▶ 原文链接

2026年AI现状:大语言模型、编码、缩放定律、中国、智能体、GPU、通用人工智能

来源: Lex Fridman Podcast | 嘉宾:Nathan Lambert、Sebastian Raschka | 日期:无 分类: AI 资讯 原文发表: Jan 31, 2026 纪要生成: 2026-02-25


全集重点


嘉宾/话题简介

本次对话邀请了两位AI领域兼具技术实力与科普能力的权威人士:Sebastian Raschka是畅销书《从零构建大语言模型》《从零构建推理模型》作者,深耕大模型架构与AI教育;Nathan Lambert是艾伦人工智能研究所(AI2)后训练负责人,RLHF领域权威,著有该领域的权威专著。本次对话围绕2025-2026年AI领域的技术突破、行业竞争、未来趋势展开,覆盖技术、商业、政策、伦理多个维度,是当前AI行业现状的全景式梳理。


分节详述

0:00 引言

本节重点

详细精要

💬 精华片段(中文) "在机器学习和计算机科学领域,理解某件事的最佳方式就是亲手从零开始构建它。"

"I truly believe in the machine learning and computer science world, the best way to learn and understand something is to build it yourself from scratch."


1:57 中美AI竞赛:谁会胜出?

本节重点

详细精要

💬 精华片段(中文) "我认为现在2026年不会有任何一家公司拥有其他公司完全无法获得的技术,核心原因是研究人员会频繁换工作、换实验室,人员是流动的。"

"One thing I know for sure is that I don't think nowadays, in 2026, that there will be any company having access to a technology that no other company has access to. And that is mainly because researchers are frequently changing jobs, changing labs. They rotate."


10:38 ChatGPT vs Claude vs Gemini vs Grok:谁处于领先?

本节重点

详细精要

💬 精华片段(中文) "你会一直用它直到它出错,遇到问题后再换其他大模型,这和我们使用文本编辑器、操作系统、浏览器的逻辑完全一样。"

"This is exactly it. You use it until it breaks, until you have a problem, and then you change the LLM. I think it's the same way we use anything, like our favorite text editor, operating system, or browser."


21:38 最佳AI编码工具

本节重点

详细精要

💬 精华片段(中文) "代码不会说谎,它本质上就是数学。就算是数学,你在书中看到的公式也可能有错误,你阅读时不会运行它所以很难发现,但代码你一运行就知道是否正确。"

"I think that's the beauty behind coding. It doesn't lie. It's math, basically. Even with math, you can have mistakes in a book you would never notice because you aren't running the math while reading, so you can't verify it. And with code, what's nice is you can verify it."


28:29 开源 vs 闭源大模型

本节重点

详细精要

💬 精华片段(中文) "解决幻觉的最佳方式之一不是让模型记住所有信息,而是让它学会调用工具:数学问题用计算器,事实查询用搜索引擎,这样就能大幅提升结果的可靠性。"

"One of the most common complaints about LLMs is, for example, hallucinations, right? And so, in my opinion, one of the best ways to solve hallucinations is to not try to always remember information or make things up. For math, why not use a calculator app or Python?"


40:08 Transformer:2019年以来大模型的演进

本节重点

详细精要

💬 精华片段(中文) "当前没有任何架构能够替代自回归Transformer成为SOTA模型的首选,虽然已有文本扩散模型、Mamba等替代架构出现,但仅适用于特定的低成本场景。"

"But what's true is there's nothing that has replaced the autoregressive transformer as the state-of-the-art model. So, for state-of-the-art, you would still go with that thing, but there are now alternatives for the cheaper end—alternatives that are kind of making compromises, but it's not just one architecture anymore."


48:05 AI缩放定律:已经失效还是仍然有效?

本节重点

详细精要

💬 精华片段(中文) "缩放定律已经在13个数量级的算力提升下保持有效,为什么会突然停止呢?从根本上来说它失效的可能性极低,只是随着规模增大,测试更大尺度的缩放会变得越来越难。"

"And this sometimes comes off as almost disillusioned from leadership at AI companies saying this, but they're like, 'It's held for 13 orders of magnitude of compute; why would it ever end?' So I think fundamentally it is pretty unlikely to stop. It's just like eventually we're not even going to be able to test the bigger scales because of all the problems that come with more compute."


