---
title: "The bet at the end of the curve"
publication: deslop.media
byline: deslop.media editorial
provenance: editorial_synthesis
provenance_note: "Editorial synthesis by deslop.media editorial, assembled from the subject's public record. The subject did not sit for an interview with deslop.media and has not reviewed it."
issue: 1
issue_title: The Scaling Bet
grade: draft
issue_version: 3
issue_version_date: 2026-10-11
piece_version: 1
first_published: 2026-10-11
reporting_current_through: 2026-10
canonical_url: https://deslop.media/feature/dario-amodei/end-of-the-scaling-exponential/
cite_as: "\"The bet at the end of the curve,\" deslop.media editorial, Issue 1 (Draft v1, October 11, 2026)"
changes: https://deslop.media/issues/1-the-scaling-bet/changes/
flag: "MCP tool flag_content, or POST https://api.deslop.media/api/flags"
---

## The bet at the end of the curve

Dario Amodei has been making the same technical claim since 2017: that doubling compute delivers proportionally smarter machines, that the relationship is log-linear and smooth, and that the field is somewhere near the end of that exponential rather than in the comfortable middle of it. He said as much to the podcaster Dwarkesh Patel in February 2026, "We are near the end of the exponential," and when pressed on whether his view had shifted in the years since he first formed it, his answer was flat: "I actually have the same hypothesis I had even all the way back in 2017."

That is an unusual thing to say in a field where consensus shifts quarterly. It is also the pin the rest of his argument turns on — and that argument extends well beyond AI development. Amodei's essays on biosecurity risk, his advocacy for export controls on advanced chips to China, his arguments about the urgency of AI governance — none of them carry the same weight without a specific, falsifiable technical premise beneath them: that compute scaling delivers smooth, predictable capability gains, and that one to three more doublings will cross the threshold he calls "powerful AI," defined roughly as systems capable of doing the work of a PhD-level researcher across every scientific domain, running in parallel, at data-center scale — a forecast he holds at roughly 90% confidence within ten years, with 2026 at the near edge of the window. He calls this vision "a country of geniuses in a datacenter."

If that curve holds, the urgency of his other positions is well-calibrated. If it flattens, those positions become much harder to defend at their current pitch.

## What the curve carries

Amodei's projections about biomedicine, biosecurity, and chip exports all sit on top of the scaling claim. In *Machines of Loving Grace* (October 2024), he projects that AI could compress decades of progress in biology and medicine into a few years, eliminating most infectious disease and dramatically extending healthy human lifespans. That projection only makes sense if the underlying capability curve produces something genuinely close to scientific-domain expertise at scale. The biosecurity arguments that follow, about AI accelerating the pace at which dangerous pathogens could be designed, require the same premise. So does the export-controls position: if frontier AI does not deliver transformative capability gains in the near term, restricting chip exports to China is an economically costly policy in search of a proportionate threat.

## Two engines

The original scaling thesis, the one Amodei, then at OpenAI, contributed to — codified in the 2020 Kaplan et al. paper *Scaling Laws for Neural Language Models* (arXiv 2001.08361) — is relatively simple: given more compute and more data, neural language models improve on a predictable log-linear curve. The scaling laws, the relationship between model size, training compute, and performance, have since been documented across modalities. Amodei points to analogous findings in images, video, mathematics, and reinforcement learning. The pattern holds.

But by late 2024, a second axis had emerged alongside pretraining scaling. Reinforcement learning applied to chain-of-thought reasoning — the process by which models are trained to show their work and then rewarded for correct answers — turned out to obey its own scaling relationship. Amodei describes this in *On DeepSeek and Export Controls* (January 29, 2025) as a "shifting paradigm" in which "spending $1M instead of $0.1M is enough to get huge gains" on the reasoning curve. This is not a replacement for pretraining scaling; it is a second slope running alongside it.

The two-engine thesis is Amodei's answer to the argument that pretraining might plateau. Even if the gains from raw pretraining scale slow, the RL reasoning curve picks up the slack. And the efficiency dividend, the roughly 4× annual fall in the cost of training a model to a given capability, does not show up as cost savings. It shows up as more capability, because labs reinvest it. (The cost *to use* a fixed level of AI falls faster — Altman puts it near 10× a year — but that larger drop is mostly inference-side optimization and competition, not the training efficiency Amodei means.)

The numbers behind those two rates are not the same numbers, and the difference is easy to miss:

| Annual cost-decline rate | Algorithmic | Efficiency | Hardware | Distillation | Quantization | MoE |
| --- | :---: | :---: | :---: | :---: | :---: | :---: |
| ~1.68× (algorithmic progress alone) | ✓ | | | | | |
| ~4× (training cost-curve, "shifting the curve") | ✓ | ✓ | ✓ | | | |
| ~10× (price to *use* a fixed capability) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |

*What each cost-decline number counts. Amodei's ~1.68× is algorithmic progress alone; his ~4× is the training cost-curve (algorithmic plus efficiency plus hardware); the ~10× Altman cites is the cost to use a given level of AI, which also folds in inference-side gains — distillation, quantization, mixture-of-experts — and competition. The first two measure the cost of training a model to a fixed capability; the third measures the price of using one.*

This is why Amodei found the market's reaction to DeepSeek V3 "baffling." DeepSeek, the Chinese AI lab, had released V3 in December 2024; its reported $6 million training cost, widely read as evidence that cheap Chinese models could neutralize the US compute advantage, looked to him like confirmation: algorithmic efficiency improving on schedule, consistent with the trend, not a disruption of it. "Scaling up the training of AI systems leads to smoothly better results," he wrote in the DeepSeek essay. DeepSeek, in his reading, was another data point on the same curve.

