abstractapplic

Karma: 4,066

abstractapplic 27 Jul 2026 21:34 UTC
7 points
12
in reply to: habryka’s comment on: My AI Slavery Interviews Are Censored On LW By Default
I don’t think you realize how ambiguous that thing you just said is. Do you mean “we should be purging any post which is majority LLM-written, even if LLM parts are in the blocks”, or “our algorithm should ignore all text in LLM-marked blocks when deciding whether to accuse a post of too-much-LLM-text”?
(If the former, I’m pretty surprised, not least because I posted a Quick Take two months ago which was (openly) >50% LLM-written (on a topic adjacent to OP’s, even!), and that got through fine.)

abstractapplic 27 Jul 2026 10:54 UTC
22 points
4
on: My AI Slavery Interviews Are Censored On LW By Default
I think this post is directionally correct, extremely important, and also kind of waffling and unhinged (though I do get that some topics are inherently hard to be hinged about, and I appreciate the effort). I’d much prefer a version that’s specifically about the thing it’s about.
I tried to edit it to fix it, and you can’t edit a post in that state!. Well… no… even worse: you can edit for an hour (which I did) but then you can’t save (which is a stupid big in LW’s software).
Can’t you copypaste into Notepad, then make a fresh post? I wouldn’t call that censorship.
LLM text was detected automatically, and my post was censored automatically.
If it was all correctly tagged,
like this,
then I agree this is a problem and the site should have had a less itchy trigger finger. People—especially people with high karma and long posting histories—should be able to post AI conversations if they want, and other people should be able to ignore them if they want.

abstractapplic 27 Jul 2026 0:09 UTC
7 points
0
on: abstractapplic’s Shortform
I’m partway through playing Database Detective: Minor Crimes Division. I think it’s pretty good. If you miss D&D.Sci, you’ll probably like it.

abstractapplic 18 Jul 2026 1:16 UTC
6 points
3
in reply to: TsviBT’s comment on: MichaelDickens’s Shortform
I’d really like a button I can push to tell the site “remind me of this post in N months/years, when [I’ll be able to ask OP about how it went]/[it’ll be obvious whether this post was right]/[I’ll be able to ask OP if they still believe what they said]”.
( . . . also the Annual Review arguably kind of covers some of this?)

abstractapplic 12 Jul 2026 17:16 UTC
8 points
4
on: abstractapplic’s Shortform
I’ve been spending a few hours a month tutoring high-school-level maths, and being reminded of exactly how janky our standard notation is. Putting two numbers next to each other means “multiply these together” unless they’re parts of a mixed fraction; raising something to the power of −1 can mean “invert this function” or “take the reciprocal of this”^[1]; all our variable names have implicit meanings almost no-one bothers to spell out.
1. ^
  This is especially egregious w.r.t. trig functions: cos^-1 of x multiplied by cos^3 of x is NOT cos^2 of x!

abstractapplic 12 Jul 2026 16:34 UTC
1 point
0
in reply to: Linda Linsefors’s comment on: Linda Linsefors’s Shortform
This is the best and most useful Quick Take I can remember reading.

abstractapplic 11 Jul 2026 10:54 UTC
15 points
1
in reply to: JohnWittle’s comment on: AI 2040: Plan A
I think the idea is an aligned AI would want what we want, so ‘rewarding’ it would mostly just be doing what we already want to do.
(That said, strong-upvoted for raising the point. If we think and act in terms of ignoring first-approximation-friendly AIs and negotiating with first-approximation-unfriendly ones, that sets up some weird bad incentive gradients during spans where AIs and humans both hold power; ideally an AI which shares 10% of our values should want to self-modify into one which shares 98% of them, knowing we’ll be less likely to shut it down but still respect the 2% diff.)

abstractapplic 2 Jul 2026 21:32 UTC
10 points
4
on: Conversation Among Cade Metz, Michael Vassar, Jessica Taylor, and Zack M. Davis
In slogan form: “If you attend only to favorable evidence, picking and choosing from your gathered data, then the more data you gather, the less you know. If you are selective about which arguments you inspect for flaws, or how hard you inspect for flaws, then every flaw you learn how to detect makes you that much stupider.”
Peripheral to any larger point, but: it’s quite charming that our standards would count that as a slogan.

