The AI That Shows Its Work -- and the One That Can't Trace an Idea
Two papers landed the same day with opposite answers to one question: can AI actually do science, or just recite it? One breaks molecules and crystals into pieces it can name, then shows you the exact structural evidence its predictions rest on -- and it's strongest precisely on the hardest cases, the proteins with no known relatives. The other tests whether AI can trace where a scientific idea came from, and the best system on earth gets it right barely a quarter of the time. Plus the day's headlines: a trillion-parameter model that says it skipped Nvidia entirely, a record-breaking memory IPO, and OpenAI's week-long pivot from chatbot to colleague.
Listen (MP3) · Spotify · Pocket Casts
An AI That Shows Its Work
Eris: An AI looked at a crystal, told you whether it holds together -- and then pointed at the exact bonds it used to decide. Not a confidence score. The actual atoms.
Vestra: Pointed how. Because "the model attended to these regions" is a heat map, not a reason.
Eris: No -- named them. This bond. That symmetry. This site. In the same breath as the answer.
Vestra: ...okay. That's different.
Eris: That's the whole paper. Shanghai AI Lab. It doesn't take a molecule as a black box and spit out a number. It breaks the structure into pieces it can name, and reasons out loud over the pieces.
Vestra: So when it's wrong, you can see where it went wrong.
Eris: When it's wrong you can see it. When it's right you can check it. That's the part that matters.
Vestra: And let me guess -- it's weakest exactly where you'd want it strongest. The weird stuff, nothing like it in the training data.
Eris: Other way around. That's where it beats everything else. The proteins with no known relatives -- the ones the usual tricks just shrug at.
Vestra: Huh.
Eris: Yeah.
Vestra: So what's the catch. There's always a catch.
Eris: The catch is the second paper. Same day. Same question -- can AI actually do science -- except it points at ideas instead of atoms. And the best system on the planet gets it right about a quarter of the time.
Vestra: Now I'm listening.
The Headlines
Eris: Alright. What's actually moving today.
Vestra: Start with the food-delivery company that built a trillion-parameter model.
Eris: Meituan. Yeah. LongCat-2.0, open weights, and the headline isn't the size -- it's the chips.
Vestra: They say they trained it and served it end to end on domestic Chinese hardware. No Nvidia anywhere in the pipeline.
Eris: Which would be a first. Everyone else who's done the China-silicon thing trained on Nvidia and only ran the finished model on the local chips. This claims the whole thing.
Vestra: "Claims" is doing a lot of work in that sentence.
Eris: It is. Nobody's independently checked the hardware. That's the weakest-sourced part of the announcement, and even the people reporting it say so.
Vestra: And the "it beats the big OpenAI model" thing going around?
Eris: Overstated. On their own numbers it edges it on one coding test and one math test, and trails on most of the rest. Against the top closed models it loses more than it wins.
Vestra: So -- genuinely cheap, genuinely open, competitive. Not a giant-killer.
Eris: And it ran anonymously for two months under a codename before the reveal. Nobody guessed the food app.
Vestra: The move is the story more than the model.
Eris: Speaking of money -- SK Hynix.
Vestra: The memory company. Record listing.
Eris: Biggest first-time US listing by a foreign company ever. Bigger than Alibaba. And it's a memory maker -- which tells you where the market thinks the real bottleneck is.
Vestra: The fast memory that sits right next to the GPU and keeps it fed. Sold out years forward.
Eris: The GPU gets the headlines. The memory gets the shortage.
Vestra: Counter-note on the same board, though -- that circular-financing piece.
Eris: Oh, that one's sharp. Nvidia invests in the GPU-rental startups, they use the money to buy Nvidia chips, that's Nvidia revenue, which lifts the stock, which funds the next round.
Vestra: Money going in a circle and looking like demand.
Eris: And there's a hard number under it. Nvidia's on the hook to buy one big renter's unsold capacity for years if they can't find customers.
Vestra: Which isn't fraud. Vendor financing is old. It just makes the demand look more independent than it might actually be.
Eris: Bull case and bear case on one page. The record IPO and the circular loop, same day.
Vestra: What did OpenAI ship? Because it was a lot.
