AGI timeline

TypeSafe AI (Diogo Almeida) launches System One and Jev: typed, calibrated decisions for software—not chat generation. Claims large speed/cost gains on decision tasks. Milestone framing: decision APIs as a first-class layer beside generative models on the path to capable agent / AGI stacks, even while the classifier-vs-new-architecture debate continues.

The phrase moves from research slides into product announcements — and the debate shifts to what, if anything, has actually changed.

Routing between fast and reasoning modes becomes a product-level feature rather than a research topic.

Native multimodality and mixture-of-experts become standard equipment for open-weight models.

One model switches between fast answers and extended reasoning.

Governments shift from debating risk to competing on adoption and investment.

Open reasoning models reach the frontier, and the open/closed gap narrows sharply.

Strong open-weight performance at a fraction of the training cost, challenging assumptions about the cost curve.

A frontier-lab leader argues publicly for AI's potential to compress decades of progress.

Thinking longer at inference buys accuracy — scaling gains a second axis.

A mid-tier model matching or beating larger ones; coding and agentic ability gain weight.

Real-time voice, vision and text in one model — interaction moves beyond the text box.

The first comprehensive AI regulation becomes law, with risk tiers and obligations for general-purpose models.

Frontier capability across three tiers, with long context and vision as baseline expectations.

Context length becomes a headline capability axis in its own right.

The world's first comprehensive AI law clears its decisive political hurdle.

Cheaper long-context inference plus custom assistants push models from product to platform.

The first state body dedicated to evaluating frontier-model risk.

Frontier-AI risk moves onto the international agenda with government-level commitments.

Safety testing, reporting duties and standards for federal use are set in motion.

Open-weight models become a serious alternative to closed APIs, reshaping the competitive landscape.

Among the first binding rules aimed specifically at generative AI services.

A founding figure of deep learning turns critic — the debate acquires unusual authority.

The first broadly signed public letter on frontier-AI risk, and a turning point in the safety debate.

Multimodal input and expert-level benchmark scores; capability and evaluation both step up.

Written principles plus model self-critique scale alignment beyond human labelling capacity.

A conversational interface brings large models to a hundred million people in two months.

提出算法歧视、数据隐私等五项原则，是美国"以权利为框架"监管思路的代表。

与闭源的 DALL·E 2 形成对照，开源权重让文生图迅速普及到个人设备。

图像生成从研究演示变成大众创作工具。

An AI-safety-first lab is created by former OpenAI researchers, marking safety as a competitive axis.

For a fixed compute budget, more data beats more parameters — a directive that reshapes training runs.

A smaller aligned model is preferred by humans over a larger unaligned one.

面向研发者与使用者的操作级规范，是中国 AI 伦理治理的具体落地文件。

Low-rank adaptation makes customising large models cheap, democratising the last mile.

Stanford's report names and frames “foundation models” as a category with its own risks.

The draft enters the legislative process, setting a global reference point for AI law.

Text and images share one representation space; generation and zero-shot recognition advance together.

The first attempt at comprehensive, risk-tiered AI regulation.

A long-standing scientific grand challenge falls, and AI's value beyond language becomes hard to dispute.

few-shot 能力并非随参数线性增长，而是在某个规模后突然出现，为"规模出智能"提供了直接证据。

收到超过 1200 份意见，是 AI 立法过程中公众参与度最高的一次。

Few-shot prompting works without fine-tuning; “large model” becomes an industry organising principle.

Loss falls predictably with compute and data — turning “make it bigger” into an engineering plan.

提出"负责任 AI"八项原则，是中国参与全球 AI 治理框架的早期动作。

The first intergovernmental set of AI principles, later adopted by the G20.

The first staged release justified by misuse risk — where capability and release ethics first collide in public.

Bidirectional pretraining resets the state of the art on language understanding benchmarks.

Pretrain on raw text, then adapt: the recipe that leads to today's foundation models.

