back
user profile

dijksterhuis

4,117karma·1,383submissions·March 13, 2019
about
data/ops/infra/foss/distributed systems/rust

sometimes makes weird machine music. sometimes makes machines do weird things. sometimes makes music machines do weird things.

recent activity (1,383 total)
comment
adversarial examples , or test-time attacks, was a whole field of machine learning security way before LLMs came around. give the model a specially crafted bad input at inference time so attacker can…
2mo ago·view thread
comment
the main reason these images turn up is because theyre in the training data. and the images are common enough in the training data for the content to come out without being explicitly asked for (in …
2mo ago·view thread
comment
it’s not a problem that came into existence a few years ago. we’ve known about these sorts of test time attacks for decades now. prompt injection is just the LLM variant where people use less math to …
2mo ago·view thread
comment
> The missing image was described as "graphic" or "violent." not in the first prompt. which kicked the whole thing off. no mention of type of content was provided. the model gen…
2mo ago·view thread
comment
normal y = f(x) prompt injection / adversarial example (same thing really) bad_y = f(x+badness) tweak badness enough you will get bad outputs. no matter the defences. the only w…
2mo ago·view thread
comment
> I do wonder why openai didn't screen obvious gore from the training set of a general purpose model more expensive / would take longer / didn’t care / line must go up / we…
2mo ago·view thread
comment
non-coding manager spotted
2mo ago·view thread
comment
> dike-ster-house yep, basically. 3rd one down https://www.howtopronounce.com/dijksterhuis the story behind it: i was looking for a name to write music under. and i was reading aca…
2mo ago·view thread
comment
one would also hope there'll be less pressure to "make line go up", i.e. not having to do attention-engineering via deliberate sycophancy to trap individuals into using it more and more…
2mo ago·view thread
comment
ASML wiki for those like me who were confused https://en.wikipedia.org/wiki/ASML (despite the name, i am not dutch)…
2mo ago·view thread
comment
giving a shit == having taste. giving a shit != perfect. i recently had a PR which had a comment explaining a change of an import: "// Changed imports to add Foo as it's needed fo…
2mo ago·view thread
comment
were you under some sort of impression that ad-tech surveillance platforms (“social media”) weren’t already fingerprinting the majority of their users? like, the facebook pixel used to track you acros…
2mo ago·view thread
comment
> Yes, it removes the "free" internet as it was 20 years ago, social media is not, and never was the "free internet" we all get nostalgic about. maybe the first couple of years …
2mo ago·view thread
comment
just one more model. just one more prompt. just one more guardrail. just one more. just. one. more.
2mo ago·view thread
comment
i had a boss. before he was my boss, he was a friend. he took me under his wing, musically speaking. he showed me new music. told me what gear he was interested in. we went to some gigs. he used to sa…
2mo ago·view thread
comment
- minus symbol for negative
2mo ago·view thread
comment
long, low: copypasta; rant; sales; slop. (brevity, purposeful /s).
2mo ago·view thread
comment
Did you see the part in my big story time comment where i convinced the ceo his business strategy was wrong? do you think i did that with a flippant and dismissive comment? https://news.y…
2mo ago·view thread