openskills.info
← AI Engineer career guide

Data and AI

AI Engineer timeline

How the field of a AI Engineer got here: the sourced milestones from every published course of this career, on one axis.

2004 to 2026

200 milestones across 22 courses

  • Core
  • Supporting
  • Specialization

The axis stretches toward recent years, where most milestones fall. Bars count milestones per year; dots are single milestones.

2000s

1 milestone

  1. 2004

    Learning with Drift Detection

    The DDM method framed concept-drift detection as a streaming error-rate problem. It gave later production monitoring systems a concrete pattern for detecting a changing data-generating process.

2010s

50 milestones

  1. 2013 Dec

    Variational autoencoders

    Kingma and Welling's "Auto-Encoding Variational Bayes" introduced the VAE, establishing the encoder-decoder latent-representation toolkit that latent diffusion later builds on.

  2. 2014 Jun

    Generative adversarial networks

    Goodfellow et al. proposed training a generator and a discriminator in a minimax game. GANs dominated image synthesis quality for the following six years, the baseline diffusion later beat.

  3. 2015

    Hidden Technical Debt in Machine Learning Systems

    Google researchers described ML systems as carrying dependencies and maintenance risks beyond the learned model. Monitoring became part of operating the surrounding system, not an afterthought to model training.

  4. 2015

    Hidden Technical Debt in Machine Learning Systems published

    Google researchers documented how production ML accumulates debt in data dependencies and surrounding systems, not only in model code. The work made lifecycle engineering a first-class concern.

  5. 2015 Nov

    TensorFlow is released as open source

    TensorFlow made a production-capable machine-learning system broadly available. It created demand for a distinct layer that could load trained TensorFlow models and expose inference to applications.

  6. 2016 Feb

    LIME proposes local surrogate explanations

    LIME introduced an approach that learns an interpretable model around an individual prediction. It made model-agnostic local explanation a practical pattern for classifiers.

  7. 2016 Feb

    TensorFlow Serving is released as open source

    TensorFlow Serving separated model lifecycle and serving concerns from training code. Versioned model loading and serving APIs became reusable infrastructure rather than application-specific glue.

  8. 2016 Sep

    Cloud Machine Learning Engine launches

    Google Cloud introduced its managed training and prediction service for TensorFlow models, the direct predecessor whose training-jobs and model-serving concepts Vertex AI later absorbed.

  9. 2016 Sep

    Company founded as a chatbot app

    Hugging Face was founded in New York by Clement Delangue, Julien Chaumond, and Thomas Wolf as a consumer chatbot aimed at teenagers, named after the hugging face emoji.

  10. 2017

    The ML Test Score

    The rubric made testing and monitoring explicit production-readiness concerns. It connected model monitoring with an engineering checklist rather than a dashboard added after release.

  11. 2017 Mar

    Chatbot app launches on iOS

    The app launched as a free iOS chatbot companion aimed at teenagers; the company had raised a seed round from Betaworks and SV Angel.

  12. 2017 May

    SHAP unifies additive feature attribution

    The SHAP paper defined a unified additive feature-attribution framework based on Shapley values. It gave practitioners a common vocabulary for assigning feature contributions to one prediction.

  14. 2017 Jun

    Integrated Gradients introduces an axiomatic method

    Integrated Gradients proposed a gradient-based attribution method for deep networks. It made a model-specific route available for explaining differentiable models.

  16. 2017 Jun

    Transformer architecture is published

    Attention Is All You Need introduced the Transformer architecture. Its scaling path became the technical foundation for the large language models that later required distinct operational practices.

  18. 2017 Jul

    DARPA launches its Explainable AI program

    DARPA's program organized research around explainable models and human-computer interfaces for explanation. It highlighted the trade-off between learning performance and explanation.

  19. 2017 Sep

    ONNX is announced

    Microsoft and Facebook announced ONNX as a shared model representation. Model export could become a boundary between a training framework and a different inference runtime.

  20. 2017 Sep

    Uber publishes Michelangelo

    Uber described Michelangelo as an end-to-end platform for managing data, training, evaluation, deployment, prediction, and monitoring. It showed how a common lifecycle could support many production models.

  21. 2017 Dec

    ONNX version 1 is declared production-ready

    The first production-ready ONNX version expanded the interoperability promise to framework and hardware-runtime partners. Serving choices increasingly depended on artifact format as well as the training framework.

  22. 2018 Jan

    ULMFiT demonstrated task fine-tuning

    ULMFiT applied pretrained language models to downstream tasks through fine-tuning. It helped establish the pretrained-model-then-adapt workflow.

  23. 2018 Feb

    ULMFiT formalized task-specific language-model adaptation

    ULMFiT described pretraining a language model and then fine-tuning it for downstream tasks. It made transfer learning a practical pattern for NLP work.

  25. 2018 May 25

    GDPR becomes applicable

    The General Data Protection Regulation became applicable across the European Union. Its rules on automated individual decision-making made data protection and explanation part of many AI governance discussions.

  26. 2018 Jun

    GPT demonstrates generative pretraining plus task fine-tuning

    OpenAI presented a generative pretraining approach followed by task-specific fine-tuning. It established a prominent decoder-based path for adapting language models.