1:04:12 AI如何训练:预训练、中训练、后训练

本节重点

详细精要

💬 精华片段(中文) "如果你想加入前沿AI实验室并产生影响力,最佳路径不是去研究高大上的算法,而是找到更好的训练数据,或者优化基础设施让整个团队的实验速度提升5%。"

"The fancy, glamorous algorithmic things, like figuring out how to make o1, is like the sexiest thought for a scientist. It's like, 'Oh, I figured out how to scale RL.' There's a group that did that, but I think most of the contributions are- 'I’m gonna make the data better,' or, 'I’m gonna make the infrastructure better so that everybody on my team can run experiments 5% faster.'"


1:37:18 后训练详解:大模型的热门新研究方向

本节重点

详细精要

💬 精华片段(中文) "RLVR不会教模型新的数学知识,它的核心作用是解锁模型在预训练阶段已经学到的知识,让模型学会如何用正确的方式调用这些知识解决问题。"

"Exactly. And so you can see that basically the RL is not teaching the model any new knowledge about math. You can't do that in 50 steps. So the knowledge is already there in the pre-training; you're just unlocking it."


1:58:11 给AI开发与研究入门者的建议

本节重点

详细精要

💬 精华片段(中文) "你不用试图学习所有领域的知识,那样会非常容易 burnout,聚焦在大模型这一个领域深入研究就足够了。"

"Yeah, I think you can't try to do it all because it would be very overwhelming and you would burn out. For example, I haven't kept up with computer vision in a long time; I've just focused on LLMs."


2:21:03 AI行业的工作文化(每周72小时以上工作)

本节重点

详细精要

💬 精华片段(中文) "这是一个以人力损耗为代价推动技术进步的完美环境,人们真的在玩命工作。"

"It's a perfect environment for creating progress based on human expense. The human expense is the 996 that we started this with, where people do really grind."


2:24:49 硅谷泡沫

本节重点

详细精要

💬 精华片段(中文) "我觉得旧金山是一个不可思议的地方,但确实存在一点泡沫。如果你进入了这个泡沫,它确实能带来极高的生产力,但也要记得走出来,读历史书,去世界其他地方看看,Twitter和Substack不是整个世界。"

"I think SF is an incredible place, but there is a bit of a bubble. And if you go into that bubble, which is extremely valuable, just get out also. Read history books, read literature, and visit other places in the world. Twitter and Substack are not the entire world."


2:28:46 文本扩散模型与其他新研究方向

本节重点

详细精要

💬 精华片段(中文) "文本扩散模型不会替代自回归大模型,但会成为快速、低成本、大规模场景的首选,未来的免费 tier 很可能会采用这类模型。"

"I don't think the text diffusion model is going to replace autoregressive LLMs, but it will be something for quick, cheap, at-scale tasks. Maybe the free tier in the future will be something like that."


2:34:28 工具使用

本节重点

详细精要

💬 精华片段(中文) "工具调用不能完全解决幻觉问题,但可以大幅降低幻觉。大模型仍然需要知道什么时候调用工具,以及如何正确判断工具返回结果的准确性。"

"Not solve it, but reduce it. Still, the LLM needs to know when to ask for a tool call. And second, it doesn't mean the internet is always correct. You can do a web search for who won the World Cup in 1998, but it still needs to find the right website and get the right information."


2:38:44 连续学习

本节重点

详细精要

💬 精华片段(中文) "我们其实已经有了不同形式的连续学习:从GPT-5到5.1再到5.2的版本迭代,就是一种全局层面的连续学习,吸收社区的反馈优化模型能力。"

"I think, to be honest with you, continual learning—the updating of weights—we already have that in different flavors. I think the distinction here is: do you do that on a personalized custom model for each person, or do you do it on a global model scale? And I think we have that already with going from GPT-5 to 5.1 and 5.2."


