June Hadley: Roy, before we get into anything — has the phrase 'scaling laws' started to sound like incantation to you, or is that just me?
Roy Halliday: Only when people use it to mean everything at once.
June Hadley: Right — and I think that's exactly the problem I want to crack open. There's a specific fact I came across that made it concrete for me. A 13-billion-parameter model at 4-bit precision outperforms a 7-billion-parameter model at full 16-bit precision. More parameters, dramatically less numerical fidelity, still wins. Which means raw capacity is doing something that careful precision cannot undo.
Roy Halliday: That's the compression story — but the deeper implication is that scale isn't one thing.
June Hadley: Exactly. So — let's see — Kaplan et al. at OpenAI established in 2020 that you get power-law relationships between model size, training data, compute budget, and loss, and they hold across many orders of magnitude and are largely independent of architectural specifics. That's the smooth, reproducible, almost boring part.
Roy Halliday: Boring and powerful. Depth versus width barely matters — scale is the dominant lever. That's Kaplan's actual finding and it's underrated as a result.
June Hadley: But then Jason Wei defines emergent abilities as capabilities that are absent, or near-random, in smaller models — and appear abruptly in larger ones at thresholds that cannot be predicted by smoothly extrapolating the curve. That's not the same phenomenon at all.
Roy Halliday: And Schaeffer et al., 2023, come back at Wei with this: swap your coarse binary metric for a continuous one and the sharp jump flattens. What looks like a phase transition might just be your measurement choosing to see a cliff.
June Hadley: The mirage argument. I think — okay, I don't think it's fully defeated, but chain-of-thought reasoning is the hard case for Schaeffer. The ability to hold multi-step reasoning structure across a long problem. That doesn't feel like a metric shadow.
Roy Halliday: Frankly, Wei's camp would agree with you. The debate remains structurally open.
June Hadley: Which is what I find philosophically alarming — in a useful way. We have a reproducible law and simultaneously no reliable way to forecast what emerges when. Those two things are both true at once.
Roy Halliday: You can model compute spend to the basis point. You cannot model what thinking appears at the other end. That's not a science communication problem — that's a structural one. For every institution scaling right now.
June Hadley: But that structural problem only lands hard if you understand why the laws felt so solid in the first place. And I think most people stop at the wrong level of that explanation.
Roy Halliday: The exam analogy. Kaplan said: get a bigger brain. More parameters, better performance. That was the whole frame.
June Hadley: Right — but Hoffmann et al. at DeepMind came back in 2022 and said: actually, you also need to study more. The brain size and the study time need to grow together.
Roy Halliday: That's Chinchilla optimality. Ten to forty tokens per parameter. Under a fixed compute budget, you scale both proportionally — not parameters alone.
June Hadley: And the proof was a 70-billion-parameter model that beat much larger models trained the old way.
Roy Halliday: Chinchilla, yes. Seventy billion parameters, optimally balanced data — outperforms models with significantly more parameters that were trained under the Kaplan framing. That's not a marginal win.
June Hadley: So — I mean, how do you make that concrete? Because ten-to-forty tokens per parameter is almost absurdly specific as a ratio.
Roy Halliday: Warehouse logistics company, 2021. They're budgeting their next model run. Under Kaplan, they spend almost everything on parameters, almost nothing proportionally on data. Same compute dollars. After Chinchilla, they split that budget roughly evenly. Same spend, meaningfully better model.
June Hadley: Wait — same compute, different allocation, better output. That's what moved OpenAI and DeepMind's actual capital decisions.
Roy Halliday: One empirical finding. Years of parameter-centric industry practice, revised.
June Hadley: And that's what I think is the real hook here — not just that Kaplan et al. established power-law relationships that hold across architectures. It's that the field read those laws, built around them for years, and was still reading them wrong.
Roy Halliday: The laws were reproducible. The interpretation was off. Architecture-independence, scale as the dominant lever — all of that held. But the allocation the field derived from it? Skewed badly toward parameters.
June Hadley: So the boring, reproducible fact — the one that felt settled — was still generating bad decisions at scale. Which means the question isn't only whether the laws hold. It's whether we know what they're actually telling us.
Roy Halliday: And that gap — between what the laws tell us and what we think they're telling us — is exactly where Jason Wei's emergence finding lands hardest.
June Hadley: Say more about that. Because the formal definition is precise — absent or near-random in smaller models, appears abruptly at larger ones, cannot be predicted by extrapolating the smooth loss curve. That's Wei's claim. But I want to know what 'abruptly' is actually doing in that sentence.
Roy Halliday: It means the loss curve gives you no warning. You're watching loss decline smoothly — predictably — and then at some scale threshold, chain-of-thought reasoning just appears. Multi-step problem, intermediate steps generated, answer correct. Wasn't there at smaller scale. No gradient in the metric that told you it was coming.
June Hadley: Which is where Schaeffer et al. push back — because 'wasn't there' might mean the metric couldn't see it yet.
Roy Halliday: Right. Schaeffer's argument: swap the coarse binary scoring for a continuous measure, the cliff flattens. The capability was improving all along — you were just measuring with a blunt instrument.
June Hadley: Okay but — I mean, that works for some tasks. For something like arithmetic, sure, partially-right answers exist. But in-context learning? The model either performs a new task from prompt examples without weight updates, or it doesn't. There's not an obvious halfway state.
