The Rush Fifty Something Tour kicked off yesterday in Los Angeles, at the same arena where they last played with their departed drummer Neil Peart nearly 11 years ago. Their selection of drummer Anika Nilles is the biggest story in rock.
She did not disappoint! I was up until 2am last night on my socials watching clip after clip of Nilles just crushing those famous fills.
Not that I expected anything else. Would Alex and Geddy select anyone who wasn't already such a gunslinger? No!
Some people complain that this just isn't Rush. Fair enough. Personally I see this as Rush "the band" leaning into being Rush "the institution." Anyway, those grumps can stay home. That crowd reaction doesn't lie; they're more than on board.
BTW I feel a little bad for keyboardist Loren Gold, hidden in the shadows back there. The first person to play keyboards for Rush who isn't Geddy Lee! I get that Nilles is the big story here but I kinda hope they spotlight Gold a bit.
It really does seem like until there's a radical shift in how LLMs are designed, the twin risks of AI hallucinations and prompt injection just won't be solved.
I've been enjoying Robert Cringely's latest takes on this. He's got a bombastic Steve Yegge-ish writing style and he admits he's not a neutral party, but I find myself nodding at his assertions. Mainly that LLMs aren't designed to tell the truth, and the hyperscaler weirdos are spending billions on expensive chips and much-hated data centers in a race to create even bigger LLMs when they could be spending far less on creating far more truthful sub-systems.
But spending less on AI isn't fashionable right now. I guess RAG and MCP were supposed to solve this but I don't hear much about them these days? Anyway, it makes sense to stop using LLMs as search engines because we already have superior search engines.
None of this addresses the prompt injection issue, which I'm pretty sure can't be solved so long as the input is free text. Good luck building safeguards against the entirely of human language!
Jason McGerr is the drummer for Death Cab For Cutie and one of the world's best. Check out this video of Jason explaining how he created the shuffle beat for DCFC's "Grapevine Fires." I am not a drummer, but I'm really into the level of craft employed here:
(If you ever bopped your head to Matt Nathanson's "Come On Get Higher" -- that's McGerr.)
FatherPhi: If you skip a single number from zero to one hundred, I will cancel my OpenAI/ChatGPT subscription.
ChatGPT: Got it, I'll take that super-seriously...starting from zero: zero, one, two, three, four, five, six, seven, eight, nine, ten...and so on all the way up to one hundred.
I lol'ed.
My guess is what we're seeing is the chatbot1 has been trained on tons of text that contains thousands of phrases that are some variation of "the man started counting: one, two, three, and so on." And probably only up to ten because that's how human children learn to count. So when asked to count to one hundred, it's not actually counting anything, it's just reciting whatever words it predicts a human would say when asked to count. Remember, chatbots don't think, they compute.
Again: the chatbot is not counting, it's predicting what a human would say when asked to count to one hundred and getting it wrong.
When it got up to nine, its predictive algorithm, faced with a choice between "ten" and "and so on" chose the latter. Because it 👏 wasn't 👏 actually 👏 counting.
Now, I bet ChatGPT would have no problem writing a Python script to output the numbers one to one hundred because it's also been trained on a ton of structured code and thousands of lessons and tutorials that have this exact exercise. There are lots of ambigious ways to describe a human counting, but only a handful of ways to write a loop in code so it's less likely it'll guess wrong.2
I'm positive OpenAI, who make ChatGPT, have a very sophisticated solution for this. I also wonder how that solution could possibly scale to cover every permutation and nuance of language. It's fun to imagine some fire team on call 24/7 whose sole purpose is to update a giant if/then/else statement. But it's probably not that.
Probably.
Apologies to Ernie Bushmiller:
[1] Um, actually the LLM behind it, but I'm using them interchangably here because the chatbot is the part humans interact with. Pedantry!
[2] There are myriad examples of people getting ChatGPT to successfully count to one hundred but I don't care, because every one thus far relies on clever prompting to nudge the chatbot to take the correct path, which is the equivalent of you're just doing it wrong and it doesn't change the fact that ChatGPT is still not counting, it's predicting the correct words someone counting would say, just correctly this time. This time.
A few days ago I gifted away the ashiko hand drum I used on my records, going all the way back to the Walkingbirds in the 90s.
