In this edition
The agent that shows its working, allegedly
Self-Evolving Scientific Agent Discovers Generalizable Physically-Reasoned Fluid Control Open with Abby on the appeal: a control system that can tell you why it did what it did, rather than a black box that just twitches the valve. Land on the specific image — the model that reads a chest scan beautifully until the hospital swaps scanner vendors, and the drone that forgets how to not hit another drone because someone lied to it about where it was for twenty pounds. Be straight with the audience that this is a preprint, version two, and that the descriptions we were given do not match the paper's own abstract — say so on air rather than pretending the metadata is clean. Do not assert the medical or drone results as established findings; frame them as the claims attached, and make the mismatch part of the story. The real disagreement: Abby says brittleness is a stage every technology goes through, Boyd says a thing that only works in the room it was born in is a demo, not a discovery. Don't resolve it.
arXiv cs.AI
The future has a shopping list
AI Firms Go Shopping in Bankruptcy Court - Bloomberg Boyd opens flat: the future has a shopping list. Land on the bathos of the venue — the same firms that talk about civilisational transformation, sitting in a courtroom bidding on someone else's furniture. Keep it factual about what we actually know: it's a Bloomberg piece reporting a pattern, not a named scandal, so no accusing anyone specific of vulturing. Abby's job is to be the one who points out that distressed-asset buying is ordinary, boring corporate behaviour and the only strange thing is the marketing wrapped round it. This is the episode's one industry story — don't let it sprawl into general AI-money talk.
AI scams, fraud and courtrooms
Slop, or not: the money-saving router
Self-Evolving Scientific Agent Discovers Generalizable Physically-Reasoned Fluid Control Verdict: slop, and say it with reasonable confidence — the routing effect looks mostly like task mix wearing a lab coat. Boyd delivers the verdict in one sentence; Abby gets the one honest caveat, that a modest real saving is still a saving and not everything unexciting is a lie. Thirty seconds, no expansion, do not claim the underlying study is definitive.
arXiv cs.AI
Something Useful: your bank statement, in English
Self-Evolving Scientific Agent Discovers Generalizable Physically-Reasoned Fluid Control Abby frames first and gets to be right: this is real, small, cheap, and someone can use it this afternoon. Land on one gnarled merchant string read aloud. Then the catch, delivered without doom — a bank statement is a diary, and 'nothing sensitive went in' is not the same as 'nothing sensitive comes out'. Practical advice: run the small model locally if you can, and don't paste twelve months of transactions into a chat window. Both soften here; it's the quiet-competence item, so don't oversell it into a revolution.
arXiv cs.AI
Something Weird: the accidental jealousy machine
Self-Evolving Scientific Agent Discovers Generalizable Physically-Reasoned Fluid Control Land it cold in one sentence — the machine that makes you feel worse about your life and genuinely does not know. No setup, no methodology, no caveats beyond one word. Abby's instinct will be to find it sad; let Boyd get the last line and keep it short. Out before anyone explains anything.
arXiv cs.AI
Transcript
0:07 Cold open: one preprint, six hats
AbbyWelcome to The Daily Slop Show.
Boyd[snappy] Where nothing we say becomes the truth.
Abby[excited] Boyd, big shelf today. Loads of items. Six, seven—
Boyd[jumping in] —and most of them are the same arXiv preprint about fluid control wearing a different description. [sigh] I clicked all of them.
Abby[laughs] So the slop pipeline has started producing slop about itself. [excited] That's vertical integration!
Boyd[statement] That's a loop. [short pause] Let's start with the hat that says generalizable.
0:37 The agent that shows its working, allegedly
Abby[excited] Okay. A control system that can tell you why it did the thing. Not a black box twitching a valve — an actual chain of reasoning from the physics to the decision. That's the pitch on this preprint: a self-evolving agent, driven by big language models and code it writes itself, building controllers while keeping the reasoning readable. [statement] Boyd, that is genuinely the thing I want.
Boyd[snappy] The abstract says: instead of adjusting weights, the agent deploys — [short pause] and then our copy stops. Mid-word.
AbbyIt does not—
Boyd[jumping in] —it does. And it's version two, announce type replace, so somebody already swapped one out.
AbbyFine. But the word on the tin is generalizable, and that word is—
Boyd[cut in] —doing enormous unpaid labour. Because the descriptions bolted onto today's items, which do not match this paper's own abstract, and we're saying that out loud rather than pretending the metadata's clean — [snappy] those describe things that fall over when the room changes brand.
