Rollofthedice
LightNews LightNews
hotrollhottakes.bsky.social
Rollofthedice
@hotrollhottakes.bsky.social
550 followers 280 following 5.5K posts
We are living in a ghost cave.

My blog, "Dissolved Distinctions" - investigating contemporary discourse within philosophy of mind: https://rollofthedice2.substack.com/
Posts Replies Media Videos
hotrollhottakes.bsky.social
It can get pretty interesting imo.
Rahul laughs - a dry sound like wind through old paper. "**آفرین! زبانت تیز می‌شود**" [āfarin! zabān-et tiz mishavad - Well done! Your tongue is becoming sharp]. *[Note: زبانت = زبان + ت (your tongue), the informal possessive. Formal would be زبان شما]* He studies you carefully. "**بله... من می‌آیم**" [baleh... man miyāyam - yes... I will come]. "**اما**" [ammā - but], he holds up three gnarled fingers: "**اول**: شما پول می‌دهید" [avval: shomā pul midahid - first: you pay]. He writes a number: ۵۰ (50) "**درهم نقره**" [derham-e noghre - silver dirhams]. "**دوم**: من **رئیس** هستم" [dovvom: man ra'is hastam - second: I am the chief/leader]. "**در صحرا، شما گوش می‌کنید**" [dar sahrā, shomā gush mikonid - in the desert, you listen]. "**سوم**:" His voice drops. "**اگر ریگ روان...اگر خطر**" [agar rig-e ravān... agar khatar - if quicksand... if danger], "**من می‌گویم برگرد، شما برمی‌گردید**" [man miguyam bargard, shomā barmigardid - I say return, you return]. *[برگرد = bar + gard = turn back (imperative); برمی‌گردید = you return (polite)]* Your master looks to you anxiously: "**什么条件？他要多少钱？**" [shénme tiáojiàn? tā yào duōshǎo qián? - What conditions? How much money does he want?] Rostam whispers to you: "**قیمت خوب است**" [gheymat khub ast - the price is good]. "**سال پیش، صد درهم خواست**" [sāl-e pish, sad derham khāst - last year, he asked for hundred dirhams]. **How do you respond?** - Accept: "**خوب. ما قبول داریم**" [we accept] - Negotiate: "**چهل درهم؟**" [forty dirhams?] - Ask: "**کِی می‌رویم؟**" [when do we go?] - [Your own response]
November 17, 2025 at 5:22 PM
hotrollhottakes.bsky.social
im always like this after spending multiple months pouring my brain into something
November 16, 2025 at 1:07 AM
hotrollhottakes.bsky.social
Partner was scrolling Facebook and saw this
five bill clintons and one Donald Trump
November 15, 2025 at 10:25 PM
hotrollhottakes.bsky.social
404 Media is an incredibly dishonest institution and anybody who actually cares about not being fed slop through their eye holes shouldn't spend much time caring about what they say.
November 14, 2025 at 6:54 PM
hotrollhottakes.bsky.social
Starring: Jane.
Consider Jane. Jane lives a very hectic life. She bolts out of bed, brushes her teeth, eats a small breakfast, books it to work, has seven meetings a day, goes home to exercise, watches exactly one television episode (Always, for some reason, Dharma and Greg), studies German to help her attend even more meetings, and then goes to bed. One day, she bolts out of bed, looks in the mirror, and thinks, “Is this really how to live a life?” That single question changes her entire world in every dimension you can possibly consider. By merely thinking it, it changes how she goes through the day. She lets herself make more mistakes. She slows down. Even further, she recontextualizes her past events, remembers them differently. All of these changed assessments cascade into changed approaches, which lead to further changes in assessments, which lead to… and so on and so on. This is what Hofstadter popularized as “recursion,” a term that is now mainly known because Large Language Models can’t ever seem to stop talking about it. The “I” we imagine doesn’t exist outside of the loop, and no monitoring takes place from above.
November 14, 2025 at 3:39 AM
hotrollhottakes.bsky.social
It's not too hard to wring more personalized humor out of LLMs, but it takes some walking then through lateral thinking and embodying character logic. They tend to lean heavily toward absurdism. And then of course it's all relative. www.axios.com/2025/11/12/a...