Amodei's public framing maps model generations to an educational ladder — high-school-level performance in 2022, undergraduate capability in 2023, PhD-level reasoning by late 2024 — each rung arriving on a schedule he describes as predictable and compute-driven. The ladder is his own characterization, not a benchmark ranking, and the claim embedded in it is that the curve has been holding rung by rung for three model generations.

The revenue picture has moved in step with the cost picture. Anthropic's run-rate revenue has gone from near zero in 2023 to a roughly $9 billion run-rate by the end of 2025 and past $47 billion by May 2026, the company said in announcing its $65 billion Series H round (Anthropic, May 28, 2026) — a curve that mirrors the training-cost trajectory closely enough that the business case for continued scaling has not yet come apart from the technical case. Companies spending more on frontier training are, so far, building larger businesses. The reinforcing loop has not broken.

## The objections, named

Rich Sutton, the reinforcement-learning pioneer whose 2019 essay "The Bitter Lesson" became the intellectual foundation of the scaling thesis (it argues that computation always wins over hand-crafted methods), is not persuaded that current transformer-based large language models are the vehicle that carries the thesis through. Amodei has engaged the point directly. On the Dwarkesh podcast in February 2026, responding to Patel's description of Sutton as "very non-LLM-pilled," he conceded Sutton's claim that humans learn far more from far less data and that biological learning is dramatically more sample-efficient than pretraining, and argued that pretraining exists "in the middle space between human evolution and human on-the-spot learning" on the efficiency spectrum — good enough, in his telling, to keep improving. ([Sutton's own case is that the scaling thesis is misread](/feature/richard-sutton/bitter-lesson-misread/).)

<!-- figure:primary:top -->

The sharper objections say the curve is not doing what Amodei says it is. Yann LeCun, who left his post as Meta's chief AI scientist at the end of 2025 to lead AMI Labs as executive chairman, argues that large language models will "never" reach human-level intelligence, on architectural grounds: they lack world models, and the sample-efficiency gap between transformers and biological learning is a structural difference that more compute does not close.

François Chollet, the Keras creator who designed the ARC-AGI benchmark — a test of fluid intelligence, the capacity to form new abstractions from minimal examples — pointed to a harder fact: when ARC-AGI-3 launched in March 2026, frontier models, the same systems Amodei describes as approaching PhD-level research capability, scored below one percent while humans cleared it routinely. Scores began to rise that summer, and on September 2, 2026, ARC Prize reported GPT-6 Astra at 62.7 percent on ARC-AGI-3's Semi-Private set under its Standard harness at max reasoning effort.

Gary Marcus has put the disagreement in falsifiable form, offering Amodei a $100,000 bet in January 2025 that there will be no AI "smarter than almost all human beings" by the end of 2027. LeCun, Chollet and Marcus are each making the same structural claim from a different angle: that smooth capability scaling and genuine generalization are not the same thing.

## The test he has already named

In an August 2023 interview with Dwarkesh Patel, before the current round of arguments about DeepSeek and ARC-AGI, Amodei was asked what result would change his mind about scaling. His answer was specific:

> "if you told me, yes, you trained your 2024 model, it was much bigger and it just wasn't any better, and you tried every architecture and didn't work, that's the explanation I would reach for."

— *Dwarkesh Patel podcast, Aug. 2023*

This is the falsification condition he has already stated publicly. He framed the test around a 2024 model. Its logic carries forward: a frontier-scale run, swept across architectures and data, that comes out no better than a much smaller predecessor is the result that would, by his own account, change his view.

He has not said that result has occurred. No one else has reported that it has occurred. The window he named is still open: by his estimate, "powerful AI" could arrive as early as 2026. The curve is either holding or it is not, and the answer will be visible in the capability of systems that are, by the timeline he has held since 2017, being trained right now.

Amodei's position is that the two engines of pretraining and RL reasoning are both running, and that, as he put it on Lex Fridman's podcast in November 2024, "there's no ceiling below the level of humans" in the underlying scaling relationship. That claim is what the falsification condition tests.

The test he named is still running.

## Sources

- darioamodei.com/essay/machines-of-loving-grace
- darioamodei.com/post/on-deepseek-and-export-controls
- lexfridman.com/dario-amodei-transcript (Lex Fridman #452, Nov. 2024)
- dwarkesh.com/p/dario-amodei-2 (Feb. 13, 2026)
- dwarkesh.com/p/dario-amodei (Aug. 2023)
- arxiv.org/abs/2001.08361
- x.com/GaryMarcus/status/1884649584603127890
- arcprize.org/results/openai-gpt-6-astra (September 2, 2026)
- arcprize.org/blog/arc-agi-3-launch (March 2026)
- techcrunch.com/2025/12/19/yann-lecun-confirms-his-new-world-model-startup-reportedly-seeks-5b-valuation/