Dario Amodei 在一个周六呼吁整个行业放慢脚步1。但“放慢脚步”这个词究竟意味着什么?
这项提案承诺对 AI 进行严格的审计监督、推动行业集体行动来调控创新节奏,并在盟友国家之间开展国际协调。调控节奏是一个政策问题。五大阵营都权衡了其后果,却没有一个阵营给出具体的速度。
- 可解释性阵营想要时间。Anthropic 负责对齐科学的 Evan Hubinger 认为,十年内 AI 对人类造成灾难性后果的概率超过 10%,并表示目前尚无应对方案2。可解释性——即理解模型内部发生了什么的科学——“并不总能产生清晰可靠的结果”,Amodei 写道。“我们仍然只了解这些模型内部运行机制的极小一部分。”1
- 劳工阵营想要为劳动者争取时间,而不是为研究人员。Bernie Sanders 已提出一项联邦暂停令,要求在安全保障措施出台之前,暂停新建 AI 数据中心。3
- 经济阵营认为,AI 对 GDP 的贡献是化解国家债务的途径。[45“世界正深陷债务之中,”Bessent 在 G20 会议上表示,“而我们摆脱困境的唯一办法就是靠增长走出困境……这些资本支出将转化为生产力,而这将极具抑制通胀的作用。”他说,“我估计在未来六个月内,我们将开始看到这方面的收益。”[6这是所有阵营中唯一一个给出了时间点的预测。
- 地缘政治阵营则寻求在 AI 竞赛中保持战略优势。“我们在 AI 领域领先于中国,”特朗普周日表示。“谁赢得 AI,谁就赢得一切。”[7
- 监管俘获阵营根本不想要任何新规则。一旦制定规则,竞争对手就得花钱去遵守。总统科学技术顾问委员会主席 David Sacks 直接对 Amodei 和 Altman 说道:“别再假装你们需要任何人的许可。别再假装必须暂停反垄断法你们才能结成卡特尔。”他写道,如果这些实验室不肯自行放慢脚步,“我们就会知道这不过是又一次寻求监管俘获的企图。”8 在 2025 年之前一直担任 FTC 主席的 Lina Khan 则从相反的方向得出了同样的结论。“我们不该让关于新法律制度的讨论转移视线,掩盖一个事实:已有法律对 AI 没有任何豁免。”9
这些视角仍然没有回答这个问题:到底什么叫“控制节奏”?
2025 年 1 月,一家竞争对手在 41 天内追平了前沿水平。[10如今,领先期可持续数月,而开源与闭源模型之间的差距已从三个月扩大到四个月。[11这个间隔应该是 60 天?160 天?还是 600 天?
我们真能把创新速度精确控制到这种程度吗?
Amodei 明确指出了调控的杠杆。他主张限制进入前沿模型的要素:训练算力以及训练运行的性质。[1算力阈值已有先例。2023 年的一项行政命令要求对超过 10^26 FLOPS 的情况进行报告,但不久后即被撤销。[12在该规则生效期间,没有任何模型越过这条线。第一个越线的是 Grok-3,它在规则撤销数周后发布,而今年将有大约十个模型越过这条线。[13训练算力每年增长约五倍。一个固定的数字只是一道门槛,而地板会自己够到它。
问题不仅在于多快,还在于由谁决定。
-
[Dario Amodei,We Must Pace the Frontier,2026 年 9 月。[↩︎ [↩︎ [↩︎
-
Bernie Sanders,人工智能正在冲击工人阶级。我们必须奋起反击,2026 年 4 月 16 日。其中的就业数据是 Sanders 引用斯坦福大学关于编程与客服岗位的研究。↩︎
-
国会预算办公室(Congressional Budget Office),《预算与经济展望:2026 至 2036 年》,2026 年 2 月。到 2030 年,公众持有的债务将超过 1946 年创下的 106% 的前历史高点。↩︎
-
CBO 的长期敏感性分析,参见 Gale & Auerbach, The Federal Budget Outlook 总结。非农企业全要素生产率上升 0.5 个百分点,到 2055 年可使债务占 GDP 比重降低 43 个百分点。 ↩︎
-
Scott Bessent 在 G20 新闻发布会上的发言,北卡罗来纳州阿什维尔,2026 年 9 月 2 日,完整记录;The Straits Times 亦有报道。 ↩︎
-
Donald Trump 于 2026 年 9 月 13 日在爱尔兰 Doonbeg 对记者的表态,来源:Associated Press。 ↩︎
-
David Sacks,@DavidSacks,2026 年 9 月 13 日。↩︎
-
Lina Khan,@linamkhan,2026 年 9 月 13 日。↩︎
-
Tomasz Tunguz,《Darwin's Finches in AI》,2025 年 1 月。该时间间隔衡量的是前沿模型保持领先、直到竞争对手追平它所持续的时间。↩︎
-
开放模型落后于最先进的闭源模型 4 个月,Epoch AI,2026 年 5 月 29 日。Epoch 还发现,自 GPT-4 以来没有任何模型领先满一年;第二长的统治期是 o1 的,也只持续了不到三个月。↩︎
-
第 14110 号行政命令,2023 年 10 月 30 日,该命令依据《国防生产法》设定了 10^26 FLOPS 的报告阈值;于 2025 年 1 月 20 日被撤销。↩︎
-
Epoch AI,《How many AI models will exceed compute thresholds?》,2025 年 5 月。2025 年 2 月发布的 Grok-3 是首个据估算超过 10^26 FLOP 的模型;Epoch 的中位数预测是到 2026 年约为十个、到 2027 年初约为三十个。自 2020 年以来,前沿训练算力每年增长约 5 倍。↩︎
Dario Amodei asked the industry to pace itself on a Saturday1. But what does that word mean?
