The avalanche that hit math is coming for AI development

以下是根据您提供的英文文本翻译的简体中文摘要： 基于所提供的演讲记录，以下是包含演讲者提出的所有新闻、事实和观点的完整摘要： ### **核心论点与视角转变** * 演讲者在过去三个月里更新了他们对超级智能到来时间的看法（此前曾认为至少还要大半个十年）。 * 他们现在认为，距离达到人工智能研究领域顶尖人类水平的AI出现，仅剩几年时间——甚至可能只有几个月。 * 自动化AI研究将把如此之多的进展压缩到单单一年之内，以至于到2030年（完全有可能最早在明年），我们将拥有在各个领域至少不逊于最顶尖人类专家的AI。 * 对于自动化研究人员何时能达到最顶尖人类AI研究人员水平的估算中位数是 **2028年**（假设全速推进，没有受到人为或政府的故意减速），有25%的概率发生在 **2032年** 之后，并且有可能最早在明年发生（**2027年** 也被指出是模型规模化带来的直接推论）。 --- ### **AI加速AI进展的当前证据** * **Meter（独立评估机构）：** 在Anthropic内部驻留了三个月后得出结论：Anthropic在范围明确的任务上的AI进展速度是一年前的 **2倍**，其中“今天的大部分加速”可能归因于Anthropic广泛使用AI来进行研发（R&D）。 * **OpenAI（9月博客文章）：** 在OpenAI的研究人员中，AI使用量最大（前10%）的群体，其内部编码智能体的日均消耗从初夏时的约 **1,000美元** 飙升至8月底的超过 **7,000美元**。 * **Anthropic（9月报告）：** Claude目前主导了Anthropic **26%** 的AI研发工作，高于2月份的不到 **1%**。 --- ### **数学领域（可验证领域）能力的爆发** * 演讲者对比了数学进展： * **2023年：** 模型无法可靠地相加四位数。 * **2024年：** 模型几乎无法解决高中数学竞赛中的任何问题。 * **2025年：** 模型能够赢得国际数学奥林匹克竞赛金牌。 * **今年早些时候：** 模型终于开始解决一些小众的未解问题。 * **一个月前：** OpenAI在 **88小时** 内解决了一个千禧年大奖难题。 * **类比：** “神明的陨落有一点是确凿无疑的。他们不会缓缓坠落，而是会崩塌粉碎。” * 托比·奥特（Toby Ort）观察到，尽管当前的AI系统开始能够自动进行数学证明，但几乎没有证据表明它们能够提出有力的问题或发现深刻的新数学研究领域。然而，演讲者辩称，*在没有该能力的情况下*解决千禧年大奖难题，恰恰证明了范围明确的问题求解能力是多么强大。 * 在AI研究中，优先级与数学正好相反：AI实验室关心的是训练或启发更智能模型的经验效率（empirical efficiency），这使得随机的猜想穷举和低垂果实的发现（low-hanging fruit discovery）具有巨大的价值。 * **Transformer架构的实例：** 索尔托·道格拉斯（Sholto Douglas，AI研究员）指出，如果以不同的方式组合，Transformer架构（于 **2017年** 发表）本可以提前两年被发明出来。在Transformer之前，注意力机制和残差网络已经为人所知；将它们结合需要反复试验、犯错和坚持——而这正是当前AI擅长的任务。如果克里斯·奥拉（Chris Ola）推动此事，Claude 7本可以发明Transformer。 --- ### **一旦AI自动化AI研究，算法进展为何会加速** 1. **速度：** 今天的AI以大约人类 **10倍** 的速度工作，全天候运行（包括夜间和周末），每个日历周塞进的工作量大约是人类的 **50倍**。 2. **智能：** 每个自动化研究人员都将达到前沿实验室最顶尖人类的水平（并指出约80%的算法进展是由实验室前50名研究人员驱动的）。 3. **并行化 / 规模化：** AI研究具有独立的模块化工作和可验证的代理指标（verifiable proxies）。一群智能体可以同时在多个团队中尝试成千上万种不同的方法（预训练数据过滤、合成RL环境设计、注意力内核）。 --- ### **未来情境：纯软件奇点 vs. 火花** * **纯软件奇点（如果被逼问，有50%的概率）：** 递归自我改进（RSI）在不增加算力的前提下迅速制造出更智能的版本，直到超级智能在几个月内撞击到物理芯片的极限。 * **火花（如果被逼问，有50%的概率）：** 3到6年的AI进展被压缩到一个日历年中，随后出现模拟速度放缓，直到构建出更多的算力（例如，在12个月内实现从GPT-3.5到GPT-6 Astra这样巨大的跨越）。 * **对比：** GPT-3.5是一个只能写写勉强过得去的打油诗和微不足道Python函数的无用玩具；而Astra则能从头构建完整的电子视频游戏，并产出达到菲尔兹奖水平的数学贡献。 * 这种跨越如果从 **2027年** 或 **2028年** 开始，将在可验证领域创造出一个超乎人类想象的超级心智（例如，在设计下一代AI加速器方面击败英伟达，在收益上超越城堡投资（Citadel）和简街资本（Jane Street），或者在寻找室温超导体方面击败全人类）。 --- ### **现实世界与难以验证的领域** * 超级智能将如何掌握难以验证的领域（例如，建立企业、打赢官司、进行交易、赢得选举）？ * 扩大强化学习（RL）训练规模将提供通用技能（寻找信息、制定计划、克服错误、协同合作）。 * **虚拟世界 / RL环境：** 在RSI期间，自动化工程师将构建比以往多成千上万倍的RL环境，模拟庞大的经济体、跨越数个世纪的企业竞争以及历史回测（例如，仅使用截至 **1700年** 的数据来弄懂广义相对论，或者使用截至 **1962年** 的报纸来预测古巴导弹危机）。 * **人类平行对照：** 拿破仑在 **27岁** 时取得胜利；洛克菲勒在 **25岁** 时经营着世界领先的炼油厂。AI将在几天内累积多年的部署语境（例如，在其上岗的第一天早晨阅读一家财富500强公司的每一封电子邮件和Slack消息，并同时面试每一位员工）。 --- ### **算力增长与规模化** * 在过去的半个十年里，全球AI算力大约每年增长 **3倍**。 * 火花时刻之后的AI将通过设计AI芯片、规划晶圆厂和远程操作机器人，进一步帮助扩大算力规模。 --- ### **附录 / 技术细节与可验证的代理指标** * **样本效率与在线学习：** 机器学习中最深奥的未解问题（即AI版本的千禧年大奖难题）具有可测量的、可验证的代理指标：预训练损失数据需求、达到RL得分的任务，以及上下文学习（in-context learning）反馈速度。 * **AI研发的三个领域：** 1. **预训练：** 通过强化学习训练一个模型，使其最小化将一个GPT-2大小的模型训练到固定预训练损失所需的算力；留出测试集（held-out test sets）上的损失是很容易验证的。更小的模型训练速度更快，从而收紧RSI循环。Magic（AI实验室）预训练了一个模型，其性能匹配领先的开源模型，但耗费的算力仅为原始GPT-3的一半（今天价值 **约500,000美元**）。 2. **后处理训练（Post-training）：** 改进算法，以在RL环境中使用更少的算力/数据/时间；衡量模型学习未见过的视频游戏或任务的速度（例如字节跳动的 *EdgeBench*）；直接的内部基准测试，如Anthropic的 *CoBench*（研究人员解决的大约500个真实问题）。 3. **推理效率：** 设计针对GPU利用率、每个Token的KB缓存大小、总浮点运算次数（flops）和延迟的模型。 * **数据改进：** 数据是一个巨大的输入。