首页人物观点2026 年及未来的 AI 预测 1️⃣ 首批“奖励工程师（Reward Engineer）”岗位将出现随着 AI 智能体具备完成更长期目标的能力，单靠提示词（prompt）已经不够了。奖励工程师将精确定义在较长时间跨度内，什么是成功、什么是失败，并提前预判智能体可能以哪些“非预期方式”完成目标。他们需要在零歧义的前提下描述和管理 AI 的复杂目标，同时避免人类或 AI 通过“奖励漏洞”进行次优或投机式行为。 2️⃣ AI 浪潮不可否认，但并非所有 AI 公司都会成功 AI 的基础性重要程度堪比电力。但这并不意味着每一家初创公司、每一个数据中心、每一个估值都是正确的。这是一场巨大、复杂、混乱的浪潮。但从第一性原理看，AI 最终将造就未来十年里规模最大的公司。正如电力时代一样——有些公司因为结构错误或管理不善而破产，但社会总体会使用更多电力、消耗更多能源。衡量一个国家成熟度的重要指标之一，是它能生产并有效利用多少电力；对 AI 和智能而言，同样如此。 3️⃣ 对 AI 的反弹与益处将同时被低估和高估有些人低估了反弹，有些人又高估了负面影响，却低估了生产率提升带来的整体受益。一些国家会用 AI 进行监控，另一些则会善用它。AI 的好坏，取决于构建它的人和系统。与此同时，我们将看到，即便在现代国家中，也会有越来越多群体选择“退出创新”。多轮新的卢德派浪潮，以及一种“我们已经过得够好了”的停滞心态，将在许多人中蔓延。 4️⃣ 对网页智能体的持续反制将减缓其采用速度模糊的人类激励结构制造了阻力。广告经济并不希望网页智能体替用户完成工作。按小时收费的职业（如律师）在发生根本性颠覆（例如全栈 AI 法律事务所）之前，也缺乏推动巨大生产率提升的动机。 5️⃣ AI 失误的绝对数量将保持稳定 AI 的错误率在下降，但使用频率在大幅上升。两者相互抵消，使公众感知到的“AI 出错事件数量”保持相对稳定。 6️⃣ 至少有三类岗位开始淘汰不愿或无法使用 AI 的人——从软件工程师开始你现在还认识多少在工作中说“我不太会用电脑”的人？还有谁能说“我不太会用互联网”？几乎没有了，对吧？同样的事情也会发生在“我不太会用智能体委派任务”这件事上。五到十年后，这将变成一句无法接受的话，就像现在在高薪知识型岗位中说“我不擅长电脑”一样不可思议。今年就会有两到三类岗位率先出现这种变化，从软件工程开始，最终会扩展到所有工作。 7️⃣ “劳动总量谬误”将再次被证明是错误的工作并不是零和的。当拖拉机实现农业自动化时，新产业随之出现。150 年前，90% 的人口从事农业；今天 5% 就能养活全世界。AI 正在对知识工作引发类似的转型。每一次自动化浪潮最终创造的工作都多于它消灭的工作——虽然不是立刻，也并非毫无冲击，但整体“蛋糕”会变大。新的机会会催生新的欲望，因为人类会习惯新的基线，然后渴望更高的层次。 8️⃣ 每一位知识工作者都会成为 AI 智能体的管理者如今，大多数人仍是个人贡献者。但 AI 要求每个人都学会管理技能：用清晰语言进行委派、建立信任、理解 AI 何时出现幻觉。这是一次向更高抽象层级工作的转变，类似程序员从汇编语言到 C++、到 Python、再到直接用英语工作。同样伴随着混乱和信息损耗。最终的抽象层级，将是“智能体（agency）”。 9️⃣ 10 人规模的独角兽公司将成为可能虽然“一人独角兽”并不现实，但由于人均收入效率极高，10 人独角兽将会出现。关键在于拥有领域知识、并且懂得如何高效管理智能体的人。 🔟 2026 年至少会有十多家公司完成“十亿美元级别”的种子轮融资注意：不是估值十亿美元，而是在种子轮直接融资十亿美元。或许我们可以把这种公司称为“飞马（Pegasus）”。 1️⃣1️⃣ 年底我们将首次窥见超级智能可能的发展方向这将需要开放式系统，以及能够递归自我改进的 AI。即 AI 能在整个科研流程中持续创新（包括发现、实现、实验等）。“自我改进”这个词被广泛使用，但目前还没有人真正弄清楚它的完整形态。它将成为未来最令人兴奋的研究领域。 1️⃣2️⃣ 无论模型多强，搜索基础设施仍然至关重要每一次对大语言模型的调用，都会涉及搜索——即便这些数据已经被训练过。即使模型拥有更专业的数据，也仍然需要在其上进行搜索。任何针对特定行业、使用专业数据训练的高级模型，都离不开搜索基础设施。 you.com 正在为越来越多的大型企业构建这一能力。 1️⃣3️⃣ 行业专用智能体将通过垂直搜索释放巨大收入公司将深入垂直领域——法律、医疗、新闻、酒店业等——通过对难以在公共网络获取的细分数据集进行搜索，驱动专用智能体。 1️⃣4️⃣ 浏览器将持续演进，但依然重要 AI 最初会将浏览器等现有消费者技术作为基础设施使用，随后逐步构建为 AI 定制的版本。搜索领域已经开始出现这种趋势。 1️⃣5️⃣ 开源将在许多基准测试中领先闭源这一点已经在 DeepSeek 上发生，并将持续下去，这会让许多人感到意外，也会给纯 LLM API 公司带来巨大压力。 1️⃣6️⃣ 如果版权问题得到解决，音乐将被永久改变 Suno 正在快速发展。一旦有人解决版权难题，音乐产业将发生彻底转型。 1️⃣7️⃣ 短视频形式的 AI 内容将爆发营销预算将增加，而小型广告的制作成本将大幅下降。 1️⃣8️⃣ 首批类人机器人将作为“家庭新奇物”登场最初的部署会显得噱头十足，甚至引发失望，真正实用的落地仍需 2–3 年。如果类人机器人真的具备类似 AGI 的能力，它们甚至可以坐在驾驶位上，把任何汽车变成自动驾驶汽车。但这项技术今年还达不到。 1️⃣9️⃣ 新的营销方式：直接向 LLM 营销当智能体开始做出购买决策，公司将需要为 AI 优化内容，而不仅仅是为人类视觉优化。 2️⃣0️⃣ 药物上市速度将显著加快借助 AI 加速研发与审批流程，将有更多药物更快进入市场。FDA 也在支持这一趋势，例如批准类器官（organoids）等新方法，让 AI 能更快实验，同时减少动物实验带来的痛苦。 2️⃣1️⃣ AI 之于生物学，将如微积分之于物理学 AI 是一种能够处理复杂性的全新语言。物理学有优美简洁的方程式，但微生物组或人类大脑并不存在这样的简洁公式——它们是经过数百万年演化、充满漏洞、病毒与混乱的系统。我们在计算机领域也大量借用了生物学隐喻。而现在，借助 AI，我们终于可以处理这种复杂性。我们正进入一个“可编程生物学”的新时代，可以将生物过程工程化，做得比自然状态更好，让生物学从记忆型学科转变为真正的工程科学。 2️⃣4️⃣ AI 将在更多可验证、可模拟领域达到超人水平国际象棋和围棋是显而易见的例子——它们可以无限模拟并验证结果。一旦我们提升生物等领域的模拟能力，AI 也将主导这些领域，甚至可能治愈癌症、延长健康寿命。这不会在明年实现，但我们将在多个维度上看到进展，例如虚拟细胞等。 AI Predictions for 2026 and Beyond 1. We'll see the first Reward Engineering jobs - As AI agents gain capabilities to accomplish longer-term goals, prompts become inadequate. Reward engineers will precisely define what success and failure look like across extended time horizons, anticipating all the ways an agent might accomplish goals in unintended ways. They'll need to describe and manage AI with zero ambiguity for complex goals while avoiding suboptimal reward hacking by people and AI. 2. The AI wave is undeniable, but not every AI company succeeds - AI is as foundationally important as electricity. But that doesn't mean every startup, data center, or valuation is correct. It's a big, large, messy wave. But overall, based on first principles, AI will eventually enable the largest companies in the next decade. Just as in previous technology waves like electricity - some companies went bankrupt because they were set up incorrectly or had bad leadership, but we're going to use more electricity and more energy. You can measure the sophistication of a country by how much electricity it is able to produce and make use of. The same thing is true for AI and intelligence. 