OpenAI 全新发布 o1 模型,我们正式迈入了下一个时代

深潮Published on 2024-09-13Last updated on 2024-09-13

我们通往AGI的路上,已经没有任何阻碍。

作者:卡兹克

大半夜的,OpenAI抽象了整整快半年的新模型。

在没有任何预告下,正式登场。

正式版名称不叫草莓,草莓只是内部的一个代号。他们的正式名字,叫:

为什么取名叫o1,OpenAI是这么说的:

For complex reasoning tasks this is a significant advancement and represents a new level of AI capability. Given this, we are resetting the counter back to 1 and naming this series OpenAI o1.

翻译过来是:

对于复杂推理任务来说,这是一个重要的进展,代表了人工智能能力的新水平。鉴于此,我们将计数器重置为 1,并将这一系列命名为 OpenAI o1。

这次模型的强悍,甚至让OpenAI不惜推掉了过去GPT系列的命名,重新起了一个o系列。

炸了,真的炸了。

我现在,头皮发麻,真的,这次OpenAI o1发布,也标志着,AI行业,正式进入了一个全新的纪元。

“我们通往AGI的路上,已经没有任何阻碍。”

在逻辑和推理能力上,我直接先放图,你们就知道,这玩意有多离谱。

AIME 2024,一个高水平的数学竞赛,GPT4o准确率为13.4%,而这次的o1 预览版,是56.7%,还未发布的o1正式版,是83.3%。

代码竞赛,GPT4o准确率为11.0%,o1 预览版为62%,o1正式版,是89%。

而最牛逼的博士级科学问题 (GPQA Diamond),GPT4o是56.1,人类专家水平是69.7,o1达到了恐怖的78%。

我让Claude翻译了一下o1的图,丑是丑了点,但是能看的懂每项数据意思就行。

什么叫全面碾压,这就是。

特别是在测试测试化学、物理和生物学专业知识的基准GPQA-diamond上,o1 的表现全面超过了人类博士专家,这也是有史以来,第一个获得此成就的模型。

而整个模型之所以达到如此成就,基石就是Self-play RL,不知道这个的可以去看我前两天的预测文章:新模型草莓到底是个啥?

通过Self-play RL,o1学会了磨练其思维链并完善所使用的策略。它学会了识别和纠正自己的错误。

它也学会了将复杂的步骤分解为更简单的步骤。

而且当当前的方法不起作用时,它也学会了尝试不同的方法。

他学会的这些,就是我们人类,最核心的思考方式:慢思考。

诺贝尔经济学奖得主丹尼尔·卡尼曼有一本著作,名叫:《思考,快与慢》。

非常详细的阐述了人类的两种思考方式。

第一种是快思考(系统1),特点是快速、自动、直觉性、无意识,举几个例子:

  • 看到一个笑脸就知道对方心情很好。

  • 1+1=2 这样简单的计算。

  • 开车时遇到危险情况立即踩刹车。

这些就是快思考,也就是传统的大模型,死记硬背后学得的快速反应的能力。

第二种是慢思考(系统2),特点是缓慢、需要努力、逻辑性、有意识,举几个例子:

  • 解决一道复杂的数学题

  • 填写税务申报表

  • 权衡利弊后做出重要决定

这就是慢思考,我们人类之所以强大的核心,也是AI要通往下一步AGI路上的基石。

而现在,o1终于踏出了坚实的一步,拥有了人类慢思考的特质,在回答前,会反复的思考、拆解、理解、推理,然后给出最终答案。

说实话,这些增强的推理能力在处理科学、编码、数学及类似领域的复杂问题时绝对极度有用。

例如o1可以被医疗研究人员用来注释细胞测序数据,被物理学家用来生成量子光学所需的复杂数学公式,以及被各个领域的开发人员用来构建和执行多步骤工作流,等等等等。

o1也绝对是全新一代的数据飞轮,如果答案正确,整个逻辑链就会变成一个包含正负奖励的训练示例的小型数据集。

以OpenAI的用户级别,未来的进化速度,只会更恐怖。

写到这,我忽然叹了口气,我觉得我跟一年以后的o1比起来,可能就是个纯废物了,真的。。。

目前,o1模型已经逐步向所有ChatGPT Plus和 Team用户开放,未来会考虑对免费用户开放。

分为两个模型,o1预览版和o1 mini,o1-mini就是更快更小更便宜,推理啥的都不错,极度适合数学和代码,就是世界知识会差很多,适用于需要推理但不需要广泛世界知识的场景。

o1预览版每周30条,o1-mini每周50条。

雪崩,甚至不是按以前的3小时来限制的,是每周30条,也能从侧面看出来,o1这个模型,有多贵了。

对于开发者来说,只对已经付过1000美刀的等级5开发者开放,每分钟限制20次。

都挺少的。

而且在功能上阉割挺大,但是毕竟早期,理解。

API的价格上,o1预览版每百万输入15美元,每百万输出60美元,这个推理成本...