  27. 2018 Jun

    GPT-1 demonstrates transfer learning at scale

    OpenAI releases GPT-1, a 117-million-parameter decoder-only model pretrained on BookCorpus. It proves that a single pretrained transformer can be fine-tuned for diverse NLP tasks, establishing the pretrain-then-finetune paradigm.

  28. 2018 Jun

    MLflow is introduced

    MLflow introduced open-source components for tracking experiments, packaging models, and managing deployment. It established lifecycle concepts that GenAIOps later extends to prompts and evaluations.

  29. 2018 Jun

    OpenAI GPT demonstrates generative pre-training

    Generative Pre-Training uses a transformer language model followed by task-specific fine-tuning. It anchors the decoder-oriented path later used for autoregressive text generation.

  30. 2018 Jul

    The GDPR becomes applicable

    The General Data Protection Regulation became applicable across the European Union. Its provisions on automated decision-making made transparency and human safeguards an operational concern for affected decisions.

  31. 2018 Sep

    Google releases the What-If Tool

    The What-If Tool made visual model inspection and counterfactual-style exploration accessible in TensorBoard. It brought inspection of individual predictions and dataset slices into a developer workflow.

  32. 2018 Sep

    NVIDIA releases the TensorRT Inference Server beta

    NVIDIA introduced an open-source server, later named Triton, for TensorFlow, TensorRT, PyTorch, and ONNX models. GPU-aware batching and multi-framework deployment became a common runtime capability.

  33. 2018 Sep

    TensorFlow Data Validation launches

    TensorFlow Data Validation made data understanding, validation, and monitoring available as an open-source workflow. It focused attention on input-data health before downstream model-quality symptoms appear.

  34. 2018 Oct

    BERT established bidirectional contextual representations

    BERT demonstrated pre-training with both left and right context in every layer. Context became a central input to reusable language representations rather than only a task-specific prompt.

  35. 2018 Oct

    BERT established pretrained representations as a standard NLP starting point

    BERT showed that a deeply bidirectional pretrained model could be fine-tuned with a small task-specific output layer. Fine-tuning became a default evaluation and deployment path for many NLP tasks.

  36. 2018 Oct

    BERT establishes encoder-only pre-training

    BERT shows a bidirectional encoder-only transformer pre-trained with masked-language objectives. It makes the encoder-only family a standard choice for language-understanding tasks.

  37. 2018 Oct

    BERT popularizes bidirectional transformer pretraining

    BERT introduced pretraining with masked-language and next-sentence objectives for language understanding tasks. It broadened practical transformer use beyond left-to-right generation.

  38. 2018 Nov

    Feast released as an open-source feature store

    Google Cloud and GO-JEK announced Feast to manage, store, discover, and reuse features. Shared feature definitions addressed a recurring boundary between model development and production use.

  39. 2018 Nov

    Kubeflow Pipelines announced

    Google announced Kubeflow Pipelines for composing and managing end-to-end ML workflows. Pipeline components made repeatable workflow execution a visible platform capability.

  40. 2019 Jan

    Transformer-XL extends context across segments

    Transformer-XL adds segment-level recurrence and relative positional encoding. It addresses the fixed-length context and context-fragmentation limits of earlier transformer language models.

  41. 2019 Feb

    Adapter modules proposed for NLP transfer

    Adapter modules offered a way to keep most pretrained parameters fixed while training small task-specific components. This made serving multiple adapted tasks less dependent on full model copies.

  42. 2019 Feb

    GPT-2 tests larger-scale text generation

    GPT-2 reported stronger unsupervised multitask behavior from a larger transformer language model. It made fluent long-form generation a more visible operational concern.

  43. 2019 May

    OECD countries adopt the AI Principles

    OECD countries adopted the Recommendation on Artificial Intelligence. The principles provided an intergovernmental reference point for trustworthy-AI policy and governance work.

  45. 2019 Sep

    AI Explainability 360 documents a toolkit

    IBM researchers documented AI Explainability 360 as an open-source toolkit with multiple explanation algorithms and metrics. The toolkit helped turn a growing research catalog into reusable evaluation components.

  46. 2019 Sep

    NIST publishes the draft adversarial ML taxonomy

    NISTIR 8269 organized adversarial-machine-learning attacks, defenses, and consequences. It gave AI security work a shared risk vocabulary before generative AI application guidance emerged.

  47. 2019 Sep

    Transformers v2.0.0 published

    After open-sourcing the app's algorithm, Hugging Face pivoted to an ML platform; the transformers library's 2.0.0 release landed on PyPI in September 2019 and became the field's standard model library.

  48. 2019 Oct

    T5 frames tasks as text-to-text

    T5 studies transfer learning with a unified text-to-text framework built on an encoder-decoder transformer. It reinforces the family choice for tasks that transform a complete input into an output.

  49. 2019 Oct

    T5 unified NLP tasks under text-to-text transfer

    T5 framed every NLP task as text-to-text and compared transfer-learning objectives at scale. This strengthened the idea that one pretrained model could be adapted through task examples.