与普林斯顿大学学生杰里·韩（Jerry Hahn）合作的研究发现，在过去的六年里，算法进展中有超过 **3倍** 的部分来自于数据改进，而非模型改进。DeepSeek的论文指出，合成数据在规模、多样性和可验证性上的系统性改进，几乎解释了所有观察到的RL收益。 * **实验算力可用性：** 前沿实验室每年将其算力翻两番。到 **2027年底**，OpenAI可以为约10,000个智能体提供足够的算力，让他们每天都能运行一次GPT-3规模的训练。最初的Transformer论文（2017年）中的模型在 **8张P100显卡上耗时3天** 训练完成（今天租用成本约为50美元，或仅为GPT-3算力的1/10,000）。 --- ### **解决瓶颈与反对意见** * **持续学习（Continuation Learning）：** 当前模型缺乏长期记忆的问题，可以通过扩大规模的上下文学习（持续数月的会话）来解决，或者作为一个带有可验证代理指标的ML研究问题来解决。（有25%的概率这会将ILLIA水平的顶级人类研究员出现的时间推迟到 **2032年** 之后）。 * **寻找可验证的代理指标 / 研究品味（Research Taste）：** 与范式无关的外循环代理指标（例如解决千禧年大奖难题、从1905年以前的数据中发现广义相对论、掌握未见过的视频游戏）绕过了这一瓶颈。 * **实验算力 / 节流（Throttling）：** 自动化agents可以用劳动力换取算力，运行小规模的消融实验（ablations）、多智能体RL运行以及基于缩放定律（scaling law）的预测（例如，使用算力减少 **10,000倍** 训练出来的模型来预测最终损失的GPT-4，或 *mu-p* 超参数预判）。 --- ### **鸣谢合作者** * 马克斯·费伦茨（Max Ferenc）、加布·弗里德（Gabe Freed）和托马斯·拉尔森（Thomas Larson）。

Based on the provided transcript, here is a complete summary including all the news, facts, and opinions presented by the speaker: ### **Main Thesis and Shift in Perspective** * The speaker has updated their views on the arrival of superintelligence over the previous three months (having previously thought it was at least a better part of a decade away). * They now believe we are only a couple of years away—and possibly only a couple of months away—from AIs that are as good as the best humans at AI research. * Automating AI research will compress so much progress into a single year that by 2030, and plausibly as early as next year, we will have AIs at least as good as the best human experts in every field. * The median year estimated for when automated researchers will be as good as the very best human AI researchers is **2028** (assuming full steam ahead without intentional or governmental slowdowns), with a 25% chance it happens after **2032**, and a possibility of it happening as early as next year (**2027** is also noted as a straightforward implication of scaling models). --- ### **Current Evidence of AI Accelerating AI Progress** * **Meter (independent evaluation organization):** Spent three months embedded inside of Anthropic and concluded that Anthropic's rate of AI progress on well-scoped tasks is **2x** what it was a year ago, with a "majority of this acceleration today" likely attributable to Anthropic's extensive use of AI to do R&D. * **OpenAI (September blog post):** The heaviest 10% of AI users amongst OpenAI’s researchers went from using about **$1,000 a day** of internal coding agents at the beginning of the summer to over **$7,000 a day** by the end of August. * **Anthropic (September report):** Claude now leads **26%** of Anthropic’s AI R&D work, up from under **1%** in February. --- ### **Explosion of Capabilities in Mathematics (Verifiable Domains)** * The speaker contrasts math progress: * **2023:** Models couldn't reliably add four-digit numbers. * **2024:** Models could barely solve any problems in a high school math competition. * **2025:** Models could win a gold medal in the International Math Olympiad. * **Earlier this year:** Models finally started solving some niche open problems. * **A month ago:** OpenAI solved a Millennium Prize problem in **88 hours**. * **Analogy:** "And there is one sure thing about the fall of gods. They do not fall a little. They crash and shatter." * Toby Ort's observation noted that while current AI systems are starting to automate mathematical