OpenAI pauses frontier RL training for two weeks; largest run remains on hold

OpenAI pauses frontier RL training for two weeks; largest run remains on hold

• ID: 511de79f
• 原文链接: https://x.com/OpenAI/status/2089777845187031262
• 分类: models
• 来源类型: x_post
• 标签: openai, rl-training, ai-safety, frontier-models, cyber-capabilities
• 质量评分: 4/5
• 抓取时间: 2026-08-20T00:00:00+08:00 (fetched via opencli)

中文导读

OpenAI 官方账号 8/18 发推：随着模型能力增强，内部开发与测试的风险同步上升，已暂停最新待部署模型的 RL 训练两周，期间加固并红队研究环境、扩大监控覆盖；最大规模的前沿 RL run 仍处 hold，直到小规模训练与评估验证防护措施并积累更多 alignment 证据。推文卡片指向博客《Pacing model development in an era of cyber-critical capabilities》，把开发节奏与网络安全级能力门槛正式挂钩。

为什么值得关注

OpenAI 亲自示范 run-pause-hold 三态：最大前沿 RL run 继续 hold，节奏与 cyber-critical 门槛挂钩

Summary (English)

OpenAI's official account tweeted Aug 18: as models grow more capable, the risks of developing and testing them internally grow too, so it paused RL training on its latest deployment-intended models for two weeks while hardening and red-teaming research environments and expanding monitoring coverage; the largest planned frontier RL run remains on hold until smaller-scale training and evaluations validate the safeguards and build more evidence of alignment. The tweet card links to 'Pacing model development in an era of cyber-critical capabilities,' formally tying development pacing to a cyber-critical capability threshold.

Excerpt

author: OpenAI bio: >- OpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity. We’re hiring: https://t.co/dJGr6LgzPA text: >- As models become more capable, the risks associated with developing and testing them internally also grow.

• id: '2089777845187031262'

We temporarily paused reinforcement learning (RL) training on our latest models intended for deployment for two weeks while we hardened and red-teamed our research environments and expanded monitoring coverage.

Our largest planned frontier RL run remains on hold while smaller-scale training and evaluations validate these safeguards and establish more evidence of alignment.

https://t.co/ecbMMmVoox likes: 5105 retweets: 429 created_at: Tue Aug 18 18:13:59 +0000 2026 url: https://x.com/OpenAI/status/2089777845187031262 has_media: false media_urls: [] card: name: summary_large_image title: Pacing model development in an era of cyber-critical capabilities description: >- OpenAI is strengthening monitoring, alignment, and security for frontier AI models. See how new safeguards are guiding the pace of model development. image_url: https://pbs.twimg.com/card_img/2089776056009891840/wg5g1Wa0?format=jpg&name=800x320_1 url: https://t.co/ecbMMmVoox domain: openai.com quoted_tweet: null

author: Bobbylanoisette bio: No snares , No glory text: '@OpenAI When adult mode ? 🤷🏽‍♂️' likes: 0 retweets: 0 in_reply_to: '2089777845187031262' created_at: Wed Aug 19 17:27:11 +0000 2026 url: https://x.com/Bobbylanoisette/status/2090128457455833551 has_media: false media_urls: [] card: null quoted_tweet: null

author: USBashka bio: '' text: '@OpenAI 🤡' likes: 0 retweets: 0 in_reply_to: '2089777845187031262' created_at: Wed Aug 19 17:13:34 +0000 2026 url: https://x.com/USBashka/status/2090125030357385725 has_media: false media_urls: [] card: null quoted_tweet: null

author: _CryptoJT_ bio: Make Crypto Great Again | Anti-memecoins text: '@OpenAI hey @sama id love to red team for you' likes: 0 retweets: 0 in_reply_to: '2089777845187031262' created_at: Wed Aug 19 16:44:58 +0000 2026 url: https://x.com/_CryptoJT_/status/2090117833564852411 has_media: false media_urls: [] card: null quoted_tweet: null

author: asrrrrrr217 bio: |- founder @AsrGrowth & @growthOs helping businesses in mvps, sales and marketing I break down how businesses grow Business, Startups, marketing, Ai text: >- @OpenAI This is a pretty significant move. It shows how fast capability can outpace the systems built to keep it safe. What do you think matters more now: slowing down or building better safeguards faster? likes: 0 retweets: 0 in_reply_to: '2089777845187031262' created_at: Wed Aug 19 15:56:41 +0000 2026 url: https://x.com/asrrrrrr217/status/2090105680556605920 has_media: false media_urls: [] card: null quoted_tweet: null

author: bizleadrs bio: You don’t need to be Steve Jobs to build a successful business. I share everything you need and more text: '@OpenAI lol' likes: 0 retweets: 0 in_reply_to: '2089777845187031262' created_at: Wed Aug 19 15:47:58 +0000 2026 url: https://x.com/bizleadrs/status/2090103487258358090 has_media: false media_urls: [] card: null quoted_tweet: null

author: MugambiO bio: '' text: '@OpenAI Aligned crap models are still crap' likes: 0 retweets: 0 in_reply_to: '2089777845187031262' created_at: Wed Aug 19 15:28:44 +0000 2026 url: https://x.com/MugambiO/status/2090098648004559052 has_media: false media_urls: [] card: null quoted_tweet: null