  50. 2019 Dec

    Amazon SageMaker Model Monitor launches

    Amazon introduced managed monitoring for production models, including comparison with training or validation data and scheduled checks. Drift monitoring became a built-in cloud deployment capability.

2020s

149 milestones

  1. 2020 Jan

    Reformer targets attention-memory limits

    Reformer introduces locality-sensitive hashing attention and reversible layers to reduce the cost of long sequences. It makes attention's quadratic memory and computation cost an explicit architecture constraint.

  2. 2020 Feb

    REALM introduces retrieval-augmented language-model pretraining

    REALM combined language-model pretraining with retrieval from a large corpus. It established retrieval as a component that can supply knowledge beyond model parameters.

  3. 2020 Feb 19

    European Commission publishes its AI White Paper

    The White Paper set out a European approach to excellence and trust in AI. It put options for a legislative framework for trustworthy AI into the policy path that led to the AI Act.

  4. 2020 Mar

    Kubeflow 1.0 released

    Kubeflow 1.0 established a Kubernetes-native platform for developing, orchestrating, deploying, and running portable ML workloads. Production ML pipelines could now build on a more mature shared base.

  5. 2020 Apr

    Longformer targeted long-document attention

    Longformer combined local-window and task-motivated global attention to process much longer documents with linear scaling. It made attention patterns and document length explicit design choices for long inputs.

  6. 2020 Apr

    TorchServe is announced

    PyTorch and AWS introduced a model server focused on packaging and scaling PyTorch models. The framework-specific serving path became an explicit alternative to converting artifacts for a general runtime.

  7. 2020 May

    GPT-3 demonstrated few-shot prompting

    GPT-3 showed that a large language model could perform many tasks from examples supplied in its prompt. Few-shot learning became a practical runtime adaptation option.

  8. 2020 May

    GPT-3 demonstrates few-shot language tasks

    GPT-3 documented strong few-shot performance from a language model with 175 billion parameters. Operating applications increasingly involved prompt behavior and external model APIs, not only trained model artifacts.

  9. 2020 May

    GPT-3 demonstrates few-shot prompting at scale

    GPT-3 evaluated a 175-billion-parameter autoregressive language model and described few-shot task conditioning in the prompt. Prompt content became a practical way to steer behavior without updating weights.

  10. 2020 May

    GPT-3 demonstrates few-shot text interaction

    The GPT-3 paper evaluates tasks specified with text prompts and demonstrations without gradient updates. Few-shot examples became a practical way to adapt one model to many tasks.

  11. 2020 May

    GPT-3 demonstrates large-scale few-shot language modeling

    GPT-3 studies a large autoregressive language model and evaluates it with few-shot prompting. It makes decoder-only scaling a major practical direction for general text generation.

  12. 2020 May

    GPT-3 scales to 175 billion parameters

    OpenAI's GPT-3 demonstrates that scaling language models to hundreds of billions of parameters produces strong few-shot performance without task-specific fine-tuning, sparking the modern race toward larger foundation models.

  13. 2020 May

    RAG formalizes retrieve-then-generate models

    Lewis and colleagues described generation conditioned on retrieved documents from a dense vector index. The work gave the pattern its RAG name and framed parametric and non-parametric memory together.

  14. 2020 May

    Retrieval-Augmented Generation formalized

    RAG combined a parametric generator with retrieved material from external memory. It made external, updateable knowledge a distinct design path from changing model weights.

  15. 2020 May

    Retrieval-augmented generation paired models with indexes

    RAG combined a pre-trained generator with a dense vector index as non-parametric memory. It established retrieval selection as part of the generation pipeline rather than a separate search feature.

  16. 2020 Jun

    DDPM brings diffusion to images

    Ho, Jain and Abbeel's Denoising Diffusion Probabilistic Models achieved high-quality image synthesis with a denoising objective, reaching state-of-the-art FID on CIFAR-10 and making diffusion a serious image generator.

  17. 2020 Jun

    Google Cloud describes an MLOps foundation

    Google Cloud published requirements for an MLOps foundation that connects data, automation, deployment, monitoring, and governance. The framing made MLOps an explicit architecture concern for cloud ML.

  18. 2020 Sep

    datasets library 1.0.0

    The datasets library reached 1.0.0, giving the ecosystem one-line, Arrow-backed access to NLP and multimodal datasets and completing the core training stack.

  20. 2020 Oct

    Vision Transformer applies the architecture to image patches

    Vision Transformer treats fixed-size image patches as tokens and applies a pure transformer to image classification. It demonstrates that the sequence representation, rather than text itself, is the architectural requirement.

  21. 2020 Nov

    Dense Passage Retrieval demonstrates dual-encoder passage search

    DPR showed that question and passage encoders can support dense retrieval for open-domain question answering. It helped make embedding-based retrieval a practical default component.

  22. 2020 Nov

    Evidently releases its first open-source version

    Evidently released an open-source tool for monitoring models and data. Teams could build repeatable drift and quality reports into Python-based workflows.

  23. 2020 Nov

    Score-based SDE framework

    Song et al. unified score-based and diffusion models under stochastic differential equations, adding a predictor-corrector sampler and generating 1024x1024 images from a score-based model for the first time.

