AI Daily · 2026-08-31

Google and DeepSeek both shipped new models: TimesFM-3 brings zero-shot forecasting to multivariate time series, while DeepSeek-V4-Flash-Vision-Exp adds multimodal capability to the V4 series and significantly improves agent benchmarks. Tencent released three ContextPilot checkpoints, 14B, 8B, and E4B, all focused on active context management for long-horizon agents through planning, memory, and unloading low-value context. Simon Willison argues that ChatGPT Work is essentially a repackaged Codex for non-programmers, with the real differences being a networked execution environment, headless browser, persistent file system, and site publishing. OpenAI also said ChatGPT Ads has reached a $1 billion annualized revenue run rate.

North America · First-hand

OpenAI

⭐⭐ [Product Update] A milestone in expanding access to AI

OpenAI announced that its advertising product ChatGPT Ads has reached a $1 billion annualized revenue run rate. The product is now expanding globally. OpenAI says the ad business supports broader access to AI through free and affordable options.

Google

⭐⭐⭐ [Model Release] TimesFM-3: A zero-shot foundation model for multivariate forecasting

Google Research introduces TimesFM-3, a zero-shot foundation model for multivariate time series forecasting with 330 million parameters, pretrained on a corpus of over 1 trillion real-world and synthetic time points. Unlike previous versions limited to univariate forecasting, TimesFM-3 natively supports multiple targets, past covariates, and past-future (dynamic) covariates, capturing cross-series dependencies without task-specific fine-tuning. Built on a decoder-only transformer, it patches data into 32-step segments and uses alternating causal temporal attention and full variate attention to blend temporal and cross-series patterns, with non-autoregressive single-pass decoding. The authors report that it significantly outperforms other forecasting models across major benchmarks, and the model is available via GitHub and HuggingFace.

East Asia · First-hand

DeepSeek

⭐⭐⭐ [Model Release] deepseek-ai/DeepSeek-V4-Flash-Vision-Exp

DeepSeek released DeepSeek-V4-Flash-Vision-Exp, its first experimental multimodal model in the DeepSeek-V4 family, built by adding visual modules to the DeepSeek-V4-Flash architecture and undergoing continued training to unlock visual understanding. Compared with DeepSeek-V4-Flash-0731, it delivers substantial gains on multimodal agent benchmarks (e.g., ApexBench Pass@1 from 26.2 to 36.5, Agents' Last Exam from 25.2 to 27.3) while maintaining comparable text-only agent performance (e.g., Terminal Bench 83.9 vs 82.7, DSBench-Hard 63.6 vs 59.6). The model is released under the MIT license, and the repository includes a tokenizer, prompt encoding reference, and a minimal PyTorch inference implementation covering the vision encoder, DFlash attention, MoE, Hyper-Connections, and more.

Tencent Hunyuan

⭐⭐ [Model Release] tencent/ContextPilot-14B

Tencent has released ContextPilot-14B, a model checkpoint based on Qwen3-14B for proactive context management in long-horizon language-model agents. The framework uses fine-grained reinforcement learning to teach agents to plan, maintain long-term memory, and offload less useful context while reasoning and using tools. It combines an extended context-management toolset, context-aware partial rollout, and fine-grained credit assignment. Evaluated on long-context QA and deep-search tasks, it achieves stronger performance than existing baselines with a more compact working context. The weights and code are open-sourced, though the checkpoint alone does not include the tool runtime.

⭐⭐ [Model Release] tencent/ContextPilot-E4B

Tencent released ContextPilot-E4B, an 8B text-generation model checkpoint based on Google's Gemma4-E4B-it, part of the ContextPilot framework for proactive context management in long-horizon language-model agents. It teaches agents to plan, maintain long-term memory, and offload less useful context while continuing reasoning and using tools. The framework combines an extended context-management toolset (planning, structured memory, retrieval, soft context offloading), context-aware partial rollout, and fine-grained credit assignment. Evaluations on long-context QA and deep-search tasks reportedly achieve stronger performance with more compact working contexts than existing baselines. The release includes a paper (EMNLP 2026) and a GitHub repository.

⭐⭐ [Model Release] tencent/ContextPilot-8B

Tencent Hunyuan released ContextPilot-8B on HuggingFace, a Qwen3-8B-based checkpoint implementing the ContextPilot framework for proactive context management in long-horizon language-model agents. Built on fine-grained reinforcement learning, the framework teaches agents to plan, maintain long-term memory, and offload less useful context while continuing reasoning and tool use. It combines an extended context-management toolset (planning, structured memory, retrieval, and soft context offloading), context-aware partial rollout, and fine-grained credit assignment that trains intermediate snapshots via downstream branch outcomes. Evaluated on long-context QA and deep-search tasks, ContextPilot achieves stronger performance with a more compact working context across base models and benchmarks. Loading code is provided, but full execution requires the tool definitions, agent runtime, and evaluation pipeline from the ContextPilot repository.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Opinion

⭐⭐⭐ [Opinion] Understanding ChatGPT Work

Simon Willison breaks down OpenAI's ChatGPT Work, arguing it is actually two products: a cloud version (Work Cloud) accessible via web and mobile apps, and a local desktop-app version (Work Local) that can access local files and run programs, essentially a re-skinned Codex made less intimidating for non-programmers. Work is currently limited to $20/month+ subscribers and offers features absent from Chat, including a choice of GPT-5.6 Sol/Luna/Terra with various reasoning levels, an internet-enabled code execution environment, a headless Chrome browser, a persistent filesystem, the ability to publish ChatGPT Sites, and sub-agent sessions. The author finds OpenAI's official guidance on when to use Chat vs Work almost useless, and instead clarifies the real differences: Work's unique capabilities. He also notes Work sessions are likely billed against the Codex allowance, while Chat sessions have a separate allowance.