Sichu Lu
@sichulu.bsky.social
e-tsundoku, supplementary info: nlab fan account, arxiv surveyor, pubmed enjoyer, two culture bridger, vacuous high gossiper, dearth of any domain expertise, reluctant g theorist, gpu poor
Pinned
Sichu Lu
@sichulu.bsky.social
· Oct 19
pgp fingerprint: EC170B8FB69F48E6C14B6BF95BE5E2C1CD617547
this is the funny thing about autism, in theory you are doing "cache to cache" communication, in practice autism narrows the contextual information being conveyed by physicality clues to nearly nothing and impedes social information
It is in principle possible to dump deep chunks of context/understanding on each other, much much bigger chunks than we actually process from each other in practice at a rate much faster than we do. If this was actually a thing we could do our current social organization wouldn't exist.
November 26, 2025 at 3:19 PM
this is the funny thing about autism, in theory you are doing "cache to cache" communication, in practice autism narrows the contextual information being conveyed by physicality clues to nearly nothing and impedes social information
November 26, 2025 at 3:17 PM
me when there is actual grammar and punctuation and commas and proper spelling in my texts
"This article is spiritually ChatGPT."
November 26, 2025 at 3:11 PM
me when there is actual grammar and punctuation and commas and proper spelling in my texts
the rate of activity on bsky keeps getting higher and higher and even with notifications i rarely actually manually check this website i just read the funny out of context notifications and giggle to myself
November 26, 2025 at 5:03 AM
the rate of activity on bsky keeps getting higher and higher and even with notifications i rarely actually manually check this website i just read the funny out of context notifications and giggle to myself
Reposted by Sichu Lu
Are LLMs Gloopy or Freepy?
November 25, 2025 at 9:08 PM
Are LLMs Gloopy or Freepy?
Reposted by Sichu Lu
We benchmarked Opus 4.5 on FrontierMath. It scored 21% on FrontierMath Tiers 1–3, continuing a trend of improvement for Anthropic models.
This score is behind Gemini 3 Pro and GPT-5.1 (high) while being on par with earlier frontier models like o3 (high) and Grok 4.
This score is behind Gemini 3 Pro and GPT-5.1 (high) while being on par with earlier frontier models like o3 (high) and Grok 4.
November 25, 2025 at 9:26 PM
We benchmarked Opus 4.5 on FrontierMath. It scored 21% on FrontierMath Tiers 1–3, continuing a trend of improvement for Anthropic models.
This score is behind Gemini 3 Pro and GPT-5.1 (high) while being on par with earlier frontier models like o3 (high) and Grok 4.
This score is behind Gemini 3 Pro and GPT-5.1 (high) while being on par with earlier frontier models like o3 (high) and Grok 4.
Imageedit @norvid-studies.bsky.social
November 25, 2025 at 2:14 PM
Imageedit @norvid-studies.bsky.social
yeah i don't think this is universally true but learning to read mathematicians is probably a bigger chunk of learning math than most of them would like to admit
I think what I'm beginning to realize is that it's not so much that I hate math as that I hate mathematicians. Single letter variable names, untyped notation, syntax based mostly on whoever defined the operations first, culture of "fuck you read the manual/prove it yourself". Status driven shitshow.
November 23, 2025 at 4:13 PM
yeah i don't think this is universally true but learning to read mathematicians is probably a bigger chunk of learning math than most of them would like to admit
Reposted by Sichu Lu
I've been dreaming of a CAS that would burn their clubhouse down since my freshman year of high school. Thank god it's finally happening, GPT goes so much farther than I ever dared to dream, than Alan Kay and Doug Engelbert and Ted Nelson ever dared to dream. Shocking how many can't see it.
November 23, 2025 at 7:28 AM
I've been dreaming of a CAS that would burn their clubhouse down since my freshman year of high school. Thank god it's finally happening, GPT goes so much farther than I ever dared to dream, than Alan Kay and Doug Engelbert and Ted Nelson ever dared to dream. Shocking how many can't see it.
wait who would turn down making everyone(including themselves hotter?)
universal basic hotness would be more controversial if it were technically feasible than universal basic income
November 23, 2025 at 4:11 PM
wait who would turn down making everyone(including themselves hotter?)
what's your average p-set tolerance rate limit
I can currently tolerate two exercises from this book per session before I get too fed up to continue. Let's see if I can push that to three by banging my head against the wall. :p
November 23, 2025 at 4:10 PM
what's your average p-set tolerance rate limit
"everything is a premature optimization problem"
“Everything is an optimization problem” seems true, but I worry it is less likely to produce good intellectual character (? 😂 ?) than “everything is about avoiding overfitting.”
November 23, 2025 at 4:10 PM
"everything is a premature optimization problem"
Reposted by Sichu Lu
hellthreads, colorized
November 23, 2025 at 1:00 AM
hellthreads, colorized
Reposted by Sichu Lu
November 22, 2025 at 6:45 PM
If Socrates was not killed would we even remember him? Maybe the greatest thing Athens did for philosophy was making Socrates drink helmlock
November 23, 2025 at 12:50 AM
If Socrates was not killed would we even remember him? Maybe the greatest thing Athens did for philosophy was making Socrates drink helmlock
Reposted by Sichu Lu
gm or whatever
November 22, 2025 at 6:18 PM
gm or whatever
Reposted by Sichu Lu
Going through my following demanding they provide me justified true knowledge lest I unfollow them
November 22, 2025 at 4:23 PM
Going through my following demanding they provide me justified true knowledge lest I unfollow them
Reposted by Sichu Lu
going through my follower list 1 at a time and demanding to know who they are and why they're "following" my account
November 22, 2025 at 4:13 PM
going through my follower list 1 at a time and demanding to know who they are and why they're "following" my account
Reposted by Sichu Lu
Non-convex loss landscapes have "horns" and as we all know horns are the devil, so if you aren't converging to a unique global minimum your immortal soul is at risk.
SVMs are the only moral type of machine learning, which is why they’re still taught in ML classes. Unlike next token prediction, they are inherently good
November 21, 2025 at 10:05 PM
Non-convex loss landscapes have "horns" and as we all know horns are the devil, so if you aren't converging to a unique global minimum your immortal soul is at risk.
Average post for a moderate sized account probably get more views than idk the average bill that dies in committee since no one else would even know it was proposed
I think it is actually quite common for popular online posters to have a bigger audience than an average member of Congress.
Nobody's saying online posters have more accountability than elected Democrats. It's that they have *some* accountability, i.e., greater than zero, which adds up over many accounts, and the problem is that accounts with large followings like Internet Hippo seem to believe it is zero.
November 21, 2025 at 6:43 PM
Average post for a moderate sized account probably get more views than idk the average bill that dies in committee since no one else would even know it was proposed
Reposted by Sichu Lu
ha ha ha yes!! (reposted on a different app) (thanks @medjed.bsky.social )
November 21, 2025 at 4:47 PM
ha ha ha yes!! (reposted on a different app) (thanks @medjed.bsky.social )
Reposted by Sichu Lu
they both seem incredibly OCEAN disagreeable so sometimes the narcissism of small differences keeps the score
November 20, 2025 at 4:28 PM
they both seem incredibly OCEAN disagreeable so sometimes the narcissism of small differences keeps the score