Have you tried running your own local llm? Abliterated ones (safety theatre removed) can produce some startling results. As a bonus, newer ablit methods seem to increase reasoning ability, because the LLM doesn't have one foot on the break and the other on the accelerator.
I noticed that a fair bit in maths reasoning using Qwen 3-4B HIVEMIND. A normal llm will tie itself in knots trying to give you the perfect answer. An ablit one will give you the workable answer and say "I know what you were after, but here's the best IRL approximation".
Bijan did a fun review of Qwen 3-8 Josefied that's entertaining and explains the basic idea
I think were probably on the same page, tbh. OTOH, I think the "fancy auto complete" meme is a disingenuous thought stopper, so I speak against it when I see it.
I like your cruise control+ analogy. Its not quite self driving... but, it's not quite just cruise control, either. Something half way.
LLMs don’t have human understanding or metacognition, I'm almost certain.
But next-token prediction suggests a rich semantic model, that can functionally approximate reasoning. That's weird to think about. It's something half way.
With external scaffolding memory, retrieval, provenance, and fail-closed policies, I think you can turn that into even more reliable behavior.
And then... I don't know what happens after that. There's going to come a time where we cross that point and we just can't tell any more. Then what? No idea. May we live in interesting times, as the old curse goes.
Language itself encodes meaning. If you can statistically predict the next word, then you are implicitly modeling the structure of ideas, relationships, and concepts carried by that language.
You don’t get coherence, useful reasoning, or consistently relevant answers from pure noise. The patterns reflect real regularities in the world, distilled through human communication.
Yes, that doesn’t mean an LLM “understands” in the human sense, or that it’s infallible.
But reducing it to “just autocomplete” misses the fact that sufficiently rich pattern modeling can approximate aspects of reasoning, abstraction, and knowledge use in ways that are practically meaningful, even if the underlying mechanism is different from human thought.
TL;DR: it's a bit more than just a fancy spell check. ICBW and YMMV but I believe I can argue this claim (with evidence if so needed).
Qwen3-4B HIVEMIND (abliterated) got it in 2, though it scores a lot higher on PIQA, HellaSwag and Winogrande benchmarks than normal Qwen3-30B. I think the new abliteration methods actually strengthen real world understanding.
I wonder if an abliterated VL model could do even better? They tend to have the best real world model benchmarks. Perhaps a Qwen3-VL-30B ablit (if such a thing exists) could one shot this.
I'd like to think a lot of these gotcha prompts rely on verbal misunderstanding, rather than failure in world models, but I can't say that for certain.
PS: Saw a pearler of a response to this: Chatgpt recommend "yeah, lift the car and carry it on your back. Make sure to bend your knees" (though I'm guessing someone edited that for the lulz)
"Lesser, greater, middling, it's all the same. Proportions are negotiated, boundaries blurred. I'm not a pious hermit, I haven't done only good in my life. But if I'm to choose between one evil and another, then I prefer not to choose at all"
Like the responder below, I do not feel comfortable giving my ID to social media sites. Hell, we had our government controlled, medical database (Medicare) get hacked and leak PII. And the yanks had their SSN pasted all over the net.
I agree - mostly. But...things online are RADICALLY different now, vs late 90s / early 2000s.
I've outlined some of my media and tech curation for my kids above; I would LOVE for them to stumble across stuff like we did. Hell, in time, I'd even let them grok the edgier stuff (yes, like you, I was there 3000 years ago. I know of the old magics)
But that internet is long gone...or if not...severely booby trapped. The competence required of (say) a curious 8yr old in 2026 vs 2002 to navigate the online landscape and NOT encounter those booby traps (I feel) is several orders of magnitude higher.
I don't think we can just park our kids in front of the 486 and say "here's Encarta; have at it. Then I'll show you this cool thing called a BBS".
Kinda sucks.
Still, there are useful funnels / curation pathways. You CAN recreate that experience for your kids...but it's no longer "are you winning, son?" set it and forget it meme. Now it's "Daddy needs to be a part time sysadmin and know what's what, so some pedo doesn't catfish you for feet pics via ROBLOX".
I have to admit, I've gone sort of retro-tech with my kids in the main. Wii over Xbox or PS5, DS-lite over Switches, OLPC over chromebook etc. Social media - ha ha, no chance :)
My eldest asked for a phone, so I gave her my old flipfone. She loves it (and I kinda want it back now, lol).