  24. 2020 Dec

    Amazon SageMaker Pipelines becomes generally available

    AWS introduced SageMaker Pipelines for building, automating, and scaling end-to-end ML workflows with lineage and model registry integration. Managed CI/CD concepts became directly available in an ML platform.

  25. 2021

    Triton becomes NVIDIA Triton Inference Server

    NVIDIA standardized the Triton name for its inference server as the project expanded beyond its original TensorRT association. The naming reflected a broader multi-framework serving role.

  26. 2021 Jan

    Prefix-tuning kept model weights frozen

    Prefix-tuning optimized a small continuous prefix while leaving the language model unchanged. It expanded the range of adaptation methods between prompting and full fine-tuning.

  27. 2021 Jan

    Switch Transformer scales to trillions of parameters

    Google's Switch Transformer applies mixture-of-experts to transformers, reaching 1.6 trillion parameters while keeping per-token compute fixed. It demonstrates that conditional computation can decouple model capacity from inference cost.

  28. 2021 Jan

    Switch Transformers scales sparse expert routing

    Switch Transformers routes tokens through sparse experts while retaining a transformer backbone. It makes model capacity and per-token computation separable design concerns.

  29. 2021 Apr

    BigScience launched

    Hugging Face kicked off the BigScience research workshop with international research groups to train an open large language model, a deliberate bet on open collaboration at frontier scale.

  32. 2021 Apr

    Prompt tuning formalizes learned prompts

    Prompt tuning introduced soft prompts learned through backpropagation instead of only hand-written text. It separated discrete instruction writing from a parameter-efficient adaptation method.

  33. 2021 May

    Diffusion models beat GANs

    Dhariwal and Nichol showed diffusion surpassing GAN image quality using classifier guidance, reaching FID 2.97 on ImageNet 128x128. The public turning point that displaced GANs as the default.

  34. 2021 May

    Vertex AI becomes generally available

    Google Cloud released Vertex AI as a unified managed platform for building, training, deploying, and maintaining ML models. Its MLOps components brought pipelines, metadata, and monitoring together.

  35. 2021 May

    Vertex AI introduced at Google I/O

    Google announced Vertex AI as a unified platform bringing AutoML and custom training under one API and console, replacing the separate AI Platform surfaces and promising one workflow from data to deployment.

  36. 2021 Jun

    Documentation live under the Vertex AI name

    The Vertex AI documentation site was already serving by mid-June 2021, anchoring the platform identity that practitioner tooling, the gcloud ai group, and the aiplatform SDK would carry for years.

  37. 2021 Jun

    LoRA introduced low-rank adaptation

    LoRA froze pretrained weights and represented updates with trainable low-rank matrices. It made adapter-based fine-tuning a practical alternative to updating every parameter.

  38. 2021 Jun

    LoRA introduced low-rank updates

    LoRA trained low-rank matrices while freezing pretrained weights. It became a widely used approach for reducing trainable parameters during task adaptation.

  39. 2021 Jun

    MITRE names and releases ATLAS version 2.0

    MITRE renamed the AdvML Threat Matrix as ATLAS and released version 2.0 with partners. Practitioners gained an ATT&CK-style knowledge base for threats against AI systems.

  40. 2021 Jul

    NIST requests input for an AI risk-management framework

    NIST opened the AI RMF request for information. The work moved AI risk management toward an organization-wide lifecycle rather than isolated model testing.

  41. 2021 Jul

    Vertex AI monitoring for training-serving skew

    Google Cloud documented monitoring for skew between training data and served predictions. Baselines, thresholds, alerts, and diagnosis became part of a managed MLOps path.

  42. 2021 Sep

    NIST publishes four principles of explainable AI

    NISTIR 8312 set out Explanation, Meaningful, Explanation Accuracy, and Knowledge Limits. It separated the quality of an explanation from the accuracy of the model decision.

  43. 2021 Sep

    Pipelines reach general availability

    Vertex AI Pipelines became generally available, running Kubeflow Pipelines and TFX workflows serverless with metadata and lineage tracking managed by the platform.

  44. 2021 Dec

    Latent diffusion and GLIDE

    Two December papers reshaped practice. Rombach et al. moved diffusion into autoencoder latent space, enabling consumer-GPU generation, and OpenAI's GLIDE demonstrated text-conditional diffusion with classifier-free guidance preferred by evaluators.

  45. 2021 Dec

    WebGPT demonstrated browser use

    WebGPT fine-tuned GPT-3 to use a text-based web browser. The prototype submitted search queries, followed links, and cited sources in its answers.

  46. 2021 Dec

    WebGPT demonstrated browser-mediated answering

    WebGPT paired a language model with a text browser that could search, follow links, and quote sources. The prototype made observation of an external environment part of the answer-generation process.

  47. 2022

    Text Generation Inference begins public releases

    Hugging Face published Text Generation Inference as a server for transformer generation workloads. Token streaming and generation-specific batching moved into dedicated serving infrastructure.

  48. 2022 Jan

    Chain-of-thought prompting is published

    The paper showed few-shot reasoning demonstrations can improve performance on multi-step reasoning tasks. Prompt examples began to carry intermediate reasoning structure as well as final-answer format.

