Industry History

Review the product, platform, and capability shifts that shaped today's AI industry.

Industry history

Major Stages

2022–2023

Foundation Models Become Products

Chat interfaces, model APIs, open weights, and copilots turned frontier research into mass-market and enterprise software.

Dec 29, 2023

Shape-IoU: More Accurate Metric Considering Bounding Box Shape and Scale

This paper proposes the Shape-IoU method, which analyzes the impact of bounding box shape and scale on regression results, designs a new loss function, and achieves state-of-the-art performance on multiple detection tasks, surpassing existing methods.

Dec 19, 2023

Gemini: A Family of Highly Multimodal Models

Google released the Gemini family of multimodal models, including three sizes: Ultra, Pro, and Nano. It achieved state-of-the-art results in 30 out of 32 benchmarks, surpassed human expert performance on MMLU for the first time, and achieved the best results on all 20 multimodal benchmarks.

Dec 18, 2023

Retrieval-Augmented Generation (RAG) Survey: From Naive RAG to Modular RAG Evolution and Evaluation: Key Technology for Knowledge Enhancement in Large Language Models

This survey systematically reviews the evolution of retrieval-augmented generation (RAG) from naive RAG, advanced RAG to modular RAG, analyzes in detail the technical details of the three core components of retrieval, generation, and augmentation, and introduces the latest evaluation frameworks and benchmarks. By introducing external knowledge bases, RAG effectively alleviates issues such as hallucinations, outdated knowledge, and opaque reasoning in large language models, improving the accuracy and credibility of knowledge-intensive tasks.

Dec 6, 2023

Gemini 1.0 Released: Native Multimodality Becomes the Standard for Frontier Model Platforms

Google released three sizes of native multimodal models—Gemini Ultra, Pro, and Nano—in December 2023.

2024

Multimodality and Inference-Time Reasoning

Native voice, image, video, long context, and reasoning-time compute widened the capability curve while governance moved into deployment decisions.

Dec 26, 2024

MEDEC: Benchmark for Medical Error Detection and Correction in Clinical Notes, LLMs Still Lag Behind Doctors

In December 2024, Microsoft Research and others released the MEDEC benchmark, comprising 488 clinical notes (3,848 sentences) from three US hospitals, covering five error types: diagnosis, management, treatment, medication, and causative microorganism. Evaluated models include o1-preview, GPT-4, Claude 3.5 Sonnet, and Gemini 2.0 Flash. Results show LLMs underperform human doctors in both error detection and correction.

Dec 26, 2024

DeepSeek-V3 Open Source: Training Efficiency Becomes a New Variable in Global Model Competition

DeepSeek released and open-sourced DeepSeek-V3 weights and technical report, featuring 671B MoE, 37B activated parameters, and FP8 training.

Dec 18, 2024

ModernBERT: Pareto Improvement for Encoder Models, Boosting Inference Efficiency and Performance

In December 2024, Answer.AI and LightOn jointly released ModernBERT, trained on 2 trillion tokens with native support for 8192 sequence length. Compared to original BERT, ModernBERT achieves SOTA on classification and retrieval tasks, with 4x faster inference and 50% less memory usage. The model incorporates modern optimizations such as rotary position encoding, GeGLU activation, alternating attention, and supports Flash Attention.

Dec 13, 2024

DeepSeek-VL2: MoE Multimodal Model with High-Resolution Dynamic Vision and Efficient Inference

In December 2024, DeepSeek released the DeepSeek-VL2 series, adopting a MoE architecture with activated parameters of 1B/2.8B/4.5B. It introduces a dynamic tiling visual encoding strategy to support high-resolution images with various aspect ratios. The language part uses DeepSeekMoE and Multi-Head Latent Attention (MLA) to compress KV cache. It achieves or surpasses open-source models of similar scale on tasks such as VQA, OCR, document/table/chart understanding, and visual grounding.