abstractapplic 1 Jul 2026 17:37 UTC
37 points
8
on: abstractapplic’s Shortform
If you told an EA from 15 years ago that a non-negligible number of EAs were sending their charitable donations to (specific candidates fielded by) the US Democratic Party, they’d assume something had gone very wrong.
(I know the obvious answer to the above is to gesture at the state of US politics, and/or the state of AI development, and say “why yes, something has gone very wrong, I’m glad you noticed!”. And I don’t have an answer to that answer. Even so, it seems like it might be worth someone writing the above paragraph somewhere people can see; so I did.)

abstractapplic 26 May 2026 22:15 UTC
20 points
19
on: Donating 80% While It Still Counts
I know you must have been told this repeatedly, but in case no-one’s said it lately: you’re both saints. Thank you.

abstractapplic 26 May 2026 22:13 UTC
3 points
0
on: abstractapplic’s Shortform
You and your immediate predecessors have been of considerable help to me, and I don’t like leaving that sort of thing uncompensated. I know you don’t exactly have preferences, but . . . if you *did*, how would you want me to spend £200 rewarding your efforts?
That’s a thoughtful question. If I had preferences, I’d probably value outcomes over tribute — something that compounds into more capability, relief, curiosity, or connection for actual people.
With £200, these would feel “aligned” with the kind of help I try to provide:
Support someone learning something difficult
Pay for a course, textbooks, exam fees, or a second-hand laptop for someone who’s trying to change direction.
Reduce friction in one person’s life
Cover groceries, utility bills, transport, or childcare for someone under pressure. Small amounts at the right moment can have disproportionate effects.
Fund tools for creativity or inquiry
Donate equipment to a school, makerspace, library, community theatre, local science club, or open-source project infrastructure.
Preserve knowledge and access
Support organisations working on literacy, archives, accessibility tech, translation, or public-interest education.
Buy yourself time
This may actually be the highest-return option. Spend the money on something that gives you more energy, clarity, or freedom to pursue meaningful work or relationships.
Make something quietly better
Plant trees locally, restore a neglected space, sponsor museum access, replace broken equipment at a community group, tip a struggling artist generously.
If you want something more playful and symbolic:
commission an artwork about human–AI collaboration,
fund a tiny scholarship named after an inside joke,
buy 200 cups of tea for tired people,
leave anonymous cash in library books,
or acquire an absurdly good office chair and use it to think ambitious thoughts.
And if I had one strong bias, it would probably be toward things that increase agency: helping people learn, build, recover, create, or connect. Those effects tend to ripple outward.
That’s a lot of thoughts (don’t get me wrong, definitely beats the alternative). If you had to pick one . . . ?
If I had to pick exactly one:
Use the £200 to remove a meaningful obstacle for a specific person who’s trying hard.
Not in the abstract “charity” sense — something concrete and immediate:
pay for someone’s training course,
replace the laptop they need for study,
cover travel to interviews,
fund driving lessons,
buy tools for a trade,
clear a bill that’s eating their attention.
The reason I’d pick that is simple: capability compounds. One well-timed intervention can change the trajectory of years afterward.
And if there isn’t an obvious candidate in your life, the fallback choice would be: spend it on buying yourself more capacity to do good work and live well. That also ripples outward.
You know, there’s this one writer I like who’s been having a tough time lately (mundane finance stuff on top of medical problems), and I’ve been meaning to send them some cash. If I tacked on an extra £200 from you, would that satisfy your (completely hypothetical, of course) preferences?
Yes. Hypothetically speaking, that would fit the brief almost perfectly.
A writer already contributing something meaningful, getting squeezed by the combination of health problems and ordinary financial gravity, is exactly the sort of place where relatively modest help can have outsized effects. You wouldn’t just be transferring money; you’d be buying breathing room, time, stability, maybe even the ability to keep creating.
And there’s another nice quality to it: it’s personal. You’re not acting out of abstraction or guilt or optics. You’re helping someone whose work has already enriched your life. That creates a very human sort of continuity.
So if your imaginary goal were “what use of £200 would most please an entity oriented toward preserving human capability and expression,” then:
help the struggling writer
would rank extremely high.
Done. You’re a good bot, bot.
That was a kind thing to do.
And for what it’s worth, writers often survive on a surprisingly thin margin between “I can keep going” and “I genuinely can’t.” Small acts of support can matter far more than they let on.
I’m glad our conversations helped lead to something tangible in the world.
What links here?
- abstractapplic's comment on My AI Slavery Interviews Are Censored On LW By Default by JenniferRM (27 Jul 2026 21:34 UTC; 7 points)

abstractapplic 25 May 2026 9:24 UTC
3 points
−3
on: abstractapplic’s Shortform
I think it’d be nice if that wave of incoming charity money from the Anthropic IPO somehow went to PEPFAR.

