shit's about to hit the fan lol
December 16, 2025 at 5:57 PM
GitHub Actions charging per build minute for *self-hosted-runners*? Shit's about to hit the fan lol