2025 H1

Agents Enter Real Workflows

Coding, research, browsing, and tool use moved from demonstrations toward bounded tasks with observable outputs and review loops.

Jun 30, 2025

UMA: A Family of Universal Models for Atoms: Meta Open-Sources Universal Atomic Simulation Model to Accelerate Materials and Chemistry Calculations

In June 2025, Meta FAIR released the UMA model family, trained on 500 million 3D atomic structures, using a mixture of linear experts architecture. The small/medium models have 1.4B parameters but only about 50M active parameters. Without fine-tuning, they achieve or exceed the performance of specialized models in multiple domains including molecules, materials, and catalysts. Code, weights, and data have been open-sourced.

Jun 30, 2025

ERNIE 4.5 Full Family Open-Sourced: Baidu Uses Heterogeneous MoE to Complete Multimodal Open Route

Baidu open-sourced 10 model variants of ERNIE 4.5 in June 2025 under the Apache 2.0 license.

Jun 27, 2025

MAI-DxO: Sequential Diagnosis with Language Models: AI Diagnostic System Achieves 80% Accuracy, Reduces Costs by 20%

In June 2025, Microsoft team proposed MAI-DxO diagnostic orchestrator, building a sequential diagnosis benchmark based on 304 NEJM challenging cases. MAI-DxO combined with o3 model achieves 80% diagnostic accuracy (general physicians 20%), with a 20% cost reduction. When configured for maximum accuracy, it reaches 85.5%, and generalizes to models such as Gemini, Claude, DeepSeek.

Jun 17, 2025

Skala: Accurate and scalable exchange-correlation with deep learning: Deep learning breaks the DFT accuracy-efficiency trade-off, ushering a new paradigm in computational chemistry

In June 2025, Meta FAIR released Skala, a deep learning-based exchange-correlation functional, achieving an error of 2.8 kcal/mol on the GMTKN55 benchmark, surpassing state-of-the-art hybrid functionals while maintaining the low computational cost of semilocal DFT. Training data comes from high-accuracy wavefunction methods, enabling systematic improvability of deep learning functionals for the first time.

2025 H2

Protocols and Product Surfaces Converge

Agent protocols, model-native tools, and application runtimes competed to own context, permissions, and complete workflows.

2026–Today

Reliability, Economics, and Long-Horizon Work

The market is testing whether agents can sustain longer tasks, recover from failure, and deliver measurable value at a defensible total cost.

Jul 9, 2026

GPT-5.6 Release: OpenAI Pushes Model Upgrades Toward Long-Term Autonomous Agents

OpenAI releases GPT-5.6 and simultaneously launches ChatGPT Work, targeting cross-app, file, and long-term tasks.

Jul 12, 2026

OpenAI's GPT-5.6 Sol Ultra reportedly solves a 50-year-old math problem in under an hour

OpenAI's GPT-5.6 Sol Ultra produced a proof of the Cycle Double Cover Conjecture in under an hour, using 64 subagents working in parallel. The conjecture had remained unsolved for 50 years. Mathematician Thomas Bloom calls the proof surprisingly elementary but criticizes the lack of citations for known prior work. The bigger question remains: Does AI just recombine existing knowledge, or does it create something new? The article OpenAI's GPT-5.6 Sol Ultra reportedly solves a 50-year-old math problem in under an hour appeared first on The Decoder.

Jul 17, 2026

SciDiagramEdit: Learning to Edit Scientific Diagrams from Paper Revisions

SciDiagramEdit is a benchmark and skill evolution framework that mines before-and-after diagram pairs from arXiv paper version history and edits scientific diagrams based on natural language revision intents.

Jul 17, 2026

Pretraining Data Can Be Poisoned through Computational Propaganda

This paper proposes a method to poison language model pretraining data by injecting malicious content into web crawler data through public discussion interfaces (e.g., forums, comment sections). It also introduces the HalfLife analysis tool to estimate the retention rate of adversarial content in web-crawled data.