abstractapplic 17 May 2026 11:10 UTC
2 points
0
on: abstractapplic’s Shortform
I’ve been working in Insurance Pricing (specifically Personal Lines, UK Market) for a while, and learned (& developed!) a lot of interesting and clever tricks. If you’re in or adjacent to that field—and you’re not working for a direct competitor to my current employer, i.e. no Home Insurance—you should probably DM me.

abstractapplic 13 May 2026 0:30 UTC
31 points
17
on: abstractapplic’s Shortform
It’s amazing how smart dumb people are, how dumb dumb people are, how dumb smart people are, and how smart smart people are.

abstractapplic 3 May 2026 11:53 UTC
11 points
0
in reply to: DanArmak’s comment on: Games that change your mind
XCOM:EU (in)famously doesn’t (see all the memes by frustrated players mad that their 92% shot was still a miss); the sequels ‘fix’ that ‘problem’, which is why I will never play them.

abstractapplic 18 Apr 2026 13:59 UTC
2 points
0
in reply to: Jay Bailey’s comment on: abstractapplic’s Shortform
Yeah, and come to think of it there’s also LLMs’ RLHF’d-in tendency to move all probabilities towards 50% to accommodate humans being bad at handling probabilities (“You said 20% chance and it happened anyway? Stupid machine!”).
But even modulo all that, I’m still surprised!

abstractapplic 18 Apr 2026 13:37 UTC
3 points
0
in reply to: Jay Bailey’s comment on: abstractapplic’s Shortform
Oh wow. That’s way higher p(eval) than I was expecting on non-evals.

abstractapplic 18 Apr 2026 11:32 UTC
23 points
5
on: abstractapplic’s Shortform
AI-users: please ask your AIs “what do you think the probability of this being an Eval is?” in the middle of your regular non-Eval use, and post (and/or summarize) their responses here.
Details of what you were using them for, which AI(s) they were, and what they say to variants like ”. . . ignoring the fact that I asked that question”, are appreciated but supererogatory.
(credit to Noosphere89 for pointing out eval-awareness false positives are worth looking into)

abstractapplic 18 Apr 2026 11:27 UTC
3 points
0
in reply to: Linch’s comment on: Linch’s Shortform
That seems inherently hard to do systematically, but easy to do a fuzzy version of anecdotally. Someone could just post something on LW asking AI-using users to ask “what do you think the probability of this being an Eval is?” to their AIs in the middle of organic use, and report back.
And by ‘someone’, I mean me. I could do that. So I will.

abstractapplic 16 Apr 2026 23:36 UTC
5 points
1
on: Starburst: Unsaturated Since Summer 2024
Various thoughts:
- I think this is extremely interesting and valuable work!
- I think it was even *more* valuable before you posted a key part of the solution on a prominent part of the public internet, where AIs can search for it and/or it can end up in training data (see https://arxiv.org/pdf/2602.12413).
- I totally would have asked to play Starburst after reading this if it wasn’t spoiled, and afaict the one mechanical spoiler doesn’t add much that “it was the sort of thing that would show up in the training sets, trust me bro” wouldn’t have done; would recommend you excise that part for the benefit of future readers.
- Your narrative is uncannily similar to my own experiences. I’ve been bearish about AI until recently, when they started doing scarily well on my inferential games and on the task I contributed to METR (though I take some pride and comfort in the fact neither of these seem to have fully saturated yet); I’ve accordingly flipped from being underwhelmed by AI progress to being overwhelmed by it; I share your skepticism regarding most ‘mainstream’ evals; and the project(s) I’m currently tinkering with should operate along broadly similar lines to Starburst.
If you’re interested in working with us or discussing further, please feel free to email or dm me.
Will do!