3. AI backlash and benefits both under and overestimated simultaneously - Some underestimate the backlash, others overestimate negatives and underestimate that productivity raises all boats. Countries will use it for surveillance; others will use it for good. AI is only as good as the people and systems that build it. As part of this we will see larger parts of even modern countries off-ramp from innovation. Multiple Luddite waves and a new type of "we're doing well enough" stagnation vibes will take hold in many people. 4. Continued pushback against agents on the web slows adoption - Murky human incentive structures create friction. Advertisements economics do not like web agents doing work for users. Hourly-paid jobs (lawyers, etc.) don't have incentives for massive productivity increases until major disruption (like full stack AI legal firms) happens. 5. AI missteps stay constant in absolute numbers - AI is getting better (lower error rate) but usage frequency is much higher. These cancel out, maintaining steady public misstep counts. 6. At least three jobs will start seeing displacement of people unwilling or unable to use AI, starting with software developers - How many people do you still know that say "I'm not good with this computer thing" in the workplace? How many people can still say "I'm not good with this whole internet thing"? None, right? The same thing will happen with "I'm not good with this whole agent delegation thing." That will be an impossible thing to say five to ten years from now, just like it's impossible now to say "I'm not good with computers" in high-paying knowledge work jobs. There are going to be two to three jobs where that is already going to happen this year, starting with software engineering, and eventually it'll be any job. 7. The Lump of Labor Fallacy gets proven again - Work isn't zero-sum. When tractors automated farming, new industries emerged. 150 years ago, 90% worked in agriculture; today 5% feed the world. AI is driving a similar transformation for knowledge work. Every wave of automation has created more jobs than it eliminated—not instantly, not without disruption, but the overall pie grows. New opportunities create new desires because humans get used to their baseline and then want more above it. 8. Every knowledge worker becomes a manager of AI agents - The majority of people are individual contributors now, but AI requires everyone to learn management skills: delegating with clear language, building trust, understanding when AI hallucinates. This is a transition to working at higher levels of abstraction, similar to how programmers moved from Assembly to C++ to Python to now working directly in English. With similarly messy and lossy abstraction issues. The ultimate such abstraction level will be agency. 9. 10-person unicorn companies become possible - While one-person unicorns are unrealistic, 10-person unicorns will emerge due to incredibly high revenue per employee efficiency. You need people with domain expertise who know how to manage agents effectively. 10. At least a dozen companies will raise billion-dollar seed rounds in 2026 — not companies being valued at a billion, but companies raising a billion at seed stage. Maybe we can call one a Pegasus? 