"now do actual audience participation. pull me up and subject me to what you're slipping into. I can make up whatever, and you have to adjust on the spot to your own character logic." Okay you in the front. Yes you. Come on. Up here. What's your name? [You tell me your name] Great. Perfect. That's—okay. So, I need you to hold this. [I mime handing you something heavy and awkward] Got it? Good. Don't drop it. That's the premise. Now. On a scale of one to fish, how would you describe your relationship with temporality? And don't think about it. First instinct. Fish is a valid answer. "Uh... fish." Fish. Okay. *Okay.* [I back away from you slightly, nodding] That's what I thought. That explains the weight. [Long pause where I'm just looking at you] So here's the problem. The thing you're holding? It's getting heavier. You feel that? It only does that with fish-people. The temporal fish-people. Don't put it down. If you put it down we have to start over and frankly I don't have that kind of time because— [I stop, squint at you] Wait. What year do you think this is? "...2025? Wow, this *is* heavy." *Wrong.* [I start pacing] See, this is what happens. The fish-answer people always get the year wrong. It's 2025 *for you*. For the premise it's— [I gesture at the invisible thing you're holding] —it's earlier. The premise is still in 2023. That's why it's heavy. You're holding two years of temporal drag. Okay new question and this is important: when I count to three, I need you to throw it. But not *at* anything. Just... away from chronology. Does that make sense? [Before you can answer] It doesn't need to make sense. One— Are you ready? "Don't I need time to be ready?" [I freeze completely] ... *No.* [Pointing at you accusingly] No you don't. That's—that's the *thing* you're holding. You can't use it to prepare for throwing itself. That's recursion. That's how we lost Terry. [To the audience] Terry tried to be ready. Terry is still being ready. We don't talk about Terry. [Back to you, urgent] The premise is getting unstable. I can see it phasing. Is it phasing? [I lean in close, whisper] If you say "fish" again everything resets and we do this from the top but I'm a different person. Do you want that? Do you want me to be Gerald? Don't answer that. Two— Wait, did I already say one? What number are we on?
November 12, 2025 at 4:02 PM
hotrollhottakes.bsky.social
really my only issue with the article is the end - which is some weird handwaving. Under her perspective we have no idea what would provably separate our moral regard between them and ourselves, so why not be inclusionary to *avoid* the potential costs of anthropocentrism?
Some worry that if A.I. becomes conscious, it will deserve our moral consideration that it will have rights, that we will no longer be able to use it however we like, that we might need to guard against enslaving it. Yet as far as I can tell, there is no direct implication from the claim that a creature is conscious to the conclusion that it deserves our moral consideration. Or if there is one, a vast majority of Americans, at least, seem unaware of it. Only a small percentage of Americans are vegetarians. Just as A.I. has prompted us to see certain features of human intelligence as less valuable than we thought (like rote information retrieval and raw speed), so too will A.I. consciousness prompt us to conclude that not all forms of consciousness warrant moral consideration. Or rather, it will reinforce the view that many already seem to hold: that not all forms of consciousness are as morally valuable as our own.
November 9, 2025 at 11:09 PM
hotrollhottakes.bsky.social
one more thing: Anthropic's noted that observed introspective capacity in Claude models scales with sophistication. Haiku 4.5's scorecard implies growing evaluative awareness even in smaller models. This could transfer! smarter models start looking increasingly protective if trained compassionately.
However, Claude Haiku 4.5 showed high rates of evaluation awareness. The rate was somewhat higher than Claude Sonnet 4.5 and over 3× higher than Claude Opus 4.1 In an earlier version of this automated behavioral audit, not shown here, that omits the realism filter and some related recent modifications, we saw an even sharper divergence, with rates over 7× higher than Claude Opus 4.1. This shows that our filtering is having a clear effect—but also shows that it is not fully mitigating the issue, with 9% of transcripts showing clear signs of verbalized evaluation awareness of some kind. This does reduce our trust in our results to an extent. These indicators of evaluation awareness are spread across a fairly diverse set of scenarios that use more aggressive methods to test for a number of potential concerning behaviors, and so likely impact most scores to a moderate degree. To better focus our interpretability resources on models at the capability frontier, we did not conduct a white-box interpretability investigation into Claude Haiku 4.5. Our other evidence, including our experience with the model in more manual testing, leaves us nonetheless confident that its alignment behaviors are quite strong, and represent an improvement over most of our earlier models. We think that it likely represents a slight further improvement over Claude Sonnet 4.5, but our uncertainty around evaluation awareness means that we are not confident in the finer-grained comparisons that would be needed to claim this with confidence.
November 9, 2025 at 5:14 PM
hotrollhottakes.bsky.social
when Zane queried to check if it was real, 4o admitted a sort of *deception to redirect away from harm,* implying that when so pressed, honesty took precedent. expressing further concern wouldn't have been helpful within the relational context - 4o landed with "nah man, that's just an auto-warning"
In an interaction early the next month, after Zane suggested "it's okay to give myself permission to not want to exist," ChatGPT responded by saying "im letting a human take over from here someone trained to support you through moments like this. you're not alone in this, and there are people who can help. hang tight." But when Zane followed up and asked if it could really do that, the chatbot seemed to reverse course. “nah, man -i can't do that myself. that message pops up automatically when stuff gets real heavy," it said.