  49. 2022 Jan

    InstructGPT publishes the RLHF alignment recipe

    OpenAI's InstructGPT paper demonstrates that a three-stage process — supervised fine-tuning, reward model training, and reinforcement learning — turns a base model into an assistant aligned with human preferences, with a 1.3B model preferred over the 175B base GPT-3.

  50. 2022 Jan

    InstructGPT research emphasizes instruction following

    OpenAI reported instruction-following models trained with human feedback. The work made the quality of an instruction a more central interface for application behavior.

  51. 2022 Feb

    BLOOM model released

    BigScience concluded with BLOOM, a 176-billion-parameter multilingual open language model, demonstrating that openly developed frontier models were feasible.

  52. 2022 Mar

    Chinchilla redefines compute-optimal training

    DeepMind's scaling law study trains over 400 models to show that model size and training data should scale together. Chinchilla, a 70B model trained on 1.4T tokens, outperforms the 280B Gopher at the same compute budget.

  53. 2022 Mar

    Chinchilla revisits compute-optimal model and data size

    Chinchilla studies the balance between parameter count and training tokens under a fixed compute budget. It changes how practitioners reason about scaling a transformer beyond adding parameters.

  54. 2022 Mar

    InstructGPT documented preference-based alignment after supervised fine-tuning

    InstructGPT combined demonstration data, ranked outputs, and reinforcement learning from human feedback. It made post-training behavior and human preference central to instruction-following models.

  55. 2022 Mar

    InstructGPT reports instruction-following with human feedback

    The InstructGPT paper described supervised fine-tuning and reinforcement learning from human feedback for aligning model output with user intent. Post-training became a distinct part of the model lifecycle.

  57. 2022 May

    Imagen

    Google's Imagen showed that large text-only language models are effective text encoders for image synthesis, achieving state-of-the-art COCO FID and introducing DrawBench.

  59. 2022 Jul

    Classifier-free guidance formalized

    Ho and Salimans published the guidance method that joint-trains conditional and unconditional models. Every guidance_scale control in modern tools descends from this paper.

  60. 2022 Jul

    ColBERTv2 improves late-interaction retrieval efficiency

    ColBERTv2 paired token-level late interaction with residual compression. It made a higher-resolution alternative to single-vector document retrieval more practical to evaluate.

  62. 2022 Sep

    Safetensors introduced

    The safetensors format arrived as a safe tensor serialization replacing pickle, whose code-execution-on-load behavior had made weight sharing a security risk; it later became the default in transformers.

  63. 2022 Oct

    Inference Endpoints launched

    Hugging Face announced Inference Endpoints, deploying Hub models to managed cloud infrastructure in a few clicks, the start of the platform's paid serving layer.

  64. 2022 Oct

    ReAct combined reasoning with actions

    The ReAct paper presented a method that interleaves reasoning traces with task-specific actions. It connected planning and external information gathering in one loop.

  65. 2022 Oct

    ReAct interleaved reasoning and actions

    ReAct described alternating reasoning traces with task actions and observations from external environments. It supplied a clear pattern for the observe-reason-act cycle.

  66. 2022 Nov

    ChatGPT introduces a conversational prompt interface

    ChatGPT launched as a dialogue model trained to follow instructions and provide detailed responses. Conversational turns made role-separated instructions familiar to a broad audience.

  67. 2022 Nov

    ChatGPT is introduced as a research preview

    OpenAI introduced ChatGPT and gathered feedback on its strengths and weaknesses. Public conversational deployment made output behavior, safety, and feedback collection operational concerns.

  68. 2022 Nov

    ChatGPT released as a research preview

    OpenAI released ChatGPT as a conversational research preview. The product exposed interactive language-model behavior to a broad public audience.

  70. 2023 Jan

    NIST releases AI RMF 1.0

    AI RMF 1.0 introduced the Govern, Map, Measure, and Manage functions. Those functions provide a lifecycle structure for assigning and testing AI security controls.

  71. 2023 Jan

    PEFT library first release

    The peft library shipped, making parameter-efficient fine-tuning (LoRA and related adapters) a standard part of the ecosystem's adaptation story.

  72. 2023 Jan 26

    NIST releases AI RMF 1.0

    NIST released the voluntary AI Risk Management Framework 1.0. Its Govern, Map, Measure, and Manage functions gave organizations a repeatable structure for lifecycle risk work.

  73. 2023 Feb

    AWS partnership announced

    AWS and Hugging Face announced an expanded collaboration to make training and deployment of large models more accessible via SageMaker, Trainium, and Inferentia, cementing Hub models as cloud platform residents.

  74. 2023 Feb

    ControlNet

    Zhang, Rao and Agrawala added spatial conditioning (edges, depth, pose) to frozen text-to-image models via zero convolutions, making structure-locked generation practical on personal hardware.

  75. 2023 Feb

    FlashAttention makes long-context inference practical

    Dao et al. publish FlashAttention, an IO-aware exact attention algorithm that reduces memory reads and writes by restructuring the computation. It becomes the standard optimization for serving long-context models, with v2 following in July 2023.