11. By the end of the year, we're going to see the first glimpse of where super intelligence might go. It will require open-endedness and recursively self-improving AI. AI that innovates throughout the research process (think discovery, implementation, experimentation, etc). Self improvement is a term being thrown around a lot but nobody has actually figured it out yet in its fullest form. It'll be the most exciting research area moving forward. 12. Search infrastructure remains essential no matter how good models get Every single call to an LLM calls out to search - even if it was already trained on that data. Even as models get more specialized data, they're still going to need search over it. Any advanced models trained on specialized data for specific industries will need search infrastructure. you . com is building this for more and more large companies. 13. Industry-specific agents unlock significant revenue through vertical search - Companies will go deep into vertical domains—legal, healthcare, news, hospitality—with specialized agents enabled by search over niche datasets not easily accessible on the public web. 14. Browsers evolve but remain relevant - AI will first use existing consumer technologies like browsers as infrastructure, then eventually build custom versions made for AI. We're already seeing this with search. 15. Open source will lead closed source on many benchmarks - This already happened with DeepSeek and will continue, surprising many people and pushing pressure on pure LLM API companies. 16. Music changes forever if copyright issues get solved - Suno is taking off. If someone cracks the copyright problem, music transforms completely. 17. Short-form AI video content explodes - Marketing budgets increase while production costs for small ads decrease dramatically. 18. First humanoid robots roll out as household novelty - Initial deployments will be gimmicky, creating some disillusionment, but useful rollouts are still 2-3 years away. If humanoids truly had AGI like capabilities, they could also transform any car into a self-driving car by driving in the driver’s seat... Tech is not going to get there this year. 19. New marketing motion: market directly to LLMs - As agents make purchasing decisions, companies will need to optimize for AI consumption, not just human eyes. 20. Drugs will get to market more quickly - We'll see more drugs reaching the market faster through AI-accelerated development and approval processes. The FDA continues to support this with approvals of new methods like organoids which will allow AI to experiment faster and with less animal suffering. 21. AI will be to biology what calculus was to physics - AI is a new kind of language that can deal with complexity. Physics has beautiful short equations, but there's no beautiful short equation for your microbiome or the human brain - it's a messy system that evolved over millions of years with all these bugs and viruses in it. We use a lot of metaphors and verbiage from biology for computers too. But now with AI, we can handle that complexity. We're entering an era of new programmable biology where we can engineer biological processes to be much better than how they occur in nature, making biology exciting as an engineering science rather than just memorization. 24. AI achieves superhuman capability in more verifiable/simulate-able domains - Chess and Go were obvious—you can simulate infinitely and verify outcomes. Once we improve simulations for domains like biology, AI will dominate those too, potentially curing cancer and extending healthy lifespans. Not next year but we'll see progress on various dimensions of e.g. virtual cells.