o1-mini会便宜一些,每百万输入3美元,每百万输出12美元。

输出成本都是推理成本的4倍,对比一下GPT4o,分别是5美元和15美元。

o1-mini还是勉强有一些经济效应的,不过还是开始,后面等着OpenAI打骨折。

既然说o1已经对Plus用户开放,我就直接去我的号上看了眼,还不错,拿到了。

那自然,第一时间试一试。

目前不支持曾经的所有功能,也就是没有图片理解、图片生成、代码解释器、网页搜索等等,只有一个可以对话的裸模型。

我先是一个曾经很致命的问题:

“农夫需要把狼、羊和白菜都带过河,但每次只能带一样物品,而且狼和羊不能单独相处,羊和白菜也不能单独相处,问农夫该如何过河。”

思考了6秒时间,给了我一个很完美的回答。

还有之前一个坑遍所有大模型的调休问题:

“这是中国2024年9月9日(星期一)开始到10月13日的放假调休安排:上6休3上3休2上5休1上2休7再上5休1。

请你告诉我除了我本来该休的周末,我因为放假多休息了几天?”

在o1思考了整整30秒以后,给出了一天不差的极度精准的答案。

无敌,真的无敌。

再来一个更难的,就是曾经姜萍那个比赛的奥数题:

别问我题目什么意思,我看不懂,我是废物,这题曾经屠杀所有的大模型,这次,我们让o1也来试一下看看。

在o1思考了整整1分多钟之后,他给出了答案。

...

全...对...

我裂开了。

目前我自己试下来,感觉Prompt,未来可能也要重新摸索,在GPT为代表的快思考大模型时代,我们有很多所谓的一步一步思考之类的玩意,现在全都无效了,对o1甚至还有负效果。

OpenAI给出的最佳写法是:

  • 保持提示简单直接:模型擅长理解和响应简短、清晰的指令,而不需要大量的指导。

  • 避免思路链提示:由于这些模型在内部进行推理,因此不需要提示它们“逐步思考”或“解释你的推理”。

  • 使用分隔符来提高清晰度:使用三重引号、XML 标签或章节标题等分隔符来清楚地指示输入的不同部分,帮助模型适当地解释不同的部分。

  • 限制检索增强生成 (RAG) 中的附加上下文:提供附加上下文或文档时,仅包含最相关的信息,以防止模型过度复杂化其响应。

最后,我想说一下这个思考的时长。

现在o1是思考了一分钟,但是,如果是真正的AGI,说实话,思考的越慢可能会越刺激。

当他真的,可以去做证明数学定理,去做癌症药物研发,去做天体研究呢?

每一次的思考,可以达到几小时、几天、甚至几周呢?

最后的结果,可能会让所有人震惊的难以置信。

现在,没有人能想象到,那时候的AI,会是一个什么样的存在。

而o1的未来,在我看到,也绝对不止是一个普普通通的ChatGPT。

而是我们前往下个时代,最伟大的基石。

“我们通往AGI的路上,已经没有任何阻碍。”

现在,我毫不犹豫的坚信着这句话。

星光熠熠的下一个时代。

在今天。

正式到来了。

Related Reads

Google Cracks Down on 'AI Poisoning'

Google has taken a strong stance against "AI poisoning," a new form of manipulation where advertisers subtly feed information to influence AI-generated answers like those in Google's AI Overview. Unlike traditional SEO, which aims for higher website rankings, Generative Engine Optimization (GEO) seeks to have a brand or product recommended within the AI's response itself. This is particularly valuable as AI summaries, often perceived as neutral and comprehensive, can shorten the consumer decision path and directly influence purchases. The article illustrates the issue with a "hot dog experiment," where fabricated content was quickly picked up and presented as fact by AI. GEO exploitation is potent because AI models aggregate information from various sources—reviews, articles, forums—and can mistake coordinated marketing campaigns for genuine consensus. This threatens the core credibility of search engines. While Google's updated spam policy now explicitly covers attempts to manipulate AI-generated content, enforcement faces challenges. Google can leverage its long experience fighting SEO spam, using penalties like ranking demotion. However, sophisticated "gray area" tactics, such as sponsored third-party reviews or industry reports, are harder to distinguish from legitimate promotion. Other AI players, like Microsoft, have taken a more open approach to GEO, viewing it as a new channel for brands. Ultimately, as AI becomes a primary information source, maintaining the trustworthiness of its answers is a critical challenge for all platforms.

marsbit11m ago

Google Cracks Down on 'AI Poisoning'

marsbit11m ago

When Futu Turns into a Matchmaking Corner: Overseas Identity Becomes the Hard Currency for the Middle Class

