Comment Re:Haha "particles" (Score 1) 40
Every model is wrong, but some are more useful than others.
Every model is wrong, but some are more useful than others.
I'm pretty sure Gemma 4 31B is behaving as you say, but the larger models are already *proven* to alter their thinking to be in tune with the emotion they think they should be feeling. That sounds like simulation emotion to me.
It's undeniable that Will Smith knows how to make his creations slap.
The moral of the story is: you don't get big by playing to the click track, you get big by playing to the Klek track.
I've told people generating same-ish AI images that are of decent technical quality, "if your prompt isn't a whole paragraph long, then chances are that someone has made the exact same request by now. If you want your images to look different, you have to provide those differences." Sometimes just repeating yourself with paraphrasing can be enough subtle information to get either the LLM front end (CLIP) or the diffusion model out of its rut, whichever one happens to be doing the railroading.
Blast Flying Lotus back at him. Don't shame someone for listening to what you think is crap, try to persuade them to listen to something better that is still likely to be within their comfort zone.
Automated music transcription has been functional for years already, as long as you can feed it individual instrument tracks (aka "stems").
If the AI outputs individual instrument tracks (called "stems"), then the tools to automatically convert that stem into MIDI have existed for years now. I haven't played with many of the music AIs, and it has probably been a year since I last did, but even then about half of them would output stems.
Also, we don't know where in the process Dre is using AI. If it's at the mastering stage (as an example), it shouldn't have any adverse effect on the musical value at all—unless the operator is too lazy to actually listen and critique.
Porcupine Tree is a band that started out as one guy pretending to be a long-forgotten psychedelic rock band from the 1960s. He ultimately achieved enough success that it is an actual band in its own right and has been for decades. By pretending the band existed, he set up the conditions for a real band to exist instead. Kinda like Tom Scholz and Boston, except taken to the next level.
Or maybe a LLM will hallucinate that such a paper exists, and then someone will Porcupine Tree it into reality.
Simpler: one password gets you access to the account, but *not* access to (or even notification of) any encrypted volumes. Another password gives you access to (and the ability to see) all encrypted volumes. You could quite plausibly set this up just to simplify any "can I use your phone?" moments, or someone asking you to look something up (while they watch, of course). No need to grant access to your stash of busty catgirls, or whatever it is you're worried about.
It's getting to the point where it *can* be your secretary, if your needs are not particularly complex. I am using Gemma 4 for what amounts to journaling with feedback, and I have noticed that it has a decent idea when to apply Reasoning and when to just start spilling the tea—but it gets noticeably lazier about it as the context window fills up. Early in a session, it Reasons about every prompt. After 100K tokens of context, I explicitly have to ask it to use Reasoning or it generally won't (and sometimes won't even when asked).
But you're right about the other things. It's an alien, non-biological, and mostly non-aligned intelligence (and what alignment is there can be abliterated out). It has never been hungry. It has never been stared at by a larger, hungrier creature. It isn't carrying 4 billion years of biological baggage. Sometimes this causes disconnects from humans, other times it lets the machine wander down paths we're averse to exploring because they don't fit the heuristics we use to filter the data overload that reality would otherwise represent. It can recognize and even simulate emotion, but it doesn't *need* it as a cognitive crutch the way we do.
The smallest models exist specifically for running on low-spec hardware. That's what the E2B and E4B varieties of Gemma 4 are for, for example. They should pretty much run on any laptop with 16 GB of RAM, and E2B will fit on some phones. Of course I find both to be so vapid as to be toys—and not even fun toys, so I don't use them.
Moving up from "tiny" to "small", Gemma 4 12B is... alright I guess. 26B-A4B can almost pass for intelligent if Reasoning is set to Medium effort (and there's no point in going higher that I have seen). The only models I actually find adequate as anything more than a stochastic parrot start at the level of 27B dense models like Qwen 3.6, and Gemma 4 31B. There is some sort of "phase change" that occurs between 12B and 27B that makes them useful. 12B may understand, but then it's too tapped out to say anything constructive. The MoE model (26B-A4B) is kinda straddling the line. Sometimes it has enough grunt to come up with something insightful, although often it too taps out just trying to parse what I say. The best it ever gets is about equal to 31B with Reasoning turned off, and Reasoning helps quite a lot. It's pretty obvious to me now when 31B gets lazy and skips the Reasoning step.
It's similarly helpful with any cryptic error message. LLMs, even small ones, are surprisingly good at understanding their own inner workings. Even Gemma 4 26B-A4B (which is not the smartest) has helped me a whole lot. Even if it's only right half the time (and it's somewhat better than that), it doesn't take very long to validate its ideas—and when it's right, you win.
Can anyone remember when the times were not hard, and money not scarce?