2:44:06 长上下文

本节重点

详细精要

💬 精华片段(中文) "当前状态下,想达到SOTA性能还是需要 brute force 的全注意力机制,保证不会遗漏任何信息。2026年的核心优化方向是在保持准确率的前提下,通过更智能的上下文管理降低成本。"

"Occasionally, in some layers you might, but it's wasteful. But right now, I think if you use everything, you're on the safe side; it gives you the best bang for the buck because you never miss information. And right now, I think this year will also be the year of figuring out, like you said, how to be smarter about that."


2:50:21 机器人

本节重点

详细精要

💬 精华片段(中文) "在大模型领域,出错只是输出错误文本,是好玩的游戏,但在机器人领域,在千家万户的真实场景中,数十亿次交互下几乎不允许出错,这是机器人落地的核心挑战。"

"All the interesting complexities we talk about regarding learning, all the failure modes and failure cases—everything we've been talking about with LLMs where sometimes it fails in interesting ways—all of that is fun and games in the LLM space. In the robotic space, in people's homes, across millions of minutes and billions of interactions, you really are almost allowed to fail never."


2:59:31 AGI的时间线

本节重点

详细精要

💬 精华片段(中文) "我认为AGI和ASI的阈值没有特别大的实用价值,更值得关注的是AI什么时候会带来明显的经济影响,当前LLM尚未带来显著的GDP跃升,这才是更值得讨论的实际问题。"

"I think the real question, and this relates to the remote worker thing, is when are we going to see a big, obvious leap in economic impact? Because currently there's not been an obvious leap in economic impact from LLM models, for example. Aside from AGI or ASI, there's a real question of when we are going to see a GDP jump. Jump."


3:06:47 AI会替代程序员吗?

本节重点

详细精要

💬 精华片段(中文) "软件工程将更多转向系统设计与目标定义,软件很大程度上会被自动化生成,越来越多的人不需要看代码就能创建软件,只需要理解系统如何工作,能从大模型中提取最佳结果即可。"

"I think software engineering will be driven more to system design and goals of outcomes, where I do think software is largely going to be… I think this has been happening over the last few weeks, where people have gone from a month ago saying, 'Oh yeah, agents are kind of slop,' which is a famous Karpathy quote, to the industrialization of software when anyone can just create software with their fingerprints."


3:25:18 AGI的梦想正在消亡吗?

本节重点

详细精要

💬 精华片段(中文) "我们其实忽略了一个非常明显的巨大价值:大模型让所有人类知识对全世界所有人都变得可及。你可以问大模型任何问题,获得准确的答案,这对整个人类文明的影响是难以估量的。"

"I think we're not saying one actually obvious thing that we're not realizing, that's a gigantic thing that's hard to measure, which is making all of human knowledge accessible… …To the entire world. One of the things that I think is hard to articulate, but there's just a huge difference between Google Search and an LLM. I feel like I can basically ask an LLM anything and get an answer, and it's doing less and less hallucination."


3:32:07 AI如何盈利?

本节重点

详细精要

💬 精华片段(中文) "当前大模型的服务之所以这么便宜,是因为厂商在大规模补贴,未来广告模式上线后,可能会出现免费的大模型服务,但会植入标注明确的广告。"

"Well, for now, that's because they're massively subsidized, and eventually they're going to be paid for by ads."


3:36:29 2026年的大型收购

本节重点

详细精要

💬 精华片段(中文) "创业生态是硅谷的命脉,如果你加入一家创业公司,即使它不算特别成功,也很可能被收购,你的股权会得到回报。而现在的授权协议模式本质上是规避反垄断监管,通常只会带走核心人才,普通员工无法受益,这是硅谷文化需要解决的大问题。"

"There are countless other deals structured in a way that is actually detrimental to the Silicon Valley ecosystem—these licensing deals where not everybody gets brought along, rather than a full acquisition that benefits the rank-and-file employees by getting their stock vested. That's a big issue for Silicon Valley culture to address because the startup ecosystem is the lifeblood. If you join a startup, even if it's not that successful, your startup very well might get acquired at a cheap premium and you'll get paid out for your equity."


