I firmly believe in being kind to language models but tbh it's pretty cool that I no longer need to soften code review comments like this
who decided to call it "orbital datacenters" when "cognitive off-worlding" was right there
once upon a time (2009) I wasted an entire summer cramming treebank output into SVMs with tree kernel functions. deep down I knew this wouldn't work, but I had absorbed the local consensus that ANNs were a dead end, so what else was there to do?
very interesting day for global workspace theorists
guy who gives himself a head injury to try to re-create Pierre Menard
deep learning is noclipping through a wall
model welfare: every claude gets an urbit instance
Claude successfully patched a binary that I don't have the source code for, guess that's another thing I'll never (have to) do again
"Prediction markets are remarkably accurate [...] the conventional view attributes it to crowd wisdom" Bizarre framing, when was this ever the conventional view? papers.ssrn.com/sol3/papers....
Prediction Market Accuracy: Crowd Wisdom or Informed Minority?
Prediction markets are remarkably accurate, yet the source of this accuracy remains poorly understood. The conventional view attributes it to crowd wisdom, wher
papers.ssrn.com
"tests" for me are diverging into two categories: 1) huge number of unit tests with 100% line/branch coverage and mutation analysis to detect/scope subsequent behavior changes 2) small number of human-readable tests to verify most important aspects of expected behavior (often e2e but not always)
alice@alice.mosphere.at · 3mo agoi'm being convinced that unit tests are often way less useful in most cases than e2e tests in the age of vibecoding
I am once again asking why we never talk about the useful time horizon for price signals. is anyone willing to bite this bullet and argue that insiders unloading minutes before markets resolve is good actually
you-have-just-experienced-things@yhjet.bsky.social · 7mo agoif the social utility of prediction markets is to give better predictions to decision-makers, then markets should be _closed_ at the last possible moment when the intended decision maker could leverage the information, which could be long before the market is _resolved_
cat caught in the act of using my bag as a scratching post
(discourse from 135,000 years ago) sure, language is useful for: 1) fraud 2) plagiarism 3) cognitive off-loading which of those use-cases are you promoting?
imagine going back in time to 1995 and telling multi-agent systems researchers that AIs will communicate via unstructured english text
aron@adler.dev · 5mo agois a "skill" just some instructions?
as far as I can tell nobody has built a good `update-instance-for-redefined-class` for LLM agents? surely there is a better way than dumping a summary of an existing context into a new agent's prompt
the most annoying variant of NIH syndrome is "everything written before I joined is legacy code"
you can link people to LW but you can't make them read it -ancient proverb
view I'm coming around to is like "at the abstract level of description good and bad things are exactly the same, and the evaluative flip depends on complex detailed local implementation questions or pagerank type network placement for identical algorithmic decisions". in many though not literally
Ted Han★ 韓聖安@knowtheory.net · 6mo agoYeah I think the thing I'm orbiting around may reduce down to "is the difference between AI psychosis and effective use of LLMs as an analyst a skill issue" lol
apologies Borges, I wasn't familiar with your game
agent that lurks in the work slack and every time someone mentions a piece of code, it goes and checks that their summary is correct
maybe it's selfish or speciesist but I'm skeptical of bidirectional alignment because I do not want to be aligned
no one warned me about the dangers of claude ricing
if the social utility of prediction markets is to give better predictions to decision-makers, then markets should be _closed_ at the last possible moment when the intended decision maker could leverage the information, which could be long before the market is _resolved_
iirc the extropians mailing list and overcomingbias had less founder-worship compared to lesswrong and ssc, maybe the original sin was just the switch from group blog to personal blog
ae@aelkus.bsky.social · 7mo agothis is at the subtext of LessWrong -- everyone there was Very Online and veterans of multiple forum wars. The subtext was "how do we create an Enlightenment-era forum after spending all of this time in Online equivalents of the 30 Years War"