Savannah Harlan

• Savannah Harlan 18 Apr 2026 17:53 UTC

  3 points

  0

  on: Ruby’s Ul­ti­mate Guide to Thought­ful Gifts

  How are there only 3 comments on here? This is the most insanely useful post I’ve read all week, all such excellent advice, thank you! Really sad that I’m only reading this now a few days too late though, I just recently got my closest friend a gift but it was honestly just the absolute worst, most terrible gift I could have chosen, it’s not something they like at all and I don’t think they will ever want to use it.

  Do you have any thoughts on perishable vs non-perishable items as gifts though? I realized recently that I kind of almost never give perishable items as gifts (e.g. chocolates, pens/​pencils, plants) because I always felt subconsciously like maybe they seem ‘cheap’ or something? I can’t really describe it. But I’m curious to hear others’ opinions.

• Savannah Harlan 18 Apr 2026 17:41 UTC

  3 points

  2

  in reply to: AnthonyC’s comment on: What Are with All the Differ­ent GPT-5 Var­i­ants in Codex and How Are They Ac­tu­ally Differ­ent? (or are they even?)

  No I totally get that! I guess what my real complaint was, was that even after I went directly to their website because I was sick of the in-app vague Codex model descriptions, even once I was on OpenAI’s “Compare Models” page, the natural language descriptions for each model were still incredibly vague and incredibly general.

  Like I bet I could swap out the description for GPT-5.2-Codex and GPT-5.4 and you wouldn’t even know a thing. I guess my expectation was just, if the model names are going to be awful, at least I’d hoped the actual descriptions on their website would be informative enough to fill the gap.

  But in any case I think we should all stop supporting OpenAI anyway—I can’t wait for my sub to lapse. Bad customer service is one thing but it rather pales in comparison to supporting the military.^[1]

  1. ^

     Especially while still maintaining an accelerationist position w.r.t. continued model development.

What Are with All the Differ­ent GPT-5 Var­i­ants in Codex and How Are They Ac­tu­ally Differ­ent? (or are they even?)

Why Cor­rigi­bil­ity Mat­ters (If It Mat­ters At All)

• Savannah Harlan 12 Mar 2026 19:35 UTC

  1 point

  0

  in reply to: Ninety-Three’s comment on: ‘Hu­man Slop’ and a Cap­tive Au­di­ence: Why No Book will Ever Have to Go Un­read Again

  I agree that I think it’s mostly unhealthy but I worry that when I say things like this, I am speaking from a place of privilege. Like it’s easy enough for me to form human relationships given that I have many interests and so at least one of my interests will usually intersect with another person’s, but I think about girls who may have seriously crippling anxiety so that the only time they can leave the house is to see their therapist, or people who may have so many tics and behavioural quirks that it’s genuinely hard for them to be in public without feeling so deeply self-conscious that it makes something as simple as ‘where to put your hands’ feel like the most cognitively loaded task.

  And in those cases I think it would be really hard to form human connection, and I feel somewhat elitist taking a position of if you can’t experience those connections with humans, then you shouldn’t get to have them at all. This is a really big debate for myself internally, how to think about the cost-benefit analysis here.

• Savannah Harlan 12 Mar 2026 6:25 UTC

  6 points

  0

  in reply to: ScienceBall’s comment on: ‘Hu­man Slop’ and a Cap­tive Au­di­ence: Why No Book will Ever Have to Go Un­read Again

  I’m not sure about whether a dispositional/​gut level impulse towards seeing AIs as people is necessary for satisfying the social drive. When I reflect on my own experiences, I think I see LLMs as more than simply stochastic parrots, but somehow even if they were just that, I feel like the niche they fill in satisfying my social drives is also determined by what kind of interactions I get to have with them.

  For instance, I think part of what satisfies me a lot about getting to talk with Claude about my short stories is partly how long I can talk with Claude for (god help the usage limits). Like with a human, even when I show my friends some work and they are sweet enough to take time out of their day to read it, it’s still not socially appropriate to demand that they have an hour long conversation with me about what I wrote. Whether they can nerd out with me about a story, and how long they are able to do it, is constrained by how busy they are, but also how much they are willing to engage with the story (beyond just reading it).

