ChatGPT app rolls out lazy-loading update, making long conversations 10x faster
产品应用
The ChatGPT app has enabled on-demand loading for new conversations, initially retrieving only the latest few exchanges. Brent Traut, the engineer responsible for the change, said loading time and memory use for long conversations both fell by more than 90%, making them roughly 10 times faster to open.
ChatGPT has changed how it retrieves long chat histories, with new conversations already using on-demand loading and older conversations set to migrate gradually. Previously, the app loaded the entire history when a conversation was opened, so longer chats took more time and memory; it now retrieves only the latest few exchanges first and loads earlier content when the user scrolls upward. According to Brent Traut, the engineer responsible for the change, loading time and memory use for long conversations both decreased by more than 90%, opening became roughly 10 times faster, and operations such as starting a separate conversation from a specific message are also faster. The mechanism is available on Mac, Windows, Linux, CLI, the VS Code extension, and apps built on the open-source app server. Input lag during extended chats and related issues remain under optimization.
OpenAI reportedly buys tens of thousands of Mac mini and Mac Studio units for Agent training
前瞻与传闻
The Information reported that OpenAI bought tens of thousands of Mac mini and Mac Studio units over the past several months for reinforcement learning and computer-use Agent training. Anthropic was also reported to be renting Mac minis through Amazon AWS.
The Information reported that OpenAI purchased tens of thousands of Mac mini and Mac Studio units over the past several months and is currently using them mainly for reinforcement learning and computer-use Agent training. OpenAI has not formally confirmed the purchases. According to the report, the devices are considered suitable for on-device Agentic orchestration because of their unified memory architecture and low-latency Thunderbolt 5, which can bypass the standard TCP/IP networking stack. The report also said Anthropic is renting Mac minis through Amazon AWS to optimize Agent workflows. Anthropic has not formally confirmed the rentals.
字节豆包2.2 reportedly delayed to strengthen coding and Agent capabilities
前瞻与传闻
White Whale Lab cited multiple people close to ByteDance as saying that Doubao Large Model 2.2, originally scheduled for August, has been delayed, with the extra development time allocated to coding, tool-calling, and Agent capabilities and a coding target set for year-end.
White Whale Lab reported that ByteDance has delayed Doubao Large Model 2.2, which was originally scheduled for release in August. Multiple people close to ByteDance said the company intends to use the longer development cycle to improve coding, tool-calling, and Agent capabilities, with the goal of bringing its coding performance by year-end