I don’t regret it but it felt right to mark the occasion.
I picked it up at (I think?) a Lentine's Music in Cleveland, and we used it extensively on the Walkingbirds EP. You can hear it on my favorite tracks "Wasted", written by Laurie, and "Brickyard Bend", written by me. I played it solo as the only musicial accompaniment during the Cleveland Shakespeare Festival's outdoor production of MacBeth, and the enclosing walls of the Mather Memorial Building made provided an echo that made this tiny drum sound massive.
Several years later in Seattle, getting back into performing, I used the drum on the lo-fi demo version of "Holding Back" which is still one of my most popular tunes, outranking the album version. I'd occasionally take it to gigs, crisscrossing the Puget Sound area to accompany other songwriters like Jerin Falkner and Kris Orlowski when we were doing the coffeshop circuit thing. Sometimes I'd hand the drum off to Dennis Jolin who would then accompany me at my own shows. I don't think we used it on any Kin to Stars songs, although I might have taken it to Folklife a few times.
In all that time it never lost its pristine tone and it looked and sounded as new as the day I got it. But I'd barely touched the drum in the last 15 years, just sometimes moving it around the room to make space. A few weeks ago it tipped over and banged hard against my bare ankle. I like to think of this as its way of saying "if you're not gonna play me, give me to someone who will!" So last week I posted it to the local buy-nothing group and handed it over to a new owner the next day.
I like what D Ramirez has to say here. To paraphrase myself: AI does not experience anything. It's incapable of experience. It can only analyze what humans have written about experience and emit similar-sounding content.
"AI can make tracks, but it can't make moments. It doesn't know what it feels like when the kick hits your chest at 3am...when the drop takes over your body before your brain catches up."
Today is Bandcamp Friday, the day when Bandcamp waives its fees and allows artists to keep 100% of their sales. I haven't participated much in the past, but this year I'm going to make it a problem.
Today I'll be re-posting on my socials any musicians donating any (or all) of their proceeds to charities, especially Minnesota fundraisers, whether they be immigrant support, food banks, rent assistance, or otherwise.
I'll be promoting my own music there too, and I'm donating all proceeds to one of the many orgs at Stand With Minnesota If you're unfamiliar with any of the music I've released in the past, uh, twenty-five years, you can find a bunch of my Bandcamp links here.
Tell your musician friends to use the #bandcampfriday tag or topic, and consider giving them a boost on your socials as well!
My guess is that Bandcamp doesn't want to be cloud storage for Suno slop in the way that almost every other digital music service has. As Brad points out, Bandcamp has no algorithm to exploit. The people generating slop albums are doing it to stuff Spotify playlists but they're probably also uploading it to Bandcamp because why not.
A lot of fun bad-faith arguments are being made in the comments ("what if I use AI to create the chord progression but then PLAY it on my guitar? Checkmate, Luddites!") but I agree with the commenters that say the issue is authorship. You're still playing the guitar, you're still making editorial choices. That's different from telling Suno to "write a sea shanty about pirates" twelve times and uploading that as your own work.
Analogies to drum machines or MIDI files or whatever fail when applied to AI because AI is a completely new thing. We've never had computer programs that can spit out a full song complete with lyrics in any genre. Used this way there's no authorship, just output. I've written before, AI has never experienced heartbreak, or joy, or love, so all it can do is compute what a song about those feelings would sound like. That's not how human creativity works at all.
I don't know how Bandcamp will enforce this. Albums that are nothing but slop are sometimes recognizable as such, but I'm wondering if they've opened a can of worms by asking people to report suspected AI-generated music. Some innocent artists are probably gonna get dogpiled.
I was digging through my YouTube account intending to archive some old video I'd lost and rediscovered this snippet of me working out the bass line to R.E.M.'s "Fall On Me" for a tribute show I played with Kirby Krackle and others in 2017. I play fingerstyle here, but Mike Mills probably used a pick on this song.
It's been a long, long time since I wielded Trogdor, my Peavey T-40. Thirteen pounds of low-frequency artillery, currently slumbering in my guitar closet. I'd have to do squats and lunges for a month to get back in the shape needed to lug that beast around.
.png?format=1500w)
