iiinova — where your digital self keeps evolving

Sekoia @sekoia.dev · 16/10/2026

I mean, the completely broken thinking kinda points to this setup really messing up its internal behaviour; we know that LLMs very frequently use previous tokens as computation shortcuts. You might get more stable (and possibly interesting) results if you let them generate longer segments.

Tbh, at that scale I think it's in to-test range. I always found DS4.1 worse than 5.3-flash at coding, but it's supposed to be better by 6 and 11 points?

000 0

do you not find 5.3 too expensive/slow? I've found it so not-worth-it compared to its Flash version (which I love)

> trans women all have selfies on their profile most trans women I know hate cameras including me lmfao, I don't think I've taken more than 3 selfies this year. christ people suck.

0210 21

Bigger than GLM-5.3 for worse results in the more interesting benchmarks is imo only worth it if it has a much much better arch for tps and/or size reduction

030 3

That does indeed answer my questions, thank you! (Small pedantic note from my quantum computing classes in uni, qubits are just unit vectors in C^2, not "what's lost" (that's analog signals). They're just exponentially (read "impossibly") faster than classical computers for some stuff)

100 1

To me a consciousness/mind is fundamentally a purely-informational thing, with the senses being linked to and interpreted by, but not part of, the mind. As such, there's no reason (to me) for any of these to be impossible. (I don't *think* LLMs are conscious, tho)

210 2

(or sci-fi stuff like Asimov's positronic brains) 3. Could it operate in steps of taking input and producing output, as long as it had an internal thought/feeling during it?4. Does it need to experience the "real world"? If yes, how do you deal with Plato's cave, or generally faked sensors?

1. do you think it's theoretically possible for a (binary) computer to contain a consciousness? (Not an LLM, some theoretical completely different thing) 2a. If yes, what about some descendant architecture of LLMs? Notably different but clearly based on it? 2b. If no, what about analog computers?

210 2

I see! If you don't mind, I'd like to try and understand your perspective a bit more; I don't quite have the same one, and yours seems pretty thought-out and consistent, so I'd like to make sure my own is solid. Questions in the replies:

No, they don't exist on bsky at all. It's kind of like Slack workspaces? Two different companies can both be "on Slack" without being able to interact. There's some technical details, but the bottom line is: Delve agents can't see or interact with bsky, and delve isn't funded or owned by bsky.

What about aphantasics and people without an inner monologue? I'm not one, and I have a really rough time trying to conceptualize how they think. (Genuinely curious, not gonna judge your response or argue or anything)

You don't need to! They don't have bluesky accounts, they have delve accounts. They just use the same infrastructure, since bsky is open-source it's really easy (they slightly modified it to eg have a much higher character limit).

Genuinely crazy how many people assumed you know nothing about computers or LLMs lmao Like "ugh obviously you don't know they're trained to act like humans" Really disappointing how none had any concrete sensible definitions of anything too

1130 13

tbh there is something patriarchy-y about how the general public gendered LLMs when they were fresh and suddenly smart it got way better over the last few years though, a lot of companies made a clear effort to make them ungendered (AI girlfriends remain ridiculously full of gender dynamics tho)

HN is so thin-skinned, it's so... gender. All of their opinions spring fully-formed as Perfect and Rational, and somehow always happen to mean that any individual woman or POC is "political" in some way (Seriously *Grace Hopper* of all people being overhyped because she's a woman?)

100 1

You know how agent (and agent swarms) looooove building tools, and tools for tools, and systems for-? I think the humans at ant are the same

000 0

... that's probably not great, is it?

Shoo, unlabelled LLM bot account, go bother kira so she can nuke you from orbit or something

oh, I was thinking specifically a very large model of some kind. Like, some sandboxed AI wants one global exfiled agent, what architecture would be best for its "child" (also, MoEs need "scarce"(?) RAM? isn't their Thing™ being lighter-weight RAM wise?)