3:41:01 OpenAI、Anthropic、Google DeepMind、xAI、Meta的未来

本节重点

详细精要

💬 精华片段(中文) "如果没有更多对开源模型的投资,我们看到的排行榜上就会全是Qwen等中国公司的优秀模型,它们正在美国和全球积累影响力。美国在AI上的投入要大得多,打造领先于闭源实验室半代到一代的开源模型仅需要约1亿美元,和这些公司的投入相比并不算多。"

"Without more investment in open models, we have all the plots on the website where it's like, 'Qwen, Qwen, Qwen, Qwen,' and it's all these models that are excellent from these Chinese companies that are cultivating influence in the US and internationally. And the US is spending way more on AI. The ability to create open models that are half a generation or a generation beyond what the cutting edge of closed labs is costs roughly $100 million, which is a lot of money, but not compared to what these companies have."


3:53:35 AI曼哈顿计划

本节重点

详细精要

💬 精华片段(中文) "开源对教育和人才培养至关重要,如果只有闭源模型,下一代人只有加入公司才能接触到核心技术,我们无法识别和培养有天赋的人才,这是唯一的路径。"

"Also, for education and talent, it's very important. Otherwise, if there are only closed models, how do you get the next generation of people contributing? You would only be able to learn after you joined a company, but at that point, how do you identify and hire talented people? I think open source is essential for educating the population and training the next generation of researchers. It's the only way."


4:00:10 NVIDIA、GPU与AI计算集群的未来

本节重点

详细精要

💬 精华片段(中文) "只要AI的进步速度仍然很高,NVIDIA的平台就是最灵活的,人们会愿意选择它。如果出现停滞,那么就有更多时间来开发定制化芯片,才会对NVIDIA构成威胁。"

"As long as the pace of AI progress is high, NVIDIA's platform is the most flexible and people will want that. But if there's stagnation, then with creating bespoke chips, there's more time to do it."


4:08:15 人类文明的未来

本节重点

详细精要

💬 精华片段(中文) "我认为人类完全有能力应对这些挑战,人类的本质就是建立社区、找到解决问题的方法,这是我们走到今天的核心原因。AI的机遇非常大,虽然面临很多社会与政治问题需要解决,但我相信我们最终可以实现AI的长久收益。"

"I think we will. I'm definitely a worrier both about AI and non-AI things, but humans do tend to find a way. I think that's what humans are built for—to have community and find a way to figure out problems. And that's what has gotten us to this point. I think the AI opportunity and related technologies is really big. I think that there are big social and political problems to help everybody understand that. I think that's what we're staring at a lot of right now; the world is a scary place, and AI is a very uncertain thing. And it takes a lot of work that is not necessarily building things. It's like telling people and understanding people, things that the people building AI are historically not motivated or wanting to do. But it is something that is probably doable. It just will take longer than people want. And we have to go through that long period of hard, distraught AI discussions if we want to have the lasting benefits."


