In this edition
The Character Is Fictional
Kids are getting scared out of their minds by grainy AI videos of a sinister Cat in the Hat lurking outside schools, and police in multiple countries have had to issue warnings that the Dr. Seuss character is not actually coming for them. Open on the police statement, not the video — the funniest artefact in this story is an official force account confirming a fictional cat's whereabouts. Land on the specific absurdity of multiple countries independently reaching the same conclusion. The disagreement: Abby says this is old panic in a new coat and the story is really about how bored the internet is; Boyd says the shift is that the production cost of a convincing hoax has gone to roughly zero, and that is a real change, not a cute one. Do not dwell on frightened children, and do not linger on the teen who was arrested — name it once as a fact of the story and move past it. Do not overstate the scale: this is a meme panic, not a wave of crime, and neither host should reach for 'this changes everything'.
The New York Times
Meanwhile
SPAR-K: Scheduled Periodic Alternating Early Exit for Spoken Language Models
arXiv cs.CLarXiv cs.CLarXiv cs.CLarXiv cs.CL
Preserving Perceptual Quality
SPAR-K: Scheduled Periodic Alternating Early Exit for Spoken Language Models Open with the two-week claim read completely deadpan, then let Boyd stop on the word 'verified'. The detail to land on: nobody ever announces how long the verification took, only how short the design took. Where they disagree: Abby will say the underlying compression of the cycle is real and interesting even if the framing is nonsense; Boyd will say a claim that cannot be checked is not a result, it is an advertisement. Let Boyd's crack land once and then get out. Do not accuse anyone of fraud — the point is the vagueness, not dishonesty.
arXiv cs.CL
Slop, or Not?
SPAR-K: Scheduled Periodic Alternating Early Exit for Spoken Language Models Verdict: slop, and confidently so — the audit was pre-registered, which is the one detail that makes the finding hard to wriggle out of, and that is the line to land on. Abby's one concession: the people citing it mostly aren't being cynical, they just inherited the metric. Boyd's verdict is one sentence and no hedging. High confidence. Keep it to thirty seconds; do not turn it into a statistics lecture.
arXiv cs.CL
Something Useful
SPAR-K: Scheduled Periodic Alternating Early Exit for Spoken Language Models Abby frames first and should be allowed her enthusiasm — this is a clean small true story and both hosts genuinely soften on it. The useful thing: it is reproducible, the recipe is published, and a listener with a gaming PC could actually attempt it this weekend. The catch to land on: what you get at the end is small, and 'small' means the model will be charming rather than competent. Boyd's role is not to sneer but to price it honestly — the card, the hours, the power. Do not oversell it as beating anything commercial; the appeal is that it does not pretend to.
arXiv cs.CL
Something Weird
SPAR-K: Scheduled Periodic Alternating Early Exit for Spoken Language Models Read the title straight, once, with no setup and no explanation. Then one beat. Then the second read, slower. That is the whole segment. Nobody explains it, because explaining it kills it.
arXiv cs.CL
Transcript
0:07 Cold Open
AbbyWelcome to The Daily Slop Show.
Boyd[snappy] Where nothing we say becomes the truth.
Abby[excited] Today, a police force had to officially confirm the location of a cartoon cat. [laughs] Boyd, this is folklore. This is just a campfire story with a better render—
Boyd[jumping in] —that costs nothing to make. [statement] That's the bit. Not the cat. The price.
Abby[thoughtful] Hm. [short pause] Okay, hold that, because it starts with a press release nobody was trained to write.
0:36 The Character Is Fictional
Abby[excited] So somewhere out there is a police press officer who woke up, made a coffee, and typed the sentence "the character is fictional."
Boyd[snappy] Which character.
Abby[statement] The Cat in the Hat. [short pause] The New York Times reports kids are getting scared out of their minds by grainy AI videos of a sinister Cat in the Hat lurking outside schools, and police in multiple countries have had to issue warnings that the Dr. Seuss character is not actually coming for them.
Boyd[thoughtful] Hm. [short pause] Multiple countries. Independently. Arriving at the same public statement.
Abby[laughs] About a cat! In a hat! [statement] And honestly, Boyd, this is the oldest story there is. Every generation gets a thing that stands at the end of the street. Slender Man, the phantom clowns, whatever your gran was scared of—
Boyd[jumping in] —my gran was scared of the bus timetable. [short pause] But go on.
Abby[excited] The bogeyman is not new! The only new thing is the texture. It renders now. That's it, that's the whole story, the internet is bored and it has better crayons.
Boyd[snappy] Except the Times also reports real school lockdowns, and a teen arrested. [short pause] So it left the phone.
Abby[surprised] Okay. [short pause] Yeah. That part's not a campfire.
Boyd[thoughtful] And here's my number, and it isn't the scare. [statement] A convincing hoax used to cost somebody a weekend and a camera. Now it costs an afternoon and a prompt, and institutions still have to answer it at full price. [snappy] That asymmetry is the change.