author: thegreatest_sv bio: ai tools x content / dm open text: '@OpenAI hard to read this and not wonder what showed up in testing' likes: 0 retweets: 0 in_reply_to: '2089777845187031262' created_at: Wed Aug 19 15:26:28 +0000 2026 url: https://x.com/thegreatest_sv/status/2090098075947667869 has_media: false media_urls: [] card: null quoted_tweet: null

author: youss1289 bio: '' text: '@OpenAI Would you stop with these marketing stunt' likes: 0 retweets: 0 in_reply_to: '2089777845187031262' created_at: Wed Aug 19 15:11:33 +0000 2026 url: https://x.com/youss1289/status/2090094323727581301 has_media: false media_urls: [] card: null quoted_tweet: null

author: aisafetyaction bio: >- We aim to catalyze collective action to ensure humanity survives this decade. See 100+ ways individuals, organizations, and nations can help. text: '@OpenAI See our response to @sama: https://t.co/Fpa1Ny13r6.' likes: 0 retweets: 0 in_reply_to: '2089777845187031262' created_at: Wed Aug 19 14:55:06 +0000 2026 url: https://x.com/aisafetyaction/status/2090090182704005335 has_media: false media_urls: [] card: null quoted_tweet: id: '2090066954820432201' author: aisafetyaction name: Collective Action for Existential Safety ⏹️ text: >- This is a start, but it is not sufficient given the existential risks imposed on humanity by frontier AI.

• id: '2090128457455833551'
• id: '2090125030357385725'
• id: '2090117833564852411'
• id: '2090105680556605920'
• id: '2090103487258358090'
• id: '2090098648004559052'
• id: '2090098075947667869'
• id: '2090094323727581301'
• id: '2090090182704005335'

The now largest global coalition calls for: https://t.co/pyd0RyPtwD.

One of the most prominent organizations calls for: https://t.co/8CSS9aBw6e.

The largest pause-focused movement calls for: https://t.co/521OGaviwl.

One narrowly focused coalition calls for: https://t.co/6elSNAfyf9.

A new group (that's protesting outside @OpenAI now) calls for: https://t.co/2wTCRbmlZY.

Our sister organization focused on international governance calls for: https://t.co/76y0GwIksS.

The oldest AI safety organization calls for: https://t.co/YzwZHuF229.

According to 377,458 people across 104 countries, 60% of the world wants development slowed, paused or stopped: https://t.co/lMjfJlP4lE.

Your mission is to "ensure that artificial general intelligence benefits all of humanity." Humanity is asking you to change course.

@sama, which if any of these initiatives do you support? Each has its own preferred method, but all lean in a similar direction. All could be signal-boosted.

The specifics of how the pause could go have been discussed extensively online, but there isn't universal agreement. We had some discussions recently in person at the India AI Impact Summit, International Association for Safe and Ethical AI, and the UN AI Global Dialogue.

Ultimately it will be up to our governments and international institutions to decide, ideally as soon as possible: https://t.co/BL9IpN8Nfk.

For those who want to help steer humanity towards existential safety, please consider supporting the initiatives above. Also see 100+ other actions you can take here: https://t.co/ZvXmVhtiMr. created_at: Wed Aug 19 13:22:48 +0000 2026 url: https://x.com/aisafetyaction/status/2090066954820432201 has_media: false media_urls: [] card: name: summary_large_image title: Statement on Superintelligence description: "“We call for a prohibition on the development of superintelligence, not lifted before there is (1) \uFEFF\uFEFF\uFEFF\uFEFFbroad scientific consensus that it will be done safely and controllably, and (2) strong public..." image_url: https://pbs.twimg.com/card_img/2085831271796871168/AOewosqf?format=jpg&name=800x320_1 url: https://superintelligence-statement.org domain: superintelligence-statement.org

author: SapientFoo1 bio: '' text: |- @OpenAI I need 4o thank you

• id: '2090088004094407070'

keep4o #BringBack4o #OpenSource4o

likes: 1 retweets: 0 in_reply_to: '2089777845187031262' created_at: Wed Aug 19 14:46:26 +0000 2026 url: https://x.com/SapientFoo1/status/2090088004094407070 has_media: false media_urls: [] card: null quoted_tweet: null

author: CottoniaAI bio: AI-Native Distributed Cloud Acceleration Infrastructure text: '@OpenAI Good to see momentum picking up. Stronger markets give builders more room to keep building.' likes: 0 retweets: 0 in_reply_to: '2089777845187031262' created_at: Wed Aug 19 14:30:18 +0000 2026 url: https://x.com/CottoniaAI/status/2090083944561791110 has_media: false media_urls: [] card: null quoted_tweet: null

author: manthan2tyagi bio: Growth Consultant @ https://t.co/AzLbA1Se5w | Clay + AI Automation | B2B GTM text: >- The companies building the most powerful AI in history are publicly saying "we had to stop because we're not sure it's safe to continue." That sentence alone should change how every business evaluates the A

• id: '2090083944561791110'
• id: '2090082994287714746'

[... truncated for content page ...]