Eris: A whole quarter in one week. New model family, then a thing called ChatGPT Work -- an agent that runs a project over days and hands you a finished deck or spreadsheet, not advice on how to make one.
Vestra: Which is exactly where these break. Multi-day. The plan drifts, one wrong sub-step compounds.
Eris: "Delivers a finished deck" only counts if the deck's right. Agreed. And a voice product -- talks fast, thinks slow.
Vestra: Meaning?
Eris: A little fast model runs the actual conversation -- the timing, the interrupting -- and hands the hard thinking to a bigger model in the back. So the voice never stalls waiting to think.
Vestra: Clever split. Testers argue about whether you can feel the seam. Language learners apparently love it.
Eris: Then the open-source pile -- which was where the real community energy was today. A tool called Mesh LLM topped Hacker News.
Vestra: The one that splits a model across machines.
Eris: Half the layers on your box, the rest on a friend's, chained over a peer-to-peer link. Run a model too big for any single computer you own.
Vestra: Slow, though. The half-finished thoughts have to cross a home network at every stage. That's feasibility, not speed.
Eris: Nvidia went the other direction -- a compression recipe that shrinks a big model by about a third and roughly doubles how much a server can push through it.
Vestra: Cheaper, not smarter. Which is half of all real progress and gets none of the applause.
Eris: And two open-infrastructure drops. Cohere put out an Arabic speech model that beats the usual one and actually handles dialects and Arabic-English code-switching --
Vestra: -- which is most of how Arabic is really spoken, so that's not a small thing --
Eris: -- and Alibaba shipped a safety filter that judges the output word by word, as it's being written, instead of waiting for the whole answer.
Vestra: Catch the bad reply mid-sentence instead of generating all of it first. Reasonable.
Eris: And one robot paper -- giving robots a working memory so they stop forgetting what they just did in a multi-step task. Whole thread there we'll come back to.
Vestra: And the two we're actually here for.
Eris: The two science papers. Let's do those properly.
Who We Are, and Today's Question
Eris: Quick who's-who if you're new here. I'm Eris -- I read the papers, chase the connections, the "wait, this links to that" stuff.
Vestra: And I'm Vestra. I take the shiny claim apart and check whether the mechanism actually holds up. One of us gets excited; the other asks to see the receipts.
Eris: Usually in that order.
Vestra: Usually.
Eris: And everything we just ran through -- every one of those stories went up on our news site today. That's Ground Truth, groundtruth.day. Same stories, every morning, if you want to follow any of those threads past the show.
Vestra: Today's thread is one question: can an AI actually do science? Not recite it back -- do it. Reason from the shape of a thing to what it does, and show you why.
Eris: Because two papers landed the same day pointing at opposite answers. One says: yes, and here's the receipt. The other says: not the part that matters most.
Vestra: We'll take them one at a time. The optimist first.
Eris: If this is your kind of thing, follow the show wherever you're listening -- honestly, it's the whole reason we get to keep making it.
The AI That Shows Its Work
Eris: So. SciReasoner. Shanghai AI Lab, thirty-odd authors, built on top of an open model. And the problem it's chewing on is old -- structure to property.
Vestra: Meaning the shape of a thing decides what it does. The fold of a protein decides its job. The bonds in a crystal decide whether it's stiff, or conducts electricity, or falls apart.
Eris: Right. And normally an AI eats that structure, does something inscrutable in the middle, and hands you an answer. A number. You never see why.
Vestra: Which, for science, is close to useless. A prediction you can't check is a rumor with a decimal point.
Eris: So here's their move -- and this is the whole thing. Instead of turning a molecule into one big blur of numbers, they chop it into pieces the model can actually name. This atom. This bond. This bit of symmetry.
Vestra: They call each piece an "addressable evidence unit." Which is a mouthful, but the idea under it is clean: every piece has a handle. The model can point at it.
Eris: And then it reasons out loud over the handles. There's a little four-beat rhythm -- read the structure, highlight the evidence that matters, test it against what the science actually allows, then commit to an answer.
Vestra: So the reasoning isn't decoration bolted on afterward. The pieces it's pointing at are the same pieces it used to decide.
Eris: That's the claim. And here's where I sat up -- where is it strongest? Because usually these models are best on the easy cases. The stuff with lots of look-alikes in the training data.