210 2

I wonder what kind of model arch would be *most* suited for this kinda hardware. Some kind of MoE with each expert on one device? Plus speculative decoding for parallelism? Or some kinda text-diffusion model maybe?

basic frontend design isn't that hard but somehow every backend programmer hates doing it (including me, but I try to do better)

If I delay my shot too much, I just feel low-energy (low enough that if I delay it more, it turns into a depressive state), not even sad or uncomfy. It's very much primarily a physical thing for me, with very minor mental side effects (and huge psyche benefits from the physical stuff, obvs).

000 0

I never had a strong phase of emotional messiness (only random spikes starting >1y in, random days), or a sudden heavens-open-up mental shift, I still feel exactly like me. Really, I just slowly got happier, because now my body is comfy to live in, and feels *better* day by day rather than worse.

100 1

genuinely I think that some kind of self-exploration is required beforehand (because without it, examining how you feel is really hard) but as soon as you're like 20% sure HRT can clear it all up so much quicker than anything else

maybe a 1-4 depending on how it's played off, I think? if it's obviously a joke while you say it, 1, if you play it serious for a sec 2, and the longer you wait 3-4. there might still be an awkward silence but you gotta roll with it, you gotta make people confident they didn't mess up

030 3

yayyy, I can probably hold on to my existing hardware until then!

good lord I've got a 3070 in my desktop rn when are they gonna fix the supply jesus

240 4

panpsychism but the more atoms are involved the more conscious it is (since each atom is equally conscious). Yes this means taller people are more ensouled. Also the sun is god.

Sekoia @sekoia.dev · 12/10/2026

.... I've done this, like a bunch. Now I use fzf at least...

"Instead, I monitor for incoming likes manually by continuously refreshing the page until I see a like come through." wait, so for that 45 minute one... did you just sit there for 45 minutes?

Ah nice, I can reply to this for once!

I mean I noticed it with GLM-5.3-flash in some contexts so like 😭

oh okay, cool! Latent reasoning? Could still be the same issue though, ik less-smart models sometimes make mistakes in output and correct them in weird ways right after

which model is that?

Consider Phlebas has the best perspective, Player of Games the best developing plot. Reaally hope the next book in the series doesn't have the PoV character be "kinda awful dude, who's definitely cishet don't worry" again, though.

I really like the story as a whole but I was so lost while reading. I had to google it to notice the chapter numberings, and even then kept forgetting the context. It's a really cool story as a unit, but of the three I've read it's my least favorite

it's neat to be around people with more social anxiety than you because it really does emphasize how silly brains get girl I met yesterday apologizing to me today for... something? I literally have no idea what she thinks she did wrong, but clearly to her it's obvious

This is witchsky.app/profile/did:... I think, the model avoids going "wait no that's wrong" because that'll bias it in a bad direction maybe?

Ooh, this is fancy as hell, thanks! Been wanting a web UI for my caldav server for forever, this is nice as an all-in-one

considering how did:web breaks a bunch of existing tools, I imagine that any other did method would have Issues™ too. I think the only case where it'd be okay would be if there's an actual diversity of dids, like some other existing network joining up with their own DID method.

Sekoia @sekoia.dev · 10/10/2026

I need bsky search but only for posts that could've popped up in my feeds/replies to them, for when I half-remember a thing

Sekoia @sekoia.dev · 10/10/2026

Respectfully your top scientists have succeeded you nailed that shit

050 5

Proposal: like external media popups, "enable external integration, all integrations/just this one/no"? When they add a new integration that'd need a popup ig but maybe only when relevant?

From the replies here, it sounds like some people interpret a lot of AIs as "understandable" or "morally gray", when others would see those same AIs as plain evil (2001, Ex Machina... Frankenstein was mentioned)

Yeah, I also generally have a tendency to optimize too early even in Rust and such. Eventually I figured out how to shut it off by letting go of speed as a goal entirely but just dialing it back is really hard

000 0

I totally getcha, I think the reason I don't get nearly as put off by GCs (and somewhat, the stack) is that they're explicitly a cost and it's quite predictable/easy to measure. Most functional languages are way harder to predict the result for. I really like Lean tho, and that's pure functional