Abby[thoughtful] Cheap to make. Expensive to deny.
Boyd[snappy] That's it. [short pause] It's a meme panic, not a crime wave. But somebody's job now includes the sentence "he is not real, he is a character in a book."
Abby[laughs] Filed under: things they did not cover at the academy. [short pause] Right. Meanwhile.
2:39 Meanwhile
Boyd[snappy] Meanwhile. [short pause] SPAR-K: Scheduled Periodic Alternating Early Exit for Spoken Language Models, posted to arXiv, computation and language. [short pause] The paper concerns interleaved spoken language models, which alternately generate text tokens and speech tokens. [short pause] Its authors write that decoding at full transformer depth for every step becomes costly, especially due to long speech sequences. [short pause] The listing is version two, filed as a replacement for an earlier version.
3:16 Preserving Perceptual Quality
Boyd[snappy] So that same paper. [statement] The proposal is a modality-aware early exit framework, designed to accelerate this kind of model, and I quote, "while preserving perceptual quality."
Abby[statement] Sure. Faster, sounds the same.
Boyd[suspicious] Preserving. [short pause] Preserving as measured how?
Abby[laughs] Here we go—
Boyd[jumping in] —because the copy on our desk stops mid-word at "we ev—". [snappy] The abstract cuts off exactly where the evaluation begins.
Abby[disbelief] Oh, that's—[laughs] that's almost too neat.
Boyd[long pause] A claim that stops before the checking is not a result. It is a sentence.
Abby[short pause] Okay, but be fair, Boyd. The mechanism is described, and it's specific. Most speech positions exit at a fixed intermediate layer, and then there are periodic full-depth refresh steps to stop the thing drifting.
Boyd[thoughtful] Distribution shift due to early exit. [snappy] Yes. That's a real problem and that's a real answer to it.
Abby[excited] So the idea is worth something even if the press-release half of the sentence isn't—
Boyd[jumping in] —and the press-release half is the half everyone quotes. [short pause] Nobody ever announces how long the checking took. Only how short the clever bit was.
Abby[thoughtful] Hm. [short pause] Fine. Half agreed.
Boyd[snappy] Alright. Anyway.
4:47 Slop, or Not?
Boyd[snappy] Slop, or not?
Abby[excited] Slop or not! Today: the cat videos.
Boyd[statement] Slop. Confidently. [snappy] The Times describes them as grainy AI videos of a Dr. Seuss character outside schools, and the whole effect depends on them being bad enough to be mistaken for a security camera.
Abby[thoughtful] My one concession — the story underneath isn't slop. Kids scaring each other is a hundred percent artisanal, hand-made, locally sourced.
Boyd[snappy] The fear is heritage. The footage is slop. [short pause] Verdict stands.
5:25 Something Useful
Abby[excited] Something Useful! And this one actually explains a thing you've noticed. You know when you talk to a voice assistant and there's that little... gap? Before it answers?
Boyd[snappy] The gap where you say "hello?" into your own kitchen.
Abby[statement] That gap. So the paper's starting point is that these models run every single step through the full depth of the network, and that gets expensive — especially, they say, because speech sequences are long.
Boyd[thoughtful] And the fix is to not do that most of the time.
Abby[excited] Right — most of the speech bits get to leave early, part-way down—
Boyd[cut in] —and then every so often—
Abby[cut in] —every so often it does a full—
Boyd[short pause] Sorry. You go.
Abby[laughs] No, you had it. Go on.
Boyd[snappy] Every so often it does a full-depth pass to keep the thing from drifting. They call those refresh steps.
Abby[thoughtful] Which — and this is me guessing, not them — is the sort of trick that could make voice stuff cheap enough to sit on a small device instead of a data centre.
Boyd[statement] Could. [short pause] It's a preprint on arXiv, on its second version, and the abstract we have stops before the results. [snappy] Nothing shipped to your kitchen this morning.
Abby[sigh] Okay. But the gap has a reason now, and I like knowing that.
6:50 Something Weird
Boyd[snappy] Something Weird. [short pause] The title, straight: SPAR-K. Scheduled Periodic Alternating Early Exit for Spoken Language Models.
Abby[long pause] ...Again?
Boyd[snappy] Scheduled. [short pause] Periodic. [short pause] Alternating. [short pause] Early Exit. [long pause] For Spoken Language Models.
Abby[thoughtful] Every word is real.
Boyd[statement] Every word is real.
7:22 Tomorrow
Abby[excited] Tomorrow: whether that abstract ever tells us what happened after "we ev—".
Boyd[snappy] And whether the cat has an alibi.
AbbyWe're The Daily Slop Show.
Boyd[snappy] None of this becomes true just because we said it. See you tomorrow.
Listen on Apple Podcasts