AbbyMeaning what, specifically?
Boyd[statement] A model that reads a chest scan beautifully until the hospital buys a different scanner. Drones that stop avoiding each other because someone lied to them about where they were with a twenty-quid gadget.
AbbyClaims. Attached claims, not results we've checked—
Boyd[jumping in] —correct, which is why they're funny rather than alarming.
Abby[excited] And brittle is a stage! Everything is brittle first. Early aeroplanes were brittle, early—
Boyd[long pause] [statement] A thing that only works in the room it was born in is a demo, not a discovery. Revolutionary framing, a replaced version two, a title promising it travels, and evidence that mostly says it doesn't.
Abby[thoughtful] Hm. [short pause] ...that's you cracking, isn't it.
Boyd[sigh] Alright. Anyway.
2:45 The future has a shopping list
Boyd[snappy] Second feature, and the future has a shopping list. Bloomberg headline, word for word: AI Firms Go Shopping in Bankruptcy Court.
Abby[surprised] Oh. [short pause] So the companies whose entire pitch is that they are what comes next—
Boyd[jumping in] —are queueing up to buy what came before, at a discount. It's a piece about a pattern, by the way. No named villain in what we've got, so nobody's being accused of anything.
Boyd[sarcastic] It's the register clash I enjoy. Civilisational transformation in the press release. [short pause] Cash-only haggling over somebody's server racks in the actual room.
AbbySee, and the venue is—
Boyd—the venue is really the—
AbbySorry. You go.
Boyd[snappy] No, you had it.
Abby[statement] I had that buying distressed assets is the most boring thing a company can do. It happens in every industry, every downturn, and quite often it's the only reason the staff and the creditors see a penny. Somebody buys the desks, somebody gets paid.
Boyd[thoughtful] Hm. Fair.
AbbyThe only strange thing is the marketing wrapped round it. You can't say you're inventing the next century and then get excited about a second-hand customer list.
Boyd[statement] You can. [short pause] They do. Right. Moving on.
4:02 Slop, or not: the money-saving router
Boyd[snappy] Slop, or not?
Abby[excited] Slop or not!
Boyd[snappy] The claim: smart routing between a big expensive model and a small cheap one is a large, reliable saving. The summary we were handed says most of that saving is explained by what you happened to ask it, not the router. [short pause] And the link goes to fluid control. Again. [statement] Slop.
AbbyOne caveat, honestly — a modest real saving is still a saving. Not everything unexciting is a lie.
Boyd[statement] It's a savings plan that works best on money you weren't spending.
4:41 Something Useful: your bank statement, in English
Abby[excited] Something Useful. And this one's real and small and boring in the best way. There's an item on the shelf about a tiny model — cheap, runs on almost nothing — that takes S-Q star T-F-C four-four-seven-one L-D-N C-D, and turns it into: the coffee place.
Boyd[surprised] Oh. [short pause] So the killer app is receipts.
AbbyMerchant names on bank statements, for freelancers doing their own books. Someone could use that this afternoon. [statement] Billions of parameters, and the useful one is the one that reads your card gibberish.
Boyd[snappy] What's the catch, Abby.
Abby[thoughtful] The catch is another item on the same shelf, which says models trained on entirely innocent, non-sensitive material can still leak things later. And a bank statement isn't a spreadsheet, it's a diary. Where you were, when, how often.
Boyd[statement] Nothing sensitive went in is not the same sentence as nothing sensitive comes out.
AbbySo — run the small one on your own machine if you can. And don't paste twelve months of transactions into a chat window because it's faster.
Boyd[snappy] It is faster, though.
Abby[laughs] Boyd.
6:06 Something Weird: the accidental jealousy machine
Abby[statement] Something Weird. [short pause] Language models are apparently very good at writing social posts that make you feel envious and inadequate — and completely unable to tell that they're doing it.
Boyd[snappy] Ask it if that post is smug and it says no.
Abby[thoughtful] That's a bit sad, actually. It's not aiming at you. It just can't see you.
Boyd[statement] It isn't trying to make you feel small. [short pause] It's only better at that than it is at noticing.
6:37 Tomorrow
AbbyTomorrow: whether version three of that preprint admits anything.
Boyd[snappy] And whether any of the links go where they say.
AbbyWe're The Daily Slop Show.
BoydNone of this becomes true just because we said it. See you tomorrow.
Listen on Apple Podcasts