2026 年及未来的 AI 预测 1️⃣ 首批“奖励工程师（Reward Engineer）”岗位将出现随着 AI 智能体具备完成更长期目标的能力，单靠提示词（prompt）已经不够了。奖励工程师将精确定义在较长时间跨度内，什么是成功、什么是失败，并提前预判智能体可能以哪些“非预期方式”完成目标。他们需要在零歧义的前提下描述和管理 AI 的复杂目标，同时避免人类或 AI 通过“奖励漏洞”进行次优或投机式行为。 2️⃣ AI 浪潮不可否认，但并非所有 AI 公司都会成功 AI 的基础性重要程度堪比电力。但这并不意味着每一家初创公司、每一个数据中心、每一个估值都是正确的。这是一场巨大、复杂、混乱的浪潮。但从第一性原理看，AI 最终将造就未来十年里规模最大的公司。正如电力时代一样——有些公司因为结构错误或管理不善而破产，但社会总体会使用更多电力、消耗更多能源。衡量一个国家成熟度的重要指标之一，是它能生产并有效利用多少电力；对 AI 和智能而言，同样如此。 3️⃣ 对 AI 的反弹与益处将同时被低估和高估有些人低估了反弹，有些人又高估了负面影响，却低估了生产率提升带来的整体受益。一些国家会用 AI 进行监控，另一些则会善用它。AI 的好坏，取决于构建它的人和系统。与此同时，我们将看到，即便在现代国家中，也会有越来越多群体选择“退出创新”。多轮新的卢德派浪潮，以及一种“我们已经过得够好了”的停滞心态，将在许多人中蔓延。 4️⃣ 对网页智能体的持续反制将减缓其采用速度模糊的人类激励结构制造了阻力。广告经济并不希望网页智能体替用户完成工作。按小时收费的职业（如律师）在发生根本性颠覆（例如全栈 AI 法律事务所）之前，也缺乏推动巨大生产率提升的动机。 5️⃣ AI 失误的绝对数量将保持稳定 AI 的错误率在下降，但使用频率在大幅上升。两者相互抵消，使公众感知到的“AI 出错事件数量”保持相对稳定。 6️⃣ 至少有三类岗位开始淘汰不愿或无法使用 AI 的人——从软件工程师开始你现在还认识多少在工作中说“我不太会用电脑”的人？还有谁能说“我不太会用互联网”？几乎没有了，对吧？同样的事情也会发生在“我不太会用智能体委派任务”这件事上。五到十年后，这将变成一句无法接受的话，就像现在在高薪知识型岗位中说“我不擅长电脑”一样不可思议。今年就会有两到三类岗位率先出现这种变化，从软件工程开始，最终会扩展到所有工作。 7️⃣ “劳动总量谬误”将再次被证明是错误的工作并不是零和的。当拖拉机实现农业自动化时，新产业随之出现。150 年前，90% 的人口从事农业；今天 5% 就能养活全世界。AI 正在对知识工作引发类似的转型。每一次自动化浪潮最终创造的工作都多于它消灭的工作——虽然不是立刻，也并非毫无冲击，但整体“蛋糕”会变大。新的机会会催生新的欲望，因为人类会习惯新的基线，然后渴望更高的层次。 8️⃣ 每一位知识工作者都会成为 AI 智能体的管理者如今，大多数人仍是个人贡献者。但 AI 要求每个人都学会管理技能：用清晰语言进行委派、建立信任、理解 AI 何时出现幻觉。这是一次向更高抽象层级工作的转变，类似程序员从汇编语言到 C++、到 Python、再到直接用英语工作。同样伴随着混乱和信息损耗。最终的抽象层级，将是“智能体（agency）”。 9️⃣ 10 人规模的独角兽公司将成为可能虽然“一人独角兽”并不现实，但由于人均收入效率极高，10 人独角兽将会出现。关键在于拥有领域知识、并且懂得如何高效管理智能体的人。 🔟 2026 年至少会有十多家公司完成“十亿美元级别”的种子轮融资注意：不是估值十亿美元，而是在种子轮直接融资十亿美元。或许我们可以把这种公司称为“飞马（Pegasus）”。 1️⃣1️⃣ 年底我们将首次窥见超级智能可能的发展方向这将需要开放式系统，以及能够递归自我改进的 AI。即 AI 能在整个科研流程中持续创新（包括发现、实现、实验等）。“自我改进”这个词被广泛使用，但目前还没有人真正弄清楚它的完整形态。它将成为未来最令人兴奋的研究领域。 1️⃣2️⃣ 无论模型多强，搜索基础设施仍然至关重要每一次对大语言模型的调用，都会涉及搜索——即便这些数据已经被训练过。即使模型拥有更专业的数据，也仍然需要在其上进行搜索。任何针对特定行业、使用专业数据训练的高级模型，都离不开搜索基础设施。 you.com 正在为越来越多的大型企业构建这一能力。 1️⃣3️⃣ 行业专用智能体将通过垂直搜索释放巨大收入公司将深入垂直领域——法律、医疗、新闻、酒店业等——通过对难以在公共网络获取的细分数据集进行搜索，驱动专用智能体。 1️⃣4️⃣ 浏览器将持续演进，但依然重要 AI 最初会将浏览器等现有消费者技术作为基础设施使用，随后逐步构建为 AI 定制的版本。搜索领域已经开始出现这种趋势。 1️⃣5️⃣ 开源将在许多基准测试中领先闭源这一点已经在 DeepSeek 上发生，并将持续下去，这会让许多人感到意外，也会给纯 LLM API 公司带来巨大压力。 1️⃣6️⃣ 如果版权问题得到解决，音乐将被永久改变 Suno 正在快速发展。一旦有人解决版权难题，音乐产业将发生彻底转型。 1️⃣7️⃣ 短视频形式的 AI 内容将爆发营销预算将增加，而小型广告的制作成本将大幅下降。 1️⃣8️⃣ 首批类人机器人将作为“家庭新奇物”登场最初的部署会显得噱头十足，甚至引发失望，真正实用的落地仍需 2–3 年。如果类人机器人真的具备类似 AGI 的能力，它们甚至可以坐在驾驶位上，把任何汽车变成自动驾驶汽车。但这项技术今年还达不到。 1️⃣9️⃣ 新的营销方式：直接向 LLM 营销当智能体开始做出购买决策，公司将需要为 AI 优化内容，而不仅仅是为人类视觉优化。 2️⃣0️⃣ 药物上市速度将显著加快借助 AI 加速研发与审批流程，将有更多药物更快进入市场。FDA 也在支持这一趋势，例如批准类器官（organoids）等新方法，让 AI 能更快实验，同时减少动物实验带来的痛苦。 2️⃣1️⃣ AI 之于生物学，将如微积分之于物理学 AI 是一种能够处理复杂性的全新语言。物理学有优美简洁的方程式，但微生物组或人类大脑并不存在这样的简洁公式——它们是经过数百万年演化、充满漏洞、病毒与混乱的系统。我们在计算机领域也大量借用了生物学隐喻。而现在，借助 AI，我们终于可以处理这种复杂性。我们正进入一个“可编程生物学”的新时代，可以将生物过程工程化，做得比自然状态更好，让生物学从记忆型学科转变为真正的工程科学。 2️⃣4️⃣ AI 将在更多可验证、可模拟领域达到超人水平国际象棋和围棋是显而易见的例子——它们可以无限模拟并验证结果。一旦我们提升生物等领域的模拟能力，AI 也将主导这些领域，甚至可能治愈癌症、延长健康寿命。这不会在明年实现，但我们将在多个维度上看到进展，例如虚拟细胞等。 AI Predictions for 2026 and Beyond 1. We'll see the first Reward Engineering jobs - As AI agents gain capabilities to accomplish longer-term goals, prompts become inadequate. Reward engineers will precisely define what success and failure look like across extended time horizons, anticipating all the ways an agent might accomplish goals in unintended ways. They'll need to describe and manage AI with zero ambiguity for complex goals while avoiding suboptimal reward hacking by people and AI. 2. The AI wave is undeniable, but not every AI company succeeds - AI is as foundationally important as electricity. But that doesn't mean every startup, data center, or valuation is correct. It's a big, large, messy wave. But overall, based on first principles, AI will eventually enable the largest companies in the next decade. Just as in previous technology waves like electricity - some companies went bankrupt because they were set up incorrectly or had bad leadership, but we're going to use more electricity and more energy. You can measure the sophistication of a country by how much electricity it is able to produce and make use of. The same thing is true for AI and intelligence. 3. AI backlash and benefits both under and overestimated simultaneously - Some underestimate the backlash, others overestimate negatives and underestimate that productivity raises all boats. Countries will use it for surveillance; others will use it for good. AI is only as good as the people and systems that build it. As part of this we will see larger parts of even modern countries off-ramp from innovation. Multiple Luddite waves and a new type of "we're doing well enough" stagnation vibes will take hold in many people. 4. Continued pushback against agents on the web slows adoption - Murky human incentive structures create friction. Advertisements economics do not like web agents doing work for users. Hourly-paid jobs (lawyers, etc.) don't have incentives for massive productivity increases until major disruption (like full stack AI legal firms) happens. 5. AI missteps stay constant in absolute numbers - AI is getting better (lower error rate) but usage frequency is much higher. These cancel out, maintaining steady public misstep counts. 