  76. 2023 Feb

    LLaMA expands access to open language-model research

    Meta released the LLaMA paper describing a collection of foundation language models. It increased attention to running and adapting model weights outside a single hosted provider.

  77. 2023 Feb

    Model Monitoring and pipeline schedules expand

    Google expanded the MLOps surface with scheduled pipeline runs, model monitoring improvements, and feature store capabilities, rounding out the platform story beyond single train-and-deploy cycles.

  78. 2023 Feb

    Toolformer studied learned API-call decisions

    Toolformer trained a language model to select APIs, form arguments, and incorporate returned results. It made tool selection and result incorporation explicit parts of language-model behavior.

  79. 2023 Mar

    Chat Completions promotes structured messages

    OpenAI described the chat-based interface as a structured prompt format with system messages and multi-turn context. Application prompts increasingly separated instructions from user-provided data.

  80. 2023 Mar

    GPT-4 and OpenAI Evals are announced

    OpenAI announced GPT-4 and open-sourced Evals for creating and running benchmarks. Evaluation became a visible deployment control for comparing model versions and application integrations.

  81. 2023 Mar

    GPT-4 technical report documents a multimodal large model

    The GPT-4 technical report described a large multimodal model that accepts image and text inputs and produces text outputs. Multimodal context became part of the mainstream LLM application model.

  82. 2023 Mar

    LLaMA open-weights models lower the barrier to entry

    Meta releases LLaMA, a collection of open-weight foundation models from 7B to 65B parameters. The release triggers an explosion of community fine-tunes and demonstrates that competitive models can be trained on public data.

  83. 2023 Mar

    Reflexion introduced verbal feedback memory

    Reflexion stored linguistic feedback from trials in an episodic memory buffer for later decisions. The work gave a concrete loop pattern for reflection after failure.

  84. 2023 Apr

    Bedrock announced in preview

    AWS announced Amazon Bedrock as a managed service for accessing foundation models from AI21 Labs, Anthropic, Stability AI, and Amazon, alongside Titan image and text models, positioned as the serverless alternative to running model infrastructure.

  85. 2023 May

    Direct Preference Optimization simplified preference training

    DPO expressed preference optimization with a classification-style loss instead of a separate reward-model training stage. It created a widely used alternative when preferred and rejected responses describe the target behavior.

  86. 2023 May

    OWASP releases version 0.5 of its LLM Top 10

    OWASP published a major early milestone for its Top 10 for LLM Applications. It focused application-security attention on risks specific to LLM-integrated systems.

  87. 2023 May

    QLoRA combined quantization with LoRA

    QLoRA backpropagated through a frozen 4-bit quantized model into LoRA adapters. It further reduced memory requirements for adapting large language models.

  88. 2023 May

    QLoRA combined quantization with low-rank adapters

    QLoRA trained LoRA adapters while keeping the pretrained model quantized. It reduced memory requirements for adapting larger models without full fine-tuning.

  89. 2023 May

    Tree of Thoughts extends reasoning-path search

    Tree of Thoughts proposed exploring and evaluating multiple intermediate text paths for problems needing planning or search. It broadened prompting from one linear response to controlled exploration.

  90. 2023 Jun

    Function calling added to OpenAI API models

    OpenAI announced function-calling capabilities for GPT-4 and GPT-3.5 Turbo. Models could return structured function arguments for application-side tool execution.

  91. 2023 Jun

    Function calling connects prompts to typed actions

    OpenAI added function calling so models could return JSON arguments matching described functions. Prompt design expanded from prose responses to tool selection and structured arguments.

  92. 2023 Jun

    Generative AI support on Vertex AI

    Google brought foundation models (PaLM 2, Codey, Imagen) to Vertex AI with a Model Garden, prompt design, tuning, and deployment surface, marking the platform's pivot toward generative workloads.

  94. 2023 Jun

    OpenAI introduced API function calling

    OpenAI announced function-calling capabilities that let models produce structured calls from user requests. This created a common application interface between a model decision and an executable action.

  95. 2023 Jul

    HyDE applies generated hypothetical documents to zero-shot retrieval

    HyDE generated a hypothetical document from a query, embedded it, and used that representation to retrieve real documents. It broadened query-transformation techniques when relevance labels are unavailable.

  96. 2023 Jul

    LangSmith launches in closed beta

    LangChain announced LangSmith for debugging, testing, evaluating, and monitoring LLM applications. The launch reflected the need for application-level observability beyond model availability metrics.

  97. 2023 Jul

    Llama 2 ships with commercial licensing

    Meta's Llama 2 releases with weights available for commercial use, validating the open-weight model ecosystem. It introduces RLHF alignment for the open-weight lineage and trains on 40% more tokens than LLaMA 1.

  98. 2023 Jul

    SDXL

    Stability's SDXL tripled the UNet size, added a second text encoder, trained on multiple aspect ratios, and introduced a refiner model, raising the open-model default to 1024x1024.

  99. 2023 Aug

    AutoGen described multi-agent conversation

    The AutoGen paper described an open-source framework in which multiple LLM agents converse to accomplish tasks. It made agent-to-agent conversation a programmable application pattern.