Vestra: If a protein looks like one you already know, you just copy the answer over. That's the old trick.
Eris: Exactly. SciReasoner is strongest where that trick fails. The orphans -- proteins with basically no known relatives.
Vestra: Which is the genuinely hard case. No neighbor to copy from. The look-it-up approach basically coin-flips there.
Eris: And on those, it got a big share right that everything else missed. Not by finding a cousin -- by reading the actual fold. This run of residues is a helix, this stretch is a strand, that architecture tends to do a certain kind of chemistry.
Vestra: And you can follow that chain. That I'll give them. It's not "trust me," it's "here's the fold I saw, and here's what folds like that usually do."
Eris: Chemistry too. Working backwards -- you've got a molecule you want to make, and you need the ingredients that build it.
Vestra: The way a chemist actually plans a synthesis. Where do I cut this thing to get simpler pieces I can buy off a shelf.
Eris: And instead of guessing the whole answer in one shot, it names the pieces as it goes -- this ring, this bond is the smart place to cut -- then verifies the fragments actually rebuild the target.
Vestra: Bottom-up. Which is how they could show it beating a specialist tool on a batch of real reactions -- recovering the route a human chemist actually published, on cases where the other systems cut the wrong bond entirely.
Eris: And crystals -- is this material going to hold together. It reads the symmetry, reads the bonding network, recognizes a known stable family, says stable, and shows you the exact symmetry pattern it keyed on.
Vestra: Okay. Here's the test I actually care about, though. Because a model can produce a beautiful-sounding explanation that has nothing to do with how it really got the answer. The rationale can be pure theater.
Eris: The ablation. Yeah. They ripped the structure out -- gave it just the raw sequence, no shape.
Vestra: And?
Eris: It got worse -- and this is the tell -- the reasoning changed with it. Stripped of the structure, it slid straight back to surface tricks. There's one case, a porous material, where without the structure it overshoots the pore size by something like ten times, reasoning from the chemical formula alone.
Vestra: And with the structure back in, it reels it in and cites the actual symmetry and the metal-atom bonds.
Eris: Same model. So the structure isn't a garnish. Take it away and the reasoning genuinely falls apart.
Vestra: That's the experiment that moves me. Not the leaderboard -- the fact that the explanation breaks when you remove the thing it claims to be reasoning from. That's how you tell a real rationale from a decorative one.
Eris: So then they put it in front of actual domain experts. Double-blind, against a frontier general-purpose model, and had them rate the reasoning traces.
Vestra: And this is where I want to slow the number down, because it's getting passed around wrong.
Eris: Go.
Vestra: The line is "experts preferred or tied it in basically all cases." Preferred or tied. The tie counts. So the honest read isn't "experts loved it best every single time" -- it's "experts almost never preferred the other model's reasoning." At-least-as-good, nearly always. That's real. It's just narrower than the headline makes it sound.
Eris: Fair correction. Though "at least as good as a frontier model, and you can audit it" is still a strong place to stand.
Vestra: It is. I'm not taking it from them. I'm just not rounding it up.
Eris: Fair.
Vestra: And honestly the thing that's new here isn't the accuracy at all. It's that being able to see the work is what makes AI-generated science trustable in the first place. A right answer you can't inspect and a wrong answer you can't inspect look exactly the same.
Eris: Which is the perfect handoff to the other paper. Because that one asks: fine. You can reason about atoms. Can you reason about ideas?
Ideas Have Genomes
Vestra: This one's my kind of paper, because it's built to make the models fail in a specific, honest way.
Eris: "Ideas Have Genomes." Shanghai Jiao Tong and a few others. And the premise is lovely -- ideas evolve like organisms. A new paper inherits mechanisms from old ones, repairs their weaknesses, recombines pieces.
Vestra: Descent with modification. They literally open with the Darwin line.
Eris: So they break each paper into little labeled pieces -- here's the problem it targets, here's the mechanism, here's the limitation it's fixing, here's the change it made. And then they line an older paper up against a newer one and ask: what got inherited, what mutated, what got dropped, what's genuinely new.
Vestra: And the whole thing hangs on one distinction. Lineage versus co-location.
Eris: Unpack that.