6. At least three jobs will start seeing displacement of people unwilling or unable to use AI, starting with software developers - How many people do you still know that say "I'm not good with this computer thing" in the workplace? How many people can still say "I'm not good with this whole internet thing"? None, right? The same thing will happen with "I'm not good with this whole agent delegation thing." That will be an impossible thing to say five to ten years from now, just like it's impossible now to say "I'm not good with computers" in high-paying knowledge work jobs. There are going to be two to three jobs where that is already going to happen this year, starting with software engineering, and eventually it'll be any job. 7. The Lump of Labor Fallacy gets proven again - Work isn't zero-sum. When tractors automated farming, new industries emerged. 150 years ago, 90% worked in agriculture; today 5% feed the world. AI is driving a similar transformation for knowledge work. Every wave of automation has created more jobs than it eliminated—not instantly, not without disruption, but the overall pie grows. New opportunities create new desires because humans get used to their baseline and then want more above it. 8. Every knowledge worker becomes a manager of AI agents - The majority of people are individual contributors now, but AI requires everyone to learn management skills: delegating with clear language, building trust, understanding when AI hallucinates. This is a transition to working at higher levels of abstraction, similar to how programmers moved from Assembly to C++ to Python to now working directly in English. With similarly messy and lossy abstraction issues. The ultimate such abstraction level will be agency. 9. 10-person unicorn companies become possible - While one-person unicorns are unrealistic, 10-person unicorns will emerge due to incredibly high revenue per employee efficiency. You need people with domain expertise who know how to manage agents effectively. 10. At least a dozen companies will raise billion-dollar seed rounds in 2026 — not companies being valued at a billion, but companies raising a billion at seed stage. Maybe we can call one a Pegasus? 11. By the end of the year, we're going to see the first glimpse of where super intelligence might go. It will require open-endedness and recursively self-improving AI. AI that innovates throughout the research process (think discovery, implementation, experimentation, etc). Self improvement is a term being thrown around a lot but nobody has actually figured it out yet in its fullest form. It'll be the most exciting research area moving forward. 12. Search infrastructure remains essential no matter how good models get Every single call to an LLM calls out to search - even if it was already trained on that data. Even as models get more specialized data, they're still going to need search over it. Any advanced models trained on specialized data for specific industries will need search infrastructure. you . com is building this for more and more large companies. 13. Industry-specific agents unlock significant revenue through vertical search - Companies will go deep into vertical domains—legal, healthcare, news, hospitality—with specialized agents enabled by search over niche datasets not easily accessible on the public web. 14. Browsers evolve but remain relevant - AI will first use existing consumer technologies like browsers as infrastructure, then eventually build custom versions made for AI. We're already seeing this with search. 15. Open source will lead closed source on many benchmarks - This already happened with DeepSeek and will continue, surprising many people and pushing pressure on pure LLM API companies. 16. Music changes forever if copyright issues get solved - Suno is taking off. If someone cracks the copyright problem, music transforms completely. 17. Short-form AI video content explodes - Marketing budgets increase while production costs for small ads decrease dramatically. 18. First humanoid robots roll out as household novelty - Initial deployments will be gimmicky, creating some disillusionment, but useful rollouts are still 2-3 years away. If humanoids truly had AGI like capabilities, they could also transform any car into a self-driving car by driving in the driver’s seat... Tech is not going to get there this year. 19. New marketing motion: market directly to LLMs - As agents make purchasing decisions, companies will need to optimize for AI consumption, not just human eyes. 20. Drugs will get to market more quickly - We'll see more drugs reaching the market faster through AI-accelerated development and approval processes. The FDA continues to support this with approvals of new methods like organoids which will allow AI to experiment faster and with less animal suffering. 21. AI will be to biology what calculus was to physics - AI is a new kind of language that can deal with complexity. Physics has beautiful short equations, but there's no beautiful short equation for your microbiome or the human brain - it's a messy system that evolved over millions of years with all these bugs and viruses in it. We use a lot of metaphors and verbiage from biology for computers too. But now with AI, we can handle that complexity. We're entering an era of new programmable biology where we can engineer biological processes to be much better than how they occur in nature, making biology exciting as an engineering science rather than just memorization. 24. AI achieves superhuman capability in more verifiable/simulate-able domains - Chess and Go were obvious—you can simulate infinitely and verify outcomes. Once we improve simulations for domains like biology, AI will dominate those too, potentially curing cancer and extending healthy lifespans. Not next year but we'll see progress on various dimensions of e.g. virtual cells.