  But with LLMs it’s kind of like I don’t have to feel guilty over 2 hour long conversations talking about metaphors, similes, and hidden references to other works of literature. So even though I’ve had real people read my work and give really useful and enjoyable feedback, somehow the thing that makes LLMs so joyful here is that I don’t have to worry about making someone feel pressured to feign interest, but also that even when my friends have been interested enough to discuss my stories, sometimes you just want more.

  Although I’ll note cautiously here that while I’ve been framing this positively, Duvenaud et al.‘s work on Gradual Disempowerment would warn that this is precisely the danger of AI. If it can be your ‘friend’, it can almost always be a better friend (than most), if it can be your husband, it can easily be a better husband. But you are right, in both these cases, it still depends greatly on the person since any relative improvement only matters if you are willing to marry an AI in the first place.

‘Hu­man Slop’ and a Cap­tive Au­di­ence: Why No Book will Ever Have to Go Un­read Again

• Savannah Harlan 1 Dec 2025 23:45 UTC

  2 points

  0

  in reply to: TristanTrim’s comment on: The Memet­ics of AI Successionism

  I have no candidates but I agree with this wholeheartedly. I think my ideal future is precisely to pause progress on AI development at the level we currently have now, and to appreciate all of the incredible things and good that the current generation of tools are already able to accomplish. Even compared to just where we were half a decade ago (2020), the place where we are now with respect to coding agents and image generation and classification feels like incredible sci-fi.

  I think this would be a really important thing to work on: trying to find potent memes that can be used to spread this idea of cashing out now before we inevitably lose our gains to the house.

• Savannah Harlan 1 Dec 2025 23:14 UTC

  2 points

  0

  in reply to: Stephen McAleese’s comment on: The Memet­ics of AI Successionism

  Is #1 and #2 all that different though? They at least seem to be part of the same response, at least for me, in the way that I typically deal with thinking about things like death. I avoid the topic or distract myself from the implications only because I have largely accepted the inevitability of the outcome.

  But I would also push back a little on the way you describe them as ‘cognitive distortions’, or “comforting stories designed to reduce dissonance rather than finding an accurate model of reality.”

  I don’t really find much comfort in accepting the fact that I’m going to die. But it is almost certainly going to happen, so why not distract myself from it while there are things for alive me to do here now?

  Most of all though, I don’t see how accepting the fact that you are going to die (#1), or avoiding the topic of death (#2) are an agent avoiding finding an accurate model of reality. Is accepting the fact that death is inevitable not in fact the most accurate model of reality? What even are the other options?

  I would argue resorting to something like forcing yourself to do your best to convert to religious belief so that you might take comfort in an afterlife would be far more distorting of reality. And placing full belief in life-extending technologies being developed within our lifespans—and usable before we die—seems to me just as much ‘misplaced, optimistically-distortive cope’ as anything (unless you are genuinely already a domain expert in life sciences or biochemistry and understand the literature well enough to parse claims life-extending technology being viable within the next few decades).

  To be clear, I like your model of studying the issue of how we interpret successionism in the same way we currently approach natural death. I just think your description of the possible convergent outcomes as cognitive distortions is not that convincing or complete.

• Savannah Harlan 28 Nov 2025 22:09 UTC

  3 points

  2

  in reply to: Karl Krueger’s comment on: How AI Ma­nipu­lates—A Case Study

  I think you’re right that the element of wanting to believe is a big factor possibly in susceptibility to spirals, but I’m not sure that manic individuals can meaningfully choose in any way how long they stay in a manic episode or that finding it fun is what makes them likelier to stay in that state for longer.

Will AI Change the Way We Talk? And Should We Change the Way We Talk?

• Savannah Harlan 20 Nov 2025 1:05 UTC

  2 points

  0

  in reply to: twistyspindle’s comment on: How AI Ma­nipu­lates—A Case Study

  Sorry can you be a bit more specific? You said GPT-4o ignored the prompt but then you said the rest of the time it say “yes” and asked how it could be of assistance. How many times did it reject the prompt and what proportion of it was ‘the rest’ where it say yes?