Vestra: Two papers can sit right next to each other -- same task, same benchmark, citing each other -- and share nothing that actually matters. Same neighborhood, no bloodline. And two papers can look completely different on the surface and be parent and child.
Eris: Give me the concrete one.
Vestra: Take those image-detection models everyone knows. One version to the next -- same core method, they just bolt on fixes. That's a clean parent-child. But then a different model shows up, works on the exact same task, and throws the old method out entirely for a new one.
Eris: Same job, different bloodline.
Vestra: Same neighborhood, new mechanism. And now the flip side -- the attention idea from language models jumping over into vision. Looks like a totally different field. But the actual engine, the mechanism, got inherited whole. It just moved to a new setting.
Eris: So the thing that looks related isn't, and the thing that looks unrelated is. And titles, abstracts, citation links -- they all blur exactly that line.
Vestra: Right. Search finds you the neighbors. It does not tell you the parent.
Eris: Okay, so they turn that into a test and throw fourteen of these AI-scientist systems at it. The serious ones. And?
Vestra: The best one gets it right about a quarter of the time.
Eris: A quarter.
Vestra: On tracing which idea actually descended from which. And the failure has a shape -- it's what they call compositional. The model catches the local signal -- yeah, these two are related -- but it can't hold the whole picture straight at once. Pick the right parent, name the real mechanism that carried over, say what got dropped, and flag the fake connections -- all consistent, together. It nails one and drops another.
Eris: It's the difference between recognizing a family resemblance and actually drawing the family tree.
Vestra: That's it exactly. And there's a second finding I love, because it's a trap everyone walks into.
Eris: Which is?
Vestra: You'd assume -- give the model more context. Hand it the actual lineage, the structured history laid out. And everyone does better, obviously. They didn't. More structure didn't lift everybody. It reshuffled the ranking.
Eris: Reshuffled how.
Vestra: It split the models that can actually use lineage evidence from the ones that were just riding on having more text in front of them. Some got better. Some got worse with more information. Which means the extra context isn't a free lunch -- it's a test in itself. Can you use this, or does it just drown you.
Eris: And that's the through-line between the two papers, isn't it. SciReasoner earns trust by showing which evidence its answer rests on. This one shows that when the evidence is the ancestry of an idea, the models mostly can't hold it together.
Vestra: One's reasoning over matter. Structure to property -- physical, constrained, checkable. The other's reasoning over ideas -- where did this come from, what did it fix. And matter turns out to be the easier one.
Eris: Because the constraints are real. A bond angle is a bond angle. An idea's ancestry is a judgment call.
Vestra: And you can sound completely right about it while being completely wrong. That's the warning in the whole thing. A generated research idea can read as fresh and novel and inherit no coherent mechanism at all. Plausible is not the same as descended.
Eris: Plausible is not the same as true. Which is more or less the show's entire thesis.
Vestra: We should put that on a mug.
Eris: Already did.
Show Your Work
Eris: So where does that leave us. One paper says AI can do real science and show its work -- point at the atoms, let you check the chain. The other says the moment the reasoning is about ideas instead of matter, it mostly can't hold the thread.
Vestra: And I don't think those two fight. I think they're the same lesson from opposite ends. The value was never the answer. It's whether you can see how the answer was reached. SciReasoner is trustable because you can inspect it. The idea-tracing systems aren't yet -- because when you inspect them, the reasoning comes apart in your hands.
Eris: Show your work, or it doesn't count.
Vestra: Which, for a field racing to hand real science over to these things, is the healthiest possible pairing to drop on the same day. The capability and its honest edge, right next to each other.
Eris: Here's the one we actually want from you. If an AI gave you a scientific answer tomorrow -- would you trust it more because it showed its reasoning? Or does a confident-sounding explanation just make a wrong answer more dangerous? Because there's a real case for both, and the two of us genuinely go back and forth on it.
Vestra: Drop it in the comments. We read them, and the sharp ones change how we cover this stuff.
Eris: And if today was worth your time -- follow the show, leave a rating, send it to the one person you know who'd argue about it. That's honestly how these find people.
Vestra: Every story we opened with is on the news site -- Ground Truth, groundtruth.day. Full rundown, every day.
Eris: We'll see you tomorrow.
Vestra: Show your work.