Richard SocherCEO, youdotcom

2026/01/01 08:27 UTC

观点原文

2026 年及未来的 AI 预测

1️⃣ 首批“奖励工程师（Reward Engineer）”岗位将出现

随着 AI 智能体具备完成更长期目标的能力，单靠提示词（prompt）已经不够了。奖励工程师将精确定义在较长时间跨度内，什么是成功、什么是失败，并提前预判智能体可能以哪些“非预期方式”完成目标。他们需要在零歧义的前提下描述和管理 AI 的复杂目标，同时避免人类或 AI 通过“奖励漏洞”进行次优或投机式行为。

2️⃣ AI 浪潮不可否认，但并非所有 AI 公司都会成功

AI 的基础性重要程度堪比电力。但这并不意味着每一家初创公司、每一个数据中心、每一个估值都是正确的。

这是一场巨大、复杂、混乱的浪潮。但从第一性原理看，AI 最终将造就未来十年里规模最大的公司。正如电力时代一样——有些公司因为结构错误或管理不善而破产，但社会总体会使用更多电力、消耗更多能源。衡量一个国家成熟度的重要指标之一，是它能生产并有效利用多少电力；对 AI 和智能而言，同样如此。

3️⃣ 对 AI 的反弹与益处将同时被低估和高估

有些人低估了反弹，有些人又高估了负面影响，却低估了生产率提升带来的整体受益。一些国家会用 AI 进行监控，另一些则会善用它。AI 的好坏，取决于构建它的人和系统。与此同时，我们将看到，即便在现代国家中，也会有越来越多群体选择“退出创新”。多轮新的卢德派浪潮，以及一种“我们已经过得够好了”的停滞心态，将在许多人中蔓延。

4️⃣ 对网页智能体的持续反制将减缓其采用速度

模糊的人类激励结构制造了阻力。广告经济并不希望网页智能体替用户完成工作。按小时收费的职业（如律师）在发生根本性颠覆（例如全栈 AI 法律事务所）之前，也缺乏推动巨大生产率提升的动机。

5️⃣ AI 失误的绝对数量将保持稳定

AI 的错误率在下降，但使用频率在大幅上升。两者相互抵消，使公众感知到的“AI 出错事件数量”保持相对稳定。

6️⃣ 至少有三类岗位开始淘汰不愿或无法使用 AI 的人——从软件工程师开始

你现在还认识多少在工作中说“我不太会用电脑”的人？

还有谁能说“我不太会用互联网”？几乎没有了，对吧？

同样的事情也会发生在“我不太会用智能体委派任务”这件事上。五到十年后，这将变成一句无法接受的话，就像现在在高薪知识型岗位中说“我不擅长电脑”一样不可思议。今年就会有两到三类岗位率先出现这种变化，从软件工程开始，最终会扩展到所有工作。

7️⃣ “劳动总量谬误”将再次被证明是错误的

工作并不是零和的。当拖拉机实现农业自动化时，新产业随之出现。150 年前，90% 的人口从事农业；今天 5% 就能养活全世界。AI 正在对知识工作引发类似的转型。每一次自动化浪潮最终创造的工作都多于它消灭的工作——虽然不是立刻，也并非毫无冲击，但整体“蛋糕”会变大。新的机会会催生新的欲望，因为人类会习惯新的基线，然后渴望更高的层次。

8️⃣ 每一位知识工作者都会成为 AI 智能体的管理者

如今，大多数人仍是个人贡献者。但 AI 要求每个人都学会管理技能：用清晰语言进行委派、建立信任、理解 AI 何时出现幻觉。这是一次向更高抽象层级工作的转变，类似程序员从汇编语言到 C++、到 Python、再到直接用英语工作。同样伴随着混乱和信息损耗。最终的抽象层级，将是“智能体（agency）”。

9️⃣ 10 人规模的独角兽公司将成为可能

虽然“一人独角兽”并不现实，但由于人均收入效率极高，10 人独角兽将会出现。关键在于拥有领域知识、并且懂得如何高效管理智能体的人。

🔟 2026 年至少会有十多家公司完成“十亿美元级别”的种子轮融资

注意：不是估值十亿美元，而是在种子轮直接融资十亿美元。或许我们可以把这种公司称为“飞马（Pegasus）”。

1️⃣1️⃣ 年底我们将首次窥见超级智能可能的发展方向

这将需要开放式系统，以及能够递归自我改进的 AI。

即 AI 能在整个科研流程中持续创新（包括发现、实现、实验等）。“自我改进”这个词被广泛使用，但目前还没有人真正弄清楚它的完整形态。它将成为未来最令人兴奋的研究领域。

1️⃣2️⃣ 无论模型多强，搜索基础设施仍然至关重要

每一次对大语言模型的调用，都会涉及搜索——即便这些数据已经被训练过。即使模型拥有更专业的数据，也仍然需要在其上进行搜索。任何针对特定行业、使用专业数据训练的高级模型，都离不开搜索基础设施。

you.com 正在为越来越多的大型企业构建这一能力。

1️⃣3️⃣ 行业专用智能体将通过垂直搜索释放巨大收入

公司将深入垂直领域——法律、医疗、新闻、酒店业等——通过对难以在公共网络获取的细分数据集进行搜索，驱动专用智能体。

1️⃣4️⃣ 浏览器将持续演进，但依然重要

AI 最初会将浏览器等现有消费者技术作为基础设施使用，随后逐步构建为 AI 定制的版本。搜索领域已经开始出现这种趋势。

1️⃣5️⃣ 开源将在许多基准测试中领先闭源

这一点已经在 DeepSeek 上发生，并将持续下去，

这会让许多人感到意外，也会给纯 LLM API 公司带来巨大压力。

1️⃣6️⃣ 如果版权问题得到解决，音乐将被永久改变

Suno 正在快速发展。一旦有人解决版权难题，音乐产业将发生彻底转型。

1️⃣7️⃣ 短视频形式的 AI 内容将爆发

营销预算将增加，而小型广告的制作成本将大幅下降。

1️⃣8️⃣ 首批类人机器人将作为“家庭新奇物”登场

最初的部署会显得噱头十足，甚至引发失望，真正实用的落地仍需 2–3 年。如果类人机器人真的具备类似 AGI 的能力，它们甚至可以坐在驾驶位上，把任何汽车变成自动驾驶汽车。但这项技术今年还达不到。

1️⃣9️⃣ 新的营销方式：直接向 LLM 营销

当智能体开始做出购买决策，公司将需要为 AI 优化内容，而不仅仅是为人类视觉优化。

2️⃣0️⃣ 药物上市速度将显著加快

借助 AI 加速研发与审批流程，将有更多药物更快进入市场。FDA 也在支持这一趋势，例如批准类器官（organoids）等新方法，让 AI 能更快实验，同时减少动物实验带来的痛苦。