Roy Halliday: That's the defenders' strongest case. In-context learning has a structural property — no gradient updates — that makes it categorically distinct, not just quantitatively better.
June Hadley: Wait — so the argument isn't just 'the jump was sharp.' It's that the thing that appeared is qualitatively different in kind from what came before.
Roy Halliday: That's the phase transition analogy. Small quantitative change — more parameters, more tokens — sudden qualitative shift. Water doesn't get 'a little frozen.' The phase transition literature in physics uses exactly this framing. The problem is it's still analogical. Nobody has formally derived it.
June Hadley: So we're in the pre-Darwinian naturalist position again. Naming the shape of the thing without a mechanism for why it has that shape.
Roy Halliday: Correct. And that unresolved question has a consequence nobody's pricing correctly — which we'll get to, but the mechanistic black box doesn't just matter scientifically. It means no one can tell a board or a regulator what capability will appear at the next order of magnitude.
June Hadley: And that's — I think that's the real weight of the debate staying open. If Schaeffer is right, the field is more predictable than it looks. If Wei is right, zero-shot generalization just appearing at some threshold we didn't forecast — that's not a surprise you can manage. That's a structural blind spot.
Roy Halliday: The field has both possibilities on the table simultaneously. And no clean experiment that settles it.
June Hadley: Which means every institution scaling now is making a capital bet on which version of reality is true — and I'm not sure most of them have named that bet explicitly.
Roy Halliday: That bet doesn't get named because naming it requires admitting what we don't have — which is any mechanistic account of why scale produces new capability at all. Knowledge compression is the story the field tells. Larger models compress knowledge more efficiently, generalize to novel problems. Intuitive. Clean. Not uniformly endorsed in the literature.
June Hadley: Wait — not uniformly endorsed, meaning the sources themselves are hedging?
Roy Halliday: Meaning it describes the outcome. It doesn't explain the mechanism. There's a difference.
June Hadley: I think — okay, that's the pre-Darwinian problem stated precisely. We can catalog that chain-of-thought reasoning appears. We can catalog that in-context learning appears. We cannot say why scale causes either to appear. We're naturalists with a very good specimen drawer.
Roy Halliday: And the phase-transition analogy from physics — which recurs constantly — nobody has formally derived the structural equivalence. Water freezing has a thermodynamic account. Neural network emergence does not. It's a narrative filling a gap.
June Hadley: So we're borrowing the shape of an explanation without the explanation itself.
Roy Halliday: Correct. And that matters the moment you leave the lab. The compute budget — you can model that. Hoffmann et al. gave us the ratio, ten to forty tokens per parameter, compute-optimal, done. Boards can price that. Regulators can track that.
June Hadley: But the capability output from that spend — actually, no, let me say it more precisely — you cannot predict which capabilities appear, when they appear, or at what parameter threshold. Wei's point is exactly that the loss curve gives no warning.
Roy Halliday: And unpredictability of future thresholds isn't just a technical admission — it's structural. Task complexity, training data distribution, architecture. They interact in ways we cannot see through. Past emergence events don't give you a reliable map.
June Hadley: Which means a safety team that built its entire evaluation process on gradual, monitorable capability growth — I mean, that assumption is quietly load-bearing, and it may just be wrong.
Roy Halliday: Standard deployment policy assumes you see the capability coming. That you have time to evaluate, adjust, gate. If zero-shot generalization or chain-of-thought reasoning can appear suddenly at a scale nobody forecast — that entire policy architecture is built on a premise that may not hold.
June Hadley: Hold on. Is that a falsifiable worry, or — I mean, Schaeffer's argument, if it's right, actually rescues the safety team. Gradual improvement, just poorly measured.
Roy Halliday: Look — if Schaeffer is right, the problem shrinks. If Wei is right, boards and regulators have no instrument for pricing actual capability risk. They can model the compute spend to the dollar. They cannot model what thinking emerges from it. That asymmetry is the structural indictment. And the debate is still open.
June Hadley: So the scariest version isn't that we don't know the mechanism. It's that we've built governance around the part we can measure — and left the part we can't measure entirely unaccounted for.
Roy Halliday: And that's the admission embedded in the research itself. Kaplan gives you the loss curve. Hoffmann gives you the ratio. Neither gives you a forecast for when zero-shot generalization appears, or chain-of-thought, or in-context learning. The measurement framework is genuinely excellent at the thing it cannot see past.
June Hadley: I keep returning to the Schaeffer position — not because I think it settles anything, but because if it's right, that's actually a different kind of problem. It means the field has been asking a real question about a real phenomenon and the answer has been sitting in the measurement choice the whole time. And if Wei is right — I mean, genuinely discontinuous, genuine phase transition — then the next frontier isn't faster scaling. It's building predictive theory for what emerges when, before we build it. Before the capability exists to evaluate.
Roy Halliday: The mirage might be real. The mirage might be the mirage. That's where it sits.
June Hadley: I can't put that down. Whether emergence is genuine or artifact — that distinction determines whether AI development is a plannable engineering discipline or a series of surprises we're retrospectively naming. Those are not close positions. And we don't have a clean experiment.
Roy Halliday: No. We don't. Frankly, that's a reasonable place to stop.