两条路线（自回归生成 vs 双向理解）在同一年出现，此后三年主导 NLP 研究方向的选择。

在正式立法前先建立伦理框架，是"先准则、后立法"监管路径的典型。

Pure self-play surpasses the version trained on human games — a reference point for scaling without human data.

Turning human preferences into a training signal becomes the standard way to steer large models.

Attention alone, without recurrence, becomes the architecture that everything since is built on.

Shazeer 等把 MoE 引入 NLP，为后来以"总参数大、激活参数小"降低推理成本提供了路径。

2018-05-25 生效。其中对"自动化决策"的限制，后来成为 AI 治理的重要法律基础。

Go falls to deep reinforcement learning — a milestone for both capability and public awareness.

Among the first national-level policy documents on AI's economic and social implications.

以"确保 AGI 造福全人类"为使命成立，此后成为大模型时代最主要的推动者之一。

Skip connections lift the depth ceiling and become a default building block of modern architectures.

RL 领域的标准教材，为后来 RLHF 提供了方法基础。

约 4 亿英镑，是当时欧洲最大的 AI 收购，也标志着大公司开始用并购获取顶尖研究团队。

Neural machine translation becomes practical; attention later proves to be the key idea behind the Transformer.

Goodfellow's adversarial training makes generative modelling a mainstream research direction.

为三年后的 Transformer 提供关键组件。

把词映射为稠密向量且保留语义关系，是"表征学习"进入 NLP 的标志。

Hinton 团队加入 Google，工业界开始系统性吸纳深度学习人才。

GPUs plus deep convolutional networks cut the error rate roughly in half; the deep-learning era begins in earnest.

首次用大规模分布式算力做无监督特征学习，证明"规模"本身能带来质变。这条思路后来直接演化为规模定律。

The first industrial-scale victory of neural methods, well before the 2012 vision breakthrough.

A million-labelled-image benchmark gives the field something it had lacked: a shared yardstick.

此后十年算力增长的技术前提。常被忽视，但它是深度学习得以实现的物质基础。

把受限玻尔兹曼机的训练成本降到可用，是"深度网络可训练"这一结论的技术前提。

Shows deep networks can be trained — the precursor to “deep learning” as a programme.

The field acquires its own venues, benchmarks and shared standards.

Vapnik《统计学习理论的本质》出版，为泛化能力提供理论工具。

Brute-force search beats the human champion — a milestone in public perception rather than learning.

Gating solves vanishing gradients; recurrent networks can finally carry information over long spans.

Cortes 与 Vapnik 提出，此后近二十年主导机器学习。

Weight sharing and convolution make handwritten-digit recognition work at scale — today's vision stacks still follow this design.

专用硬件市场萎缩叠加维护成本高企。

Rumelhart, Hinton and Williams popularise backprop, and connectionism returns to the centre.

Physics-inspired models reconnect neural networks with statistical mechanics.

专家系统首次大规模商用，催生 AI 产业第一波热潮。

A practical way to train multi-layer networks; it takes another decade to become mainstream.

第一次 AI 寒冬的制度性成因：承诺未兑现，资金退潮。

Rule-based diagnosis at expert level — the high-water mark of symbolic AI in practice.

对 AI 过度承诺的第一次系统性反噬。

Minsky and Papert show single-layer perceptrons cannot represent XOR, helping trigger the first AI winter.

为处理不确定性提供数学框架，影响此后的控制系统与决策模型。

把梯度下降引入自适应线性单元，是反向传播的前身。

此后三十年 AI 研究的主要编程语言。

The first trainable neural model, and the first AI idea to draw wide public attention.

McCarthy coins the term “artificial intelligence”; the field gains a name, a research agenda and a community.

用 40 个突触模拟赫布学习，是最早的神经网络硬件实现之一。

Proposes the imitation game, reframing “can machines think?” as an operational test — the philosophical foundation of AI.