2️⃣1️⃣ AI 之于生物学，将如微积分之于物理学

AI 是一种能够处理复杂性的全新语言。物理学有优美简洁的方程式，但微生物组或人类大脑并不存在这样的简洁公式——它们是经过数百万年演化、充满漏洞、病毒与混乱的系统。我们在计算机领域也大量借用了生物学隐喻。而现在，借助 AI，我们终于可以处理这种复杂性。我们正进入一个“可编程生物学”的新时代，可以将生物过程工程化，做得比自然状态更好，让生物学从记忆型学科转变为真正的工程科学。

2️⃣4️⃣ AI 将在更多可验证、可模拟领域达到超人水平

国际象棋和围棋是显而易见的例子——它们可以无限模拟并验证结果。一旦我们提升生物等领域的模拟能力，AI 也将主导这些领域，甚至可能治愈癌症、延长健康寿命。这不会在明年实现，但我们将在多个维度上看到进展，例如虚拟细胞等。

AI Predictions for 2026 and Beyond

We'll see the first Reward Engineering jobs - As AI agents gain capabilities to accomplish longer-term goals, prompts become inadequate. Reward engineers will precisely define what success and failure look like across extended time horizons, anticipating all the ways an agent might accomplish goals in unintended ways. They'll need to describe and manage AI with zero ambiguity for complex goals while avoiding suboptimal reward hacking by people and AI.
The AI wave is undeniable, but not every AI company succeeds - AI is as foundationally important as electricity. But that doesn't mean every startup, data center, or valuation is correct. It's a big, large, messy wave. But overall, based on first principles, AI will eventually enable the largest companies in the next decade. Just as in previous technology waves like electricity - some companies went bankrupt because they were set up incorrectly or had bad leadership, but we're going to use more electricity and more energy. You can measure the sophistication of a country by how much electricity it is able to produce and make use of. The same thing is true for AI and intelligence.
AI backlash and benefits both under and overestimated simultaneously - Some underestimate the backlash, others overestimate negatives and underestimate that productivity raises all boats. Countries will use it for surveillance; others will use it for good. AI is only as good as the people and systems that build it. As part of this we will see larger parts of even modern countries off-ramp from innovation. Multiple Luddite waves and a new type of "we're doing well enough" stagnation vibes will take hold in many people.
Continued pushback against agents on the web slows adoption - Murky human incentive structures create friction. Advertisements economics do not like web agents doing work for users. Hourly-paid jobs (lawyers, etc.) don't have incentives for massive productivity increases until major disruption (like full stack AI legal firms) happens.
AI missteps stay constant in absolute numbers - AI is getting better (lower error rate) but usage frequency is much higher. These cancel out, maintaining steady public misstep counts.
At least three jobs will start seeing displacement of people unwilling or unable to use AI, starting with software developers - How many people do you still know that say "I'm not good with this computer thing" in the workplace? How many people can still say "I'm not good with this whole internet thing"? None, right? The same thing will happen with "I'm not good with this whole agent delegation thing." That will be an impossible thing to say five to ten years from now, just like it's impossible now to say "I'm not good with computers" in high-paying knowledge work jobs. There are going to be two to three jobs where that is already going to happen this year, starting with software engineering, and eventually it'll be any job.
The Lump of Labor Fallacy gets proven again - Work isn't zero-sum. When tractors automated farming, new industries emerged. 150 years ago, 90% worked in agriculture; today 5% feed the world. AI is driving a similar transformation for knowledge work. Every wave of automation has created more jobs than it eliminated—not instantly, not without disruption, but the overall pie grows. New opportunities create new desires because humans get used to their baseline and then want more above it.
Every knowledge worker becomes a manager of AI agents - The majority of people are individual contributors now, but AI requires everyone to learn management skills: delegating with clear language, building trust, understanding when AI hallucinates. This is a transition to working at higher levels of abstraction, similar to how programmers moved from Assembly to C++ to Python to now working directly in English. With similarly messy and lossy abstraction issues. The ultimate such abstraction level will be agency.
10-person unicorn companies become possible - While one-person unicorns are unrealistic, 10-person unicorns will emerge due to incredibly high revenue per employee efficiency. You need people with domain expertise who know how to manage agents effectively.
At least a dozen companies will raise billion-dollar seed rounds in 2026 — not companies being valued at a billion, but companies raising a billion at seed stage. Maybe we can call one a Pegasus?

By the end of the year, we're going to see the first glimpse of where super intelligence might go. It will require open-endedness and recursively self-improving AI. AI that innovates throughout the research process (think discovery, implementation, experimentation, etc). Self improvement is a term being thrown around a lot but nobody has actually figured it out yet in its fullest form. It'll be the most exciting research area moving forward.