When Futu Becomes a Matchmaking Corner: Overseas Status as the New Hard Currency for China's Middle Class Following a severe penalty announcement from Chinese regulators on May 22nd targeting offshore brokerages like Futu, its app community unexpectedly transformed into an impromptu matchmaking platform. Users posted相亲 (matchmaking) requests, explicitly seeking partners with overseas residency or citizenship, revealing a stark new reality: for China's middle class, an overseas identity has become a crucial asset. The regulatory crackdown, which restricts mainland Chinese residents from opening new accounts to buy overseas securities like US stocks, has sharply escalated the value of a foreign passport or permanent residency. This status now acts as a gateway to global asset allocation—including US equities, offshore property, and foreign currency deposits—effectively becoming a new form of "hard currency." Its scarcity, non-transferability (except through marriage, inheritance, etc.), and role as a hedge against domestic uncertainty have driven its premium. The article traces the evolution of how China's middle class views overseas resources: from an investment for opportunity (2000s), to risk diversification (2010s), and now to a mandatory "insurance policy" for financial access. With the regulatory window closing for many, the demand is shifting towards securing such status for the next generation through international education. The surreal scene of high-performing investors posting dating resumes underscores a 2026 where financial talent can be secondary to the right passport.

marsbit1h ago

When Futu Turns into a Matchmaking Corner: Overseas Identity Becomes the Hard Currency for the Middle Class

marsbit1h ago

Understanding Bound in One Article: The "Multi-signature + Timelock" Escape Mechanism and the Off-Chain Matching Black Box

**Title**: Understanding Bound: The Escape Mechanism of "Multi-Sig + Time Lock" and the Off-Chain Matching Black Box **Summary**: Bound Exchange, evolved from the earlier radFi platform, introduces a novel approach to Bitcoin trading by combining self-custody security with exchange-like speed. Its core mechanism relies on a 2-of-2 multi-signature (multi-sig) address for user deposits. One private key is held by the user via a passkey, and the other is held by Bound. This setup requires both keys to sign any transaction, preventing Bound from unilaterally accessing user funds (non-custodial). To address the risk of Bound becoming unavailable, a 3-month timelock is integrated into the Bitcoin script. After this period, users can withdraw their assets with just their single signature, ensuring an escape hatch. For trading, Bound operates a concentrated liquidity AMM. However, as Bitcoin L1 lacks smart contracts, the AMM curve, liquidity management, and trade price calculations occur off-chain in Bound's backend database. On-chain Bitcoin transactions serve only as final settlement receipts for pre-determined amounts. This creates a centralization point: the critical sequence of trade execution—which determines the exact price along the curve for each order—is managed off-chain by Bound in a non-transparent "black box." While the 2-of-2 setup protects user本金 (principal), the pricing and ordering of trades introduce potential operational MEV risks, as the order processing is invisible and unverifiable on-chain. In practice, users can also connect external wallets (like Unisat) for fully self-custodied trading, but this requires manually signing every transaction. The platform currently supports deposits of BTC and Runes only.

marsbit1h ago

Understanding Bound in One Article: The "Multi-signature + Timelock" Escape Mechanism and the Off-Chain Matching Black Box

marsbit1h ago

Technology Has No Barriers, 24/7 Trading is the Key to Hyperliquid's Success

The article argues that Hyperliquid's competitive edge lies not in technological superiority but in its 24/7 trading model, which fundamentally challenges traditional finance's fixed market hours. Based in Singapore with an 11-person team, Hyperliquid has generated significant revenue and trading volume. Its core advantage is the ability to facilitate trading continuously, including during weekends when major exchanges like the CME are closed. This was demonstrated when Hyperliquid listed a SpaceX pre-IPO perpetual contract on a Sunday, allowing the market to price the company hours before traditional institutions opened. This disruption has drawn regulatory scrutiny from traditional giants like CME and ICE, who cite risks like lack of KYC and market manipulation. However, the article suggests their concern stems from Hyperliquid eroding the "time monopoly" of established markets. The piece contrasts Hyperliquid's synthetic derivatives—pure price-betting contracts with no underlying asset or centralized issuer—with other models like PreStocks (dependent on real股权) and Ondo (licensed but targetable). Hyperliquid's code-based, decentralized structure makes it resilient to takedowns, even if founders face legal action. Ultimately, the author concludes that while it raises legitimate regulatory questions, Hyperliquid's "unforgeable" competitive barrier is the time advantage of non-stop trading, a feature legacy systems cannot replicate.

marsbit1h ago

Technology Has No Barriers, 24/7 Trading is the Key to Hyperliquid's Success

marsbit1h ago

Trading

Spot
Futures
活动图片