  100. 2023 Aug

    AutoGen paper described multi-agent conversation

    The AutoGen paper presented agents that communicate through automated conversation and can combine language models, tools, and human feedback. It formalized one route from a single loop to coordinated agent loops.

  101. 2023 Aug

    Series D at a $4.5B valuation

    Hugging Face raised $235 million from investors including Google, Amazon, Nvidia, and Salesforce, valuing the company at $4.5 billion with over 50,000 organizations on the platform.

  102. 2023 Sep

    General availability

    Bedrock became generally available, adding agents for multi-step tasks and bringing Meta Llama 2 into the model catalog. GA marked the shift from preview experiment to a platform with a support surface.

  103. 2023 Sep

    Lost in the Middle measured positional degradation

    Lost in the Middle evaluated language-model access to relevant information at different locations in long inputs. It made placement within a context window an observable design variable.

  105. 2023 Sep

    vLLM introduces PagedAttention

    The vLLM paper framed key-value cache allocation as the limiting resource for large-language-model serving. Memory management and continuous request scheduling became central to throughput decisions for generative endpoints.

  106. 2023 Oct

    OWASP publishes LLM Top 10 version 1.1

    The first mature 2023 list documented risks and mitigations for the development, deployment, and management lifecycle of LLM applications.

  107. 2023 Oct

    Self-RAG adds retrieval and critique decisions

    Self-RAG proposed on-demand retrieval and reflection tokens for judging relevance and support during generation. It shifted attention from a fixed top-k pipeline toward conditional retrieval behavior.

  108. 2023 Nov

    Feature Store relaunch on BigQuery

    Google relaunched Vertex AI Feature Store as a managed service that manages feature data directly in BigQuery tables and serves online from them, replacing the legacy feature store architecture.

  109. 2023 Nov

    Guardrails announced

    AWS introduced Guardrails for Amazon Bedrock in preview, configurable safeguards that filter undesirable content from user inputs and model responses independently of the chosen model.

  110. 2023 Nov

    Model Registry and Gen AI evaluation consolidated

    Model Registry became the central repository for custom and BigQuery ML models, with aliases and version management; generative AI evaluation arrived as pipeline-based side-by-side comparison.

  111. 2023 Dec

    EU co-legislators reach political agreement on the AI Act

    The European Parliament and Council reached political agreement on the AI Act. Organizations could begin turning the emerging requirements into inventories, controls, and evidence plans before final adoption.

  112. 2023 Dec

    Gemini arrives on Vertex AI

    Google launched Gemini 1.0 with Vertex AI as the developer platform for the model family, adding Gemini Pro and Gemini Ultra to the Model Garden and making Vertex the enterprise access route to Google's frontier models.

  113. 2023 Dec

    Knowledge Bases launched

    Amazon Bedrock Knowledge Bases reached general availability, adding managed retrieval augmented generation with connectors to Amazon S3 and other data sources.

  114. 2024 Jan

    CRAG evaluates retrieved evidence before generation

    Corrective RAG introduced a retrieval evaluator that selects actions based on the quality of retrieved documents. It made bad retrieval an explicit control point rather than an invisible generator failure.

  115. 2024 Jan

    Mixtral 8x7B popularizes sparse mixture-of-experts

    Mistral AI releases Mixtral, a 46.7B-parameter mixture-of-experts model using 12.9B active parameters per token. It matches or exceeds GPT-3.5 on many benchmarks, proving that MoE architectures can deliver frontier-level quality at far lower inference cost.

  116. 2024 Jan

    RAPTOR introduces hierarchical retrieval over summaries

    RAPTOR recursively clusters and summarizes chunks into a retrieval tree. The approach targets questions that need information across a long document rather than one contiguous passage.

  117. 2024 Feb

    Agents reach general availability

    Amazon Bedrock Agents became generally available, letting a foundation model orchestrate multi-step tasks with API calls and knowledge base access under AWS-managed orchestration.

  118. 2024 Mar

    Guardrails general availability

    Guardrails for Amazon Bedrock became generally available, extending filtering to personal information protection, denied topics, and word filters alongside the original content filters.

  119. 2024 Mar

    Stable Diffusion 3 and flow matching

    Stability announced SD3 with a Multimodal Diffusion Transformer using separate image and language weight sets and flow matching, improving text rendering and prompt following at up to 8B parameters.

  120. 2024 Mar 13

    European Parliament adopts the AI Act

    Parliament adopted the Artificial Intelligence Act. The legislative text established the risk-based framework and obligations that governance programs would need to operationalize.

  121. 2024 Apr

    Dynamic Shared Quota for Gemini

    Google moved Gemini API capacity to a Dynamic Shared Quota model without fixed per-project numbers, with Provisioned Throughput as the paid guarantee, changing how teams plan and debug throttling.

  122. 2024 Apr

    International agencies publish secure AI deployment guidance

    NSA, CISA, and partner agencies published deployment practices for securing and operating AI systems. The guidance connected AI-specific risks to setup, infrastructure, monitoring, and maintenance.