proof, there is very little evidence they can ask powerful questions or find profound new fields of mathematical inquiry. However, the speaker argues that solving a Millennium Prize problem *without* that ability demonstrates how powerful well-scoped problem-solving is. * In AI research, priorities are reversed compared to mathematics: AI labs care about empirical efficiency in training or eliciting smarter models, making random conjecture-grinding and low-hanging fruit discovery enormously valuable. * **The Transformer Architecture Example:** Sholto Douglas (AI researcher) suggests the transformer architecture (published in **2017**) could have been invented two years earlier if combined differently. Before the transformer, attention and residual networks were known; combining them required trial, error, and persistence—tasks at which current AIs excel. Claude 7 could have invented the transformer if Chris Ola pushed for it. --- ### **Why Algorithmic Progress Will Accelerate Once AI Automates AI Research** 1. **Speed:** Today's AIs work at about **10 times** human speed, operating around the clock (nights and weekends), cramming roughly **50 times** more work every calendar week than a human. 2. **Intelligence:** Every automated researcher will be at the level of the very best humans at frontier labs (noting that ~80% of algorithmic progress is driven by a lab's top 50 researchers). 3. **Parallelization / Scale:** AI research features independent modular work and verifiable proxies. A swarm of agents can try out thousands of different approaches simultaneously across multiple teams (pre-training data filtering, synthetic RL environment design, attention kernels). --- ### **Future Scenarios: The Software-Only Singularity vs. The Spark** * **The Software-Only Singularity (50% chance if pressed):** RSI (Recursive Self-Improvement) rapidly makes smarter versions of itself without increasing compute until a superintelligence scrapes against physical chip limits within months. * **The Spark (50% chance if pressed):** 3 to 6 years of AI progress compressed into a single calendar year, followed by a simulation slowdown until more compute is built (e.g., a jump as big as GPT-3.5 to GPT-6 Astra in 12 months). * Comparison: GPT-3.5 was a useless toy writing passable limericks and trivial Python functions; Astra builds whole video games from scratch and produces field-medal-worthy math contributions. * A jump like this starting in **2027** or **2028** creates a mind vastly superhuman in verifiable domains (e.g., beating NVIDIA at designing next-gen AI accelerators, out-earning Citadel and Jane Street, or beating humanity at finding a room-temperature superconductor). --- ### **Real-World and Hard-to-Verify Domains** * How will superintelligence master hard-to-verify domains (e.g., building a business, winning court cases, trading, winning elections)? * Scaling up RL training will provide general skills (seeking information, making plans, pushing through mistakes, collaborating). * **Virtual Worlds / RL Environments:** During RSI, automated engineers will build tens of thousands of times more RL environments, simulating sprawling economies, corporate competitions spanning centuries, and historical backtests (e.g., figuring out general relativity using data only up to **1700**, or predicting the Cuban Missile Crisis using newspapers up to **1962**). * **Human Parallels:** Napoleon achieved victories by age **27**; Rockefeller ran a world-leading refinery by age **25**. AIs will accumulate years of deployed context in days (e.g., reading every email and Slack message of a Fortune 500 company on its first morning and interviewing every employee in parallel). --- ### **Compute Growth and Scaling** * Global AI compute has grown roughly **3x a year** for the last half-decade. * Post-Spark AIs will help scale compute further by designing AI chips, laying out fabs, and teleoperating robots. --- ### **Appendix / Technical Details & Verified Proxies** * **Sample Efficiency and Online Learning:** The deepest open problem in ML (the AI equivalent of a Millennium Prize problem) has measurable verifiable proxies: pre-training loss data requirements, tasks to reach RL scores, and in-context learning feedback speed. * **Three Buckets of AI R&D:** 1. **Pre-training:** RL a model to minimize compute used to train a GPT-2-sized model to a fixed pre-training loss; loss on held-out test sets is easily verifiable. Smaller models train faster, tightening the RSI loop. Magic (AI lab) pre-trained a model matching leading open-source models using half the compute of original GPT-3 (**~$500,000** today). 2. **Post-training:** Improving algorithms to use less compute/data/time for RL environments; measuring how quickly models learn unseen video games or tasks (e.g., ByteDance's *EdgeBench*); direct internal benchmarks like Anthropic's *CoBench* (~500 real problems solved by researchers). 3. **Inference Efficiency:** Designing models for GPU utilization, KB cache size per token, total flops, and latency. * **Data Improvements:** Data is a massive input. Research with Jerry Hahn (Princeton student) found that over the last six years, more than **3x** the algorithmic progress came from data improvements than model improvements. DeepSeek's paper notes that systematic improvements in scale, diversity, and verifiability of synthesized data account for essentially all observed RL gains. * **Experiment Compute Availability:** Frontier labs quadruple compute yearly. By the end of **2027**, OpenAI could provide ~10,000 agents with enough compute to run a GPT-3-sized training run every single day. The original Transformer paper (2017) model took **3 days on 8 P100s** (costing ~$50 to rent today, or 1/10,000th of GPT-3 compute). --- ### **Addressing Bottlenecks and Objections** * **Continuation Learning:** The lack of long-term memory in current models might be solved via scaled-up in-context learning (sessions lasting many months) or solved as an ML research problem with verifiable proxies. (25% chance this delays ILLIA-level top human researchers beyond **2032**). * **Finding Verifiable Proxies / Research Taste:** Paradigm-agnostic outer loop proxies (e.g., solving Millennium Prize problems, discovering general relativity from pre-1905 data, mastering unseen video games) bypass this bottleneck. * **Experiment Compute / Throttling:** Automated agents can trade labor for compute, running small-scale ablations, multi-agent RL runs, and scaling-law-driven predictions (e.g., GPT-4 predicting final loss using models trained with **10,000x less compute**, or *mu-p* hyperparameter anticipation). --- ### **Collaborators Acknowledged** * Max Ferenc, Gabe Freed, and Thomas Larson.