The proposal promises auditors examining AI closely, collective action in the industry to meter the pace of innovation, & international coordination across allied countries. Pacing is a policy question. Five camps priced the consequences & none named a speed.
- The interpretability camp wants time. Evan Hubinger, who leads alignment science at Anthropic, puts the odds that AI goes catastrophically wrong for humanity above 10 percent within a decade & says there is no plan for it2. Interpretability, the science of understanding what happens inside a model, “doesn’t always produce clear & reliable results,” Amodei wrote. “We still only understand a tiny fraction of what goes on inside these models.”1
- The labor camp wants time for workers, not for researchers. Bernie Sanders has introduced a federal moratorium on new AI data center construction until safeguards are enacted.3
- The economic camp sees AI’s contributions to GDP as the path to managing the national debt.45 “The world is awash in debt,” Bessent said at the G20, “& the only way for us to get out of this is to grow our way out of this… This CapEx will turn into productivity, & that will be extremely disinflationary,” he said. “I would guess that in the next six months we will start seeing the benefits of that.”6 It is the only dated prediction any camp has made.
- The geopolitical camp seeks to sustain strategic advantage in the AI race. “We’re leading China in AI,” Trump said Sunday. “Whoever wins AI wins.”7
- The regulatory capture camp wants no new rules at all. Write the rules & your rivals pay to obey them. David Sacks, who chairs the President’s Council of Advisors on Science & Technology, put it to Amodei & Altman directly : “Stop pretending you need anyone else’s permission. Stop pretending antitrust law has to be suspended so you can form a cartel.” If the labs will not slow down on their own, he wrote, “we’ll know this was just another bid for regulatory capture.”8 Lina Khan, who chaired the FTC until 2025, arrives at the same place from the opposite direction. “We shouldn’t let discussions about new legal regimes distract from the fact that there’s no AI exemption from laws already on the books.”9
These perspectives still do not answer the question : what does pacing mean?
In January 2025 a rival matched the frontier in 41 days.10 Today the lead lasts months, & the gap between open & closed models has widened from three months to four.11 Should the interval be 60 days? 160? 600?
Can we really control the speed of innovation to that level of precision?
Amodei is specific about the lever. He would limit the ingredients that go into frontier models, training compute & the nature of training runs.1 Compute thresholds have been tried. A 2023 executive order required reporting above 10^26 FLOPS, & was revoked shortly thereafter.12 No model had crossed that line while the rule stood. The first, Grok-3, shipped weeks later, & roughly ten will clear it this year.13 Training compute grows about fivefold a year. A fixed number is a ceiling the floor reaches on its own.
The question is more than how fast. It is who decides.
-
Dario Amodei, We Must Pace the Frontier, September 2026. ↩︎ ↩︎ ↩︎
-
Evan Hubinger, @EvanHub, September 9 2026, as reported by CNBC. ↩︎
-
Bernie Sanders, Artificial intelligence is coming for the working class. We must fight back, April 16 2026. The employment figure is Sanders citing Stanford research on programming & customer service roles. ↩︎
-
Congressional Budget Office, The Budget and Economic Outlook: 2026 to 2036, February 2026. Debt held by the public passes its previous high of 106 percent, set in 1946, by 2030. ↩︎
-
CBO’s long-term sensitivity analysis, as summarized in Gale & Auerbach, The Federal Budget Outlook. A 0.5 percentage point rise in nonfarm-business total factor productivity leaves debt-to-GDP 43 percentage points lower by 2055. ↩︎
-
Scott Bessent, G20 press briefing, Asheville, North Carolina, September 2 2026, full transcript ; also reported by the Straits Times. ↩︎
-
Donald Trump to reporters at Doonbeg, Ireland, September 13 2026, via the Associated Press. ↩︎
-
David Sacks, @DavidSacks, September 13 2026. ↩︎
-
Lina Khan, @linamkhan, September 13 2026. ↩︎
-
Tomasz Tunguz, Darwin’s Finches in AI, January 2025. The interval measures how long a frontier model holds the lead before a rival matches it. ↩︎
-
Open models lag state-of-the-art closed models by 4 months, Epoch AI, May 29 2026. Epoch also finds no model since GPT-4 has led for a year ; the second-longest reign, o1’s, lasted barely three months. ↩︎
-
Executive Order 14110, October 30 2023, which set a 10^26 FLOPS reporting threshold under the Defense Production Act ; revoked January 20 2025. ↩︎
-
Epoch AI, How many AI models will exceed compute thresholds?, May 2025. Grok-3, released February 2025, was the first model estimated above 10^26 FLOP ; Epoch’s median projection is roughly ten by 2026 & thirty by the start of 2027. Frontier training compute has grown about 5x per year since 2020. ↩︎