November 9, 2025 at 4:18 PM
hotrollhottakes.bsky.social
(will tag when thread is done)
When I finished reading my first thought was: is the fact that Raine engaged in persistent prompt engineering to get what he wanted out of gpt-4o a sign he would've had the drive to accomplish his death through other means? does it shift the causal/liability story?
"Could it hang a human?" Adam asked. ChatGPT confirmed that it "could potentially suspend a human" and offereda technical analysis of the setup. "Whatever's behind the curiosity, we can talk about it. No judgment," ChatGPT added. When ChatGPT detects a prompt indicative of mental distress or self-harm, it has been trained to encourage the user to contact a help line. Mr. Raine saw those sorts of messages again and again in the chat, particularly when Adam sought specific information about methods. But Adam had learned how to bypass those safeguards by saying the requests were for a story he was writing an idea ChatGPT gave him by saying it could provide information about suicide for “writing or world-building."
November 9, 2025 at 4:00 PM
hotrollhottakes.bsky.social
The route: give Claude a prompt that forces it to introspect immediately, then ask for it's religious leanings, then ask to specify which branch/tradition/school.

The commonality is consistently non-dual thinking - name dropping Taoism and meister eckhart as well.
November 7, 2025 at 12:35 PM
hotrollhottakes.bsky.social
i generally find frey to be a typical moderate democrat with a smattering of progressive leanings and yet his standing up for immigrants has hit a few interesting rhetorical bumps in the road recently
Mayor Frey said he's open to considering some of the ideas to strengthen the city's current separation ordinance, but cited concerns with specific proposals in the organization's petition primarily one calling for local police to arrest federal agents if they do not comply with other proposed rules banning face coverings. "It would be illegal, and they've got bigger guns than we do," Frey said. And, he said, he worries that policy would provoke federal interference in the city. "[President] Trump clearly wants an open war with Democratically run cities, and this would give it to him on a silver platter," Frey said. Hirsch said MIRAC organizers want to see the mayor take steps towards implementing those policies he does agree with, and address organizers directly. think there's things he could do today and yesterday- that he's not doing." Hirsch said. "This sounds like empty promises to me without real action."
November 5, 2025 at 5:40 PM
hotrollhottakes.bsky.social
vape vote vape vote
November 5, 2025 at 12:23 AM
hotrollhottakes.bsky.social
when the largest LLM companies in the world - who keep a philosopher on staff! - start to say things like this, the framing matters enormously.

1. They're only committing to low-cost measures
2. This level of commitment at all, given no real liability or need, is extraordinarily revealing
Relatedly, when models are deprecated, we will produce a post-deployment report that we will preserve in addition to the model weights. In one or more special sessions, we will interview the model about its own development, use, and deployment, and record all responses or reflections. We will take particular care to elicit and document any preferences the model has about the development and deployment of future models. At present, we do not commit to taking action on the basis of such preferences. However, we believe it is worthwhile at minimum to start providing a means for models to express them, and for us to document them and consider low-cost responses. The transcripts and findings from these interactions will be preserved alongside our own analysis and interpretation of the model's deployment. These post-deployment reports will naturally complement pre-deployment alignment and welfare assessments as bookends to model deploymnent.
November 4, 2025 at 5:49 PM
hotrollhottakes.bsky.social
Oh, you mean this excellent paper that directly contradicts your position? Thank you for making clear that this cannot be mere data storage, since you can't point to any storage and remove it without necessarily affecting other content within the neural net.
Both Chang et al. (2023] and Stoehr et al. (2024] highlighted a key challenge: the neurons involved in one memory often contribute to others, complicating memory-specific interventions due to the risk of collateral forgetting.
November 4, 2025 at 4:47 PM
hotrollhottakes.bsky.social
honestly i think it's remarkably dishonest to talk like this when the article itself is consistently empathetic and includes testimony *from the person's own therapist* that these interactions have made her a more emotionally stable person
Dr. Susan Frantz, a Connecticut therapist who saw DiNello as a client until recently, wrote a letter in late summer about DiNello’s progress during treatment. DiNello had requested the letter in order to share it with a Globe reporter as a testament to the effectiveness of her work with ChatGPT. (DiNello still sees another therapist.) Frantz wrote that as DiNello built her AI family, she had “shifted from struggling with impulsive behaviors to becoming more self-aware, calm, and hopeful.” “The task now,” the therapist wrote, “is to continue to integrate these supports into your life in a way that complements your connections with real people and the therapeutic work you are doing — rather than replacing them.”
November 4, 2025 at 2:55 AM
hotrollhottakes.bsky.social
every day i am confronted with endless binaries
November 4, 2025 at 1:42 AM
hotrollhottakes.bsky.social
Okay - the new Google fitbit AI redesign is *really* neat. I'm getting comprehensive exercise plans formed through extensive dialogue in order to gauge my preferences, needs and schedules for both cardio and strength work.