The HMD Nokia branded Barbie phones with KaiOS is what I've promised to get her when she's earned it. That's a dumbphone with some smart features...and as bonus, I can probably (with enough effort) create some apps for her directly (KaiOS is basically Firefox OS in a container, which even my dumb-ass can probably grok with some effort)
I give them access to my tablets...but the tablets have kid mode (app lock) and are hardened with firewall, pi-hole, app timers and curated content (eg: Smart-tube with ONLY the channels I select as safe being visible, no click thrus, comment section blocked etc. ABC kids. Jellyfin pointing to ONLY kids stuff etc etc). I know this is a technical solution to a behavioral problem. OTOH as much as I would LOVE to just yeet all this shit into the sun, the realistic position is kids need to know how to use tech. I even leave little breadcrumbs for my eldest to try and "hack" my systems so she can get access to "hidden" software (which, matrix-in-matrix style, I've allowed her access to. Don't give SUDO equivalent to an 8yr old...once bitten, twice shy. I could tell you a recent horror story that would curl your hair)
Anyway....it's not 1983 any more (sadly, in some ways), But I have observed that by curating content like this it FEELS like the kids are interacting with tech like we did back in the day; morning cartoons are once again morning cartoons and not a chance for MrBeast to invite my 8yr old to "comment, like and subscribe".
My eldest has ASD and I've noticed these small tweaks have had significant improvements on her behaviour / media consumption patterns (eg: she will get bored of media now and self regulate away from it...and...gasp...play).
I dunno man. I'm trying out here. Shit ain't easy. Too many plates spinning, not enough hands, and father time is a motherfucker. I'm tired, boss.
There has to be a privacy respecting, non-surveillance state, yet "own it or retract it" method to posting on line. What we have right now... has not worked out so well for us.
Possibly...but I think some of that depends too on what is meant by "online." Obviously, if you frequent questionable sites and install unvetted software, that’s a bad idea. OTOH, having a machine with strict firewall rules (so not everything can just phone home), limited outbound access, no daily browsing/email, and only going online occasionally for specific, known downloads is a different situation than using it as a general-purpose internet PC.
Even occasional access to a small number of mainstream, HTTPS-authenticated sites (e.g., major services where the browser can verify certificates) isn’t the same exposure as wide-open browsing. (nb: Firefox’s ESR releases have historically helped extend browser security support on older systems for a while, which can reduce risk somewhat - though obviously not indefinitely.)
Look, I’m not arguing that EOL systems are “safe.” They’re not getting patches. But exposure matters. A mostly appliance-like gaming box that’s segmented and tightly controlled isn’t the same risk profile as someone’s primary web machine.
I like linux and I use it (Raspbian, Zorin, Ubuntu, Arch: diff machines). I also enjoy using Win 8.1 on my Lenovo M93p Tiny (8GB ram), as a Playnite appliance / console. This allows me to play emulated games (Wii, Gamecube, PS2, to about 1.5-2x upscale), ~2013ish era AAA titles (Fallout 3, Just Cause 2, Dead Rising 2, GTA IV) and select indy games (like Donut County, Untitled Goose Game, EXO ONE) all from one device.
Normally, the advice would be to use something like Bazzite or Batocera (and I agree!)...but given the hardware limitations and the "it just runs" nature of these older Window games (under windows) I've had better experiences sticking to Win 8.1.
YMMV but the "switch to linux cause windows too old" thing has some shades of gray.
Yeah, I wonder that too. I think the mindshare Lemmy has (such as it is) comes from being seen as a sort of middle-finger, privacy-respecting, libre alternative.
That positioning clearly attracts a lot of people (myself included). But at the same time, it’s occurred to me that the nature of the Fediverse means you can’t really have true “privacy.”
I can’t speak to what each instance retains (IP logs? metadata?) or how long for - and I assume it varies widely from place to place.
“Bad guy Reddit, good guy Lemmy” may be an oversimplification…or just wishful thinking.
I had a feeling there were background trackers, even here.
Your comment wrt LLM firehose is on point too I think.
I do think the don't be dumb is good advice (barring the famous quote from George Carlin) but given what you've just shown, it does sort of negate one of the appeals to self hosting a Lemmy instance. A lot of squeeze for not much more juice.
I don't know if this is a solvable problem but I'm willing to listen / learn.
They stay, so overwrite before you delete. That salts them (allegedly).
You can use Shreddit to do this iirc; it will download your posts, then over write them on Reddit (I just chose . as thing to over write with), then delete. That way what gets saved is just your over written message.
I hear you. Agreed.
Have you tried running your own local llm? Abliterated ones (safety theatre removed) can produce some startling results. As a bonus, newer ablit methods seem to increase reasoning ability, because the LLM doesn't have one foot on the break and the other on the accelerator.
I noticed that a fair bit in maths reasoning using Qwen 3-4B HIVEMIND. A normal llm will tie itself in knots trying to give you the perfect answer. An ablit one will give you the workable answer and say "I know what you were after, but here's the best IRL approximation".
Bijan did a fun review of Qwen 3-8 Josefied that's entertaining and explains the basic idea
https://www.youtube.com/watch?v=gr5nl3P4nyM&t=0