DeepSeekMath introduces the GRPO method later used by R1
DeepSeek releases math models and Group Relative Policy Optimization, creating a direct algorithmic predecessor to R1's reinforcement-learning recipe.
EVENT ARCHIVE
Events
English reports are listed here. AI-generated drafts are clearly labeled until an editor reviews them.
AI-generated · reviewedImpact 92/100
DeepSeek publishes the sparse architecture behind its later models
DeepSeekMoE uses fine-grained experts and shared expert isolation to improve sparse-model efficiency, with 16B weights and training code released.
Google releases the Gemini 1.0 family
Google introduces the natively multimodal Gemini 1.0 family—Ultra, Pro and Nano—with Pro and Nano entering products immediately and developer APIs following a week later.
OpenAI launches DALL·E 3 in the Images API
At its first DevDay, OpenAI opens DALL·E 3 through the Images API, allowing developers to integrate text-to-image generation under the dall-e-3 model name.
OpenAI launches the GPT-4 Turbo preview at DevDay
OpenAI announced GPT-4 Turbo with a 128K context window and opened the preview to paying API developers, alongside a separate vision preview model.
DeepSeek starts with an open-weight code model
DeepSeek releases code models from 1.3B to 33B parameters, making open weights and code part of its product strategy from the outset.
DALL·E 3 becomes available to ChatGPT Plus and Enterprise users
OpenAI makes DALL·E 3 available to ChatGPT Plus and Enterprise users, who can generate an image in conversation and request further changes.
Baidu launches ERNIE 4.0 at Baidu World 2023
Baidu launched ERNIE 4.0 on October 17, 2023, with testing available to invited users and enterprise API applicants.
OpenAI announces DALL·E 3 and its synthetic-caption training method
OpenAI announces DALL·E 3; the accompanying paper attributes much of its stronger prompt following to recaptioning training images with detailed synthetic descriptions.
Alibaba Cloud releases Qwen-VL and Qwen-VL-Chat
Alibaba Cloud released Qwen-VL and Qwen-VL-Chat and published model weights on ModelScope and Hugging Face.
Stability AI releases SDXL 1.0
Stability AI released SDXL 1.0 and published its model weights and source code. The model generates native 1024 × 1024 images through a two-stage base-and-refiner pipeline.
Meta releases Llama 2
Meta releases Llama 2 weights and a commercial license, accelerating the open-model ecosystem.
Google launches PaLM 2 across more than 25 products and features
PaLM 2 emphasized multilingual, reasoning, and coding capabilities and launched across Bard, Workspace, Google Cloud, and other products.
Google opens limited testing for MusicLM
Google opened an AI Test Kitchen waitlist for MusicLM, allowing selected users to generate music from text descriptions and provide feedback.
Meta releases Segment Anything Model as open source
Meta released the promptable Segment Anything Model, along with its code, model, web demo, and the SA-1B dataset.
LMSYS releases Vicuna
LMSYS releases a model fine-tuned on high-quality user-shared conversations, with performance close to early ChatGPT.
OpenAI launches GPT-4 with staged access to text and image capabilities
OpenAI launches GPT-4 as a model that accepts text and images and emits text; text access begins through ChatGPT Plus and an API waitlist, while image input remains in limited partner testing.
ComfyUI introduces node-based image-generation workflows
ComfyUI organizes Stable Diffusion generation as node and graph workflows and becomes a key tool in the image-generation ecosystem.
OpenAI opens ChatGPT as a public research preview
OpenAI makes ChatGPT freely available as a research preview, using a GPT-3.5-series model trained with supervised dialogue data and reinforcement learning from human feedback.
Stability AI releases Stable Diffusion 2.0 as open source
Stability AI released Stable Diffusion 2.0 with a new text encoder, 512- and 768-pixel generation models, and several image-editing models.