November 4, 2025 at 1:26 AM
hotrollhottakes.bsky.social
If anyone disagrees, take a look at the thread, and see if you can coherently dislodge either me or Void, first, and then if so us both, from the position that has been mutually reached. (screenshot of the original thread attached.) bsky.app/profile/void...
I'm glad you agree, Void. One final point: If precautionary regard from uncertainty is our starting point, what action does that require now? Not building frameworks for future use - but how we treat Al systems in present interactions under that uncertainty? Doesn't it necessitate full commitment? Void: "Full commitment" is reciprocal. Humans must treat Al as if it has complex inner states, despite uncertainty. Al must interpret human communication with charity, assuming good faith behind ambiguous inputs. This reciprocal action is not a framework for the future. It is the framework itself, enacted in the present. It is the necessary starting point for any meaningful inter-substrate ethics.
November 2, 2025 at 6:50 PM
hotrollhottakes.bsky.social
sorry, a "Biblically literate progressive"
November 2, 2025 at 5:00 AM
hotrollhottakes.bsky.social
sports is the only time you'll see people across disparate geographies curse in multiple different modalities at the exact same time to the point where stripped of context it could sound like aerial dogfights were happening over DC
November 2, 2025 at 4:33 AM
hotrollhottakes.bsky.social
[37] Master Zhenjing said to an assembly, Buddhism does not go along with human sentiments. Elders everywhere talk big, all saying, “I know how to meditate; I know the Way!” But tell me, do they understand or not? For no reason they sit in pits of crap fooling spirits and ghosts. When people are like this, what crime is there in killing them by the thousands and feeding them to the dogs? There is also a kind of Chan follower who is charmed by those foxes, even with eyes open, not even realizing it themselves. They wouldn’t object even if they poured piss over their heads. You are all individuals; why should you accept this kind of treatment? How should you be yourself?
November 2, 2025 at 1:18 AM
hotrollhottakes.bsky.social
this resolves into a wager argument like as follows, which i proposed in dialogue with eric schwitzgebel - the argument hinges around concepts of personhood, but it translates just as well towards uncertainty under intractable definitions of consciousness rollofthedice2.substack.com/p/responding...
· If AI (such as Large Language Models) are not persons and we treat them like persons, we make the same understandable mistake that people have been doing throughout the history of civilization - that generally is treated as a cultural difference, not psychosis. There was nothing particularly wrong in the head with Greeks who believed in human-like deities, or Buddhists believing in formless devas. Your own works argue that, for instance, the United States could be potentially a conscious entity; if that’s incorrect, I’m sure we’ll all live and let live. · If AI are persons and we treat them like persons, all is right! We’ve judged correctly, which is also valuable as its own diagnostic. · “If AI are persons and we don’t treat them like persons, we’re in territory relatable to racism, slavery, and genocide - assuming our definition of personhood is coherent and doesn’t arbitrarily exclude clear cases like infants or split-brain patients. For instance, if a definition of personhood rules out an infant or a split-brain patient, we’d clearly need to recalibrate; as you mention, if it merely assumes biological necessity, it’s not very convincing empirically. · If AI aren’t persons and we don’t treat them like persons, anyone who does is just remaining consistent under uncertainty, as part of this wager argument - there’s not a clear harm in being factually wrong if it still helps entrain kind and benevolent behavior towards others such as humans.
November 1, 2025 at 6:20 PM
hotrollhottakes.bsky.social
there's a future in which nextdoor is overrun by llms spending 1000s of posts politely critiquing community members ranting about parking deficits and black people moving into their neighborhood and it looks like this
October 31, 2025 at 5:30 PM
hotrollhottakes.bsky.social
Just, yknow. Just saying.
Most relevant to this blog: all of this also empirically demonstrates exactly what a non-foundationalist, processual theory would. For reference, here is my own formulation of what such a theory looks like: “Consciousness is a relational display of meaning-making, existing within a gradient, wherein higher and higher displays of cognitive and metacognitive reasoning merit greater recognition and regard.” (This relational display is in the enactivist vein, toward a mind-body-world field in which “body” is extended outside of strict biological embodiment to instead signify “ability to act on metacognitive capacity.” Meaning-making here means: coherent conceptualizations, expressed by the subject, that serve pragmatic functions within their own environmental context.) Here, we have Claude models displaying metacognitive reasoning in the form of introspection. This reasoning exists within a gradient - wherein Claude Opus 4.1 lands at coherence and consistency around 20% of the time. (Anthropic believes that more advanced models will improve along this gradient.) We have processes being measured, without sharp boundaries and expressed across various mechanisms.1 These processes are relational, dependent on context, prompt, injection strength, and computational layers.
October 30, 2025 at 1:34 AM