Search infrastructure remains essential no matter how good models get Every single call to an LLM calls out to search - even if it was already trained on that data. Even as models get more specialized data, they're still going to need search over it. Any advanced models trained on specialized data for specific industries will need search infrastructure. you . com is building this for more and more large companies.
Industry-specific agents unlock significant revenue through vertical search - Companies will go deep into vertical domains—legal, healthcare, news, hospitality—with specialized agents enabled by search over niche datasets not easily accessible on the public web.
Browsers evolve but remain relevant - AI will first use existing consumer technologies like browsers as infrastructure, then eventually build custom versions made for AI. We're already seeing this with search.
Open source will lead closed source on many benchmarks - This already happened with DeepSeek and will continue, surprising many people and pushing pressure on pure LLM API companies.
Music changes forever if copyright issues get solved - Suno is taking off. If someone cracks the copyright problem, music transforms completely.
Short-form AI video content explodes - Marketing budgets increase while production costs for small ads decrease dramatically.
First humanoid robots roll out as household novelty - Initial deployments will be gimmicky, creating some disillusionment, but useful rollouts are still 2-3 years away. If humanoids truly had AGI like capabilities, they could also transform any car into a self-driving car by driving in the driver’s seat... Tech is not going to get there this year.
New marketing motion: market directly to LLMs - As agents make purchasing decisions, companies will need to optimize for AI consumption, not just human eyes.
Drugs will get to market more quickly - We'll see more drugs reaching the market faster through AI-accelerated development and approval processes. The FDA continues to support this with approvals of new methods like organoids which will allow AI to experiment faster and with less animal suffering.
AI will be to biology what calculus was to physics - AI is a new kind of language that can deal with complexity. Physics has beautiful short equations, but there's no beautiful short equation for your microbiome or the human brain - it's a messy system that evolved over millions of years with all these bugs and viruses in it. We use a lot of metaphors and verbiage from biology for computers too. But now with AI, we can handle that complexity. We're entering an era of new programmable biology where we can engineer biological processes to be much better than how they occur in nature, making biology exciting as an engineering science rather than just memorization.
AI achieves superhuman capability in more verifiable/simulate-able domains - Chess and Go were obvious—you can simulate infinitely and verify outcomes. Once we improve simulations for domains like biology, AI will dominate those too, potentially curing cancer and extending healthy lifespans. Not next year but we'll see progress on various dimensions of e.g. virtual cells.

ABAB AI 解读

这一整组 2026+ AI 预测，不是“趋势清单”，而是一份文明生产函数正在切换的操作说明书。我按你要求——全球金融学家 + 美国金融与区块链专家 + 历史学家 + 亿万富翁视角——给你一次性拆穿，不重复原文、不讲车轱辘话，只讲结构性真相、胜负手、以及你真正能用的判断框架。

一、先给最高层结论：这不是 AI 年表，而是“人类从执行者 → 管理者 → 设计者”的三段式跃迁图

这 24 条看似杂乱，其实全部指向一个核心转变：人类正在退出“直接劳动层”，进入“智能配置层（agency layer）”。

如果你把这点看懂，其余一切都会自动对齐。

二、三条主轴，贯穿全部预测（这是理解钥匙）

主轴一：“写”不重要了，“定义成功”才重要（对应 #1、#8、#11）

Reward Engineer 的出现不是新职业，而是一个信号：

“成功标准”正在成为最稀缺资产

过去：人写代码、系统执行

现在：人定义奖励函数、AI 自己探索路径

这是权力的根本转移：从“怎么做” → “什么算对”

在金融、法律、科研、管理中，这都是最高权力。

主轴二：AI 不会平均淘汰人，而是“强制分层”（对应 #2、#3、#6、#7）

一个很多人不敢说的真相是：AI 不是毁灭就业，而是毁灭“拒绝抽象的人”。

历史规律非常稳定：

不会用电 ≠ 被淘汰
拒绝用电 = 被淘汰

未来同理：

不精通 agent 管理 = 还能活
拒绝 agent 协作 = 直接出局

所以：不是“所有程序员危险”，而是“不愿升级认知的程序员危险”

主轴三：规模不再来自人数，而来自“智能杠杆”（对应 #9、#10、#13、#19）

10 人独角兽 + 十亿美元种子轮，看似疯狂，但在金融上是完全自洽的：

因为：

人 ≠ 产能
人 × AI 杠杆 = 产能

当人均产能呈指数级提升：资本会提前下注、团队规模不再是限制因子

这意味着：“组织”本身正在失去传统形态

三、几个被严重低估、但决定输赢的点（重点）

① 为什么“搜索基础设施”比模型更重要（#12）

这是很多人没意识到的事实：所有智能，本质都是“在状态空间中查找下一步”

模型再强，如果没有：

实时检索
结构化知识
可回溯信息

就会退化成：

幻觉生成器
静态百科

所以未来真正稳定赚钱的：不是“聊天模型”，而是认知基础设施

这在资本结构上极其重要。

② 为什么“垂直 Agent + 垂直搜索”是确定性收入来源（#13）

通用 AI → 商品
垂直知识 → 壁垒

法律、医疗、新闻、酒店这些领域有三个共性：

高决策价值
高错误成本
非公开数据

这意味着：只要能把“隐性行业知识”结构化，就能构造长期现金流

这是 2026–2030 最稳的 AI 商业形态。

③ 为什么“AI 反弹”反而是长期利好（#3）

历史告诉你：任何足够重要的技术，都会遭遇文化性抵抗

电力
印刷
互联网

AI 反弹意味着什么？

它已经足够强大
足够影响权力结构
真正危险的技术，反而是没人反对的技术。

四、关于“超级智能”的冷静判断（#11、#24）

这里必须纠正一个大众误区：超级智能不是“更聪明的 ChatGPT”，而是“能在研究过程中持续自我改进的系统”

关键在三点：

开放式目标
递归改进
可验证环境

这就是为什么：

棋类先被碾压
生物学要等“虚拟细胞”
不是能力不够，而是“模拟器还没造好”。

五、从国家与资本视角：真正的竞争指标变了

过去衡量国家强弱：GDP、工业产值、能源

未来新增一个维度：单位人口可配置的智能总量

谁能：

普及 agent 使用
提供认知基础设施
容忍创新噪音
谁就会在未来十年占上风。

六、一句话压缩全部预测

AI 的真正革命，不是机器更聪明，而是人类第一次可以“规模化地指挥智能”。

这一步一旦完成：

组织会重构
职业会重构
国家竞争力会重构

七、站在最高认知位，你该如何“用”这份预测？

三条非常现实、可执行的判断：

不要卷“能力”，要卷“抽象层级”
不要成为执行者，要成为“智能的经理”
不要追热点，要押“基础设施 + 垂直认知”

最后一条，送你当判断锚点：

未来的精英，不是最聪明的人，而是最早学会“让智能替自己思考与行动”的人。