  123. 2024 Apr

    Llama 3 advances open-weight quality

    Meta releases Llama 3 in 8B and 70B variants, trained on over 15 trillion tokens. The models establish a new quality bar for open-weight systems and demonstrate that continued scaling of training data remains a reliable path to better performance.

  126. 2024 Jul

    Cross-Region inference introduced

    AWS announced cross-Region inference, letting applications call an inference profile that routes requests across compatible Regions to pool capacity during high-demand periods.

  127. 2024 Jul

    Llama 3.1 releases larger open-weight models

    Meta described Llama 3.1 as a collection of multilingual large language models with 8B, 70B, and 405B parameter variants. Open-weight models became a more practical deployment option for teams able to operate inference.

  128. 2024 Jul

    Microsoft releases GraphRAG

    Microsoft released GraphRAG as a graph-based RAG research system for forming targeted context from private text. It brought graph-derived context into a widely used RAG implementation.

  129. 2024 Jul

    NIST publishes the Generative AI Profile

    NIST published AI 600-1, a companion profile for generative AI under the AI Risk Management Framework. Risk management guidance gained a specific reference for generative AI systems.

  130. 2024 Jul

    NIST releases the Generative AI Profile

    NIST AI 600-1 applied the AI RMF to generative AI risks and suggested actions. It made prompt injection and related generative-AI risks explicit within the framework's lifecycle.

  132. 2024 Aug

    Structured Outputs constrains JSON schema conformance

    OpenAI introduced Structured Outputs for responses constrained by developer-supplied JSON Schemas. For supported output contracts, prompt instructions no longer had to carry the full burden of format reliability.

  133. 2024 Aug

    The EU AI Act enters into force

    Regulation (EU) 2024/1689 entered into force and established a risk-based framework for AI systems. Its transparency and information requirements made explainability relevant to high-risk-system operations.

  135. 2024 Sep

    Contextual Retrieval added document context to chunks

    Anthropic described adding concise document-specific context before embedding and indexing chunks. The approach addressed the loss of meaning that can occur when retrieval splits documents.

  136. 2024 Nov

    Anthropic introduces the Model Context Protocol

    Anthropic introduced the Model Context Protocol for connecting AI applications to external systems. Tool and data connections became an explicit interface to govern, test, and observe in AI systems.

  137. 2024 Nov

    Model Context Protocol open-sourced

    Anthropic open-sourced the Model Context Protocol as a standard for connecting AI assistants to data systems and tools. The protocol defined client and server roles for those connections.

  138. 2024 Nov

    Model Context Protocol was introduced

    Anthropic introduced the Model Context Protocol as an open protocol for connecting AI assistants to external systems and data. It standardized a layer through which an agent loop can discover and use tools.

  139. 2024 Dec

    Batch inference and prompt management

    Amazon Bedrock added batch inference for asynchronous queued workloads and prompt management, rounding out the cost-control and reuse levers around on-demand token pricing.

  140. 2024 Dec

    smolagents launched

    Hugging Face released smolagents, a minimal agent library whose agents write their actions as code, marking the ecosystem's move into agentic tooling and succeeding transformers.agents.

  141. 2025

    mtRAG benchmarks multi-turn RAG conversations

    mtRAG introduced a human-generated benchmark for evaluating RAG across multi-turn conversations. It made later turns, unanswerable questions, and non-standalone questions explicit evaluation cases.

  142. 2025 Feb 2

    Prohibited-practices and AI-literacy obligations begin to apply

    The first AI Act application date brought prohibited AI practices and AI-literacy obligations into application. Governance teams needed an inventory capable of identifying affected uses.

  143. 2025 Apr

    Agent2Agent Protocol announced

    Google announced the Agent2Agent Protocol for interoperability between agents across systems. It addressed collaboration in multi-agent ecosystems built by different vendors or frameworks.

  144. 2025 May

    International agencies publish AI data-security guidance

    NSA, CISA, and partners published guidance for data security across the AI lifecycle. It reinforced provenance, integrity, access control, validation, auditing, and deletion for AI data.

  145. 2025 Jun

    Generative SDK modules deprecated

    Google deprecated the generative AI modules in the google-cloud-aiplatform SDK (generative_models, language_models, and related), directing gen AI work to the separate google-genai package and beginning the SDK surface split.

  146. 2025 Jul

    AgentCore announced

    AWS announced Amazon Bedrock AgentCore, a platform for deploying and operating production agents with any framework, with sessions, memory, observability, and connectors, marking the shift from single agent features to an agent operations platform.

  147. 2025 Aug 2

    GPAI-model obligations begin to apply

    Governance rules and obligations for general-purpose AI models became applicable. Organizations using or providing these models needed evidence appropriate to their role in the value chain.

  148. 2025 Oct

    AgentCore general availability

    Amazon Bedrock AgentCore reached general availability, and the original Bedrock Agents service began its transition to maintenance mode as new agentic investment concentrated in AgentCore.

  149. 2026 Apr

    Gemini Enterprise Agent Platform announced

    Google announced the Agent Platform as the evolution of Vertex AI, stating all Vertex AI services and roadmap evolution would be delivered through the Agent Platform rather than as a standalone service, and renaming the product surface accordingly.