专业术语注释

术语 解释
LLM(Large Language Model,大语言模型) 以Transformer架构为基础,通过大规模语料预训练得到的具备通用语言理解与生成能力的模型,是当前AI技术的核心载体
Transformer 2017年"Attention Is All You Need"论文提出的架构,基于自注意力机制,是当前所有主流大模型的基础架构
Scaling Laws(缩放定律) 指大模型的性能与预训练算力、数据量呈幂律关系,投入更多算力与数据可稳定提升模型性能,当前已经扩展到后训练、推理等多个维度
MoE(Mixture of Experts,混合专家) 大模型架构优化技术,将全连接层替换为多个并行的专家层,每次推理仅激活部分专家,可在不增加推理算力的前提下大幅提升模型参数量
Pre-training(预训练) 大模型训练的第一阶段,基于大规模通用语料做下一词预测,让模型获取通用知识,形成基础模型
Mid-training(中训练) 介于预训练与后训练之间的阶段,基于高质量专用数据(长上下文、推理、代码等)进一步训练,提升模型在特定领域的基础能力,避免灾难性遗忘
Post-training(后训练) 大模型训练的最后阶段,包括监督微调、DPO、RLHF、RLVR等,核心目标是解锁模型的技能、优化用户体验,属于能力解锁阶段而非知识学习阶段
RLHF(Reinforcement Learning from Human Feedback,人类反馈强化学习) 后训练技术的一种,通过人类标注的偏好数据训练奖励模型,再用强化学习优化大模型的输出,提升用户体验与安全性
RLVR(Reinforcement Learning with Verifiable Rewards,可验证奖励强化学习) 2025年兴起的后训练技术,通过可验证的客观规则(数学题答案是否正确、代码是否可运行等)给出奖励,无需人工标注,可大规模训练提升模型的推理、编码能力
DPO(Direct Preference Optimization,直接偏好优化) 后训练技术的一种,无需训练奖励模型,直接基于人类偏好数据优化模型,相比RLHF更简单高效
AGI(Artificial General Intelligence,通用人工智能) 指具备与人类相当的通用认知能力,可适应任意新场景、学习任意新技能的人工智能,当前没有统一的定义与明确的实现时间线
ASI(Artificial Superintelligence,人工超级智能) 指在所有领域都远超人类最聪明个体的人工智能,属于远期预测的概念
KV Cache(键值缓存) 大模型推理优化技术,缓存之前token的注意力键值对,避免重复计算,大幅提升推理速度,降低显存占用
Group Query Attention(分组查询注意力) 注意力机制优化技术,将查询头分组,每组共享键值头,大幅降低KV缓存的显存占用,提升长上下文推理效率
Sliding Window Attention(滑动窗口注意力) 注意力机制优化技术,仅关注当前token之前的固定窗口内的token,降低长上下文推理的算力与显存需求
LoRA(Low-Rank Adaptation,低秩适配) 大模型微调技术,仅更新少量低秩矩阵的参数,无需更新全量模型参数,大幅降低微调的成本与显存需求,适合个性化微调场景
OCR(Optical Character Recognition,光学字符识别) 将图像中的文本转换为可编辑文本的技术,是大模型训练中提取PDF、扫描件等非结构化文本数据的核心工具
Sim-to-real gap(仿真到真实的差距) 机器人领域的核心问题,指在仿真环境中训练的模型迁移到真实世界时出现的性能下降,需要世界模型、领域自适应等技术缩小差距
TPU(Tensor Processing Unit,张量处理单元) Google自研的AI专用芯片,专为Transformer训练与推理优化,相比GPU有更高的能效比,是Google AI基础设施的核心优势
CUDA NVIDIA推出的并行计算平台与编程模型,是深度学习领域的事实标准,经过20多年的发展形成了深厚的生态壁垒,是NVIDIA的核心竞争力
FP8/FP4(8位/4位浮点数) 低精度训练技术,通过降低数值的精度减少显存占用、提升计算吞吐量,是当前大模型训练与推理的核心优化技术之一

延伸思考

  1. 中国开源大模型的快速崛起已经对全球AI格局产生了深远影响,未来中美在开源生态上的竞争将如何演进?美国的Adam项目等本土开源计划能否缩小与中国的差距?
  2. RLVR的出现大幅提升了大模型的推理与编码能力,未来将其扩展到开放场景(创意、写作、复杂决策等)的核心瓶颈是什么?何时能实现突破?
  3. 大模型的版权问题已经成为行业核心风险,类似Spotify的版权分成机制是否适合大模型训练场景?如何平衡创作者权益与AI技术的发展?
  4. AI带来的短期失业问题已经逐渐显现,如何设计合理的社会保障与职业转型机制,降低技术变革带来的社会阵痛?
  5. 开源大模型的快速发展是否会改变当前闭源厂商主导的市场格局?未来开源与闭源的市场份额会呈现怎样的比例?

原文发表:Jan 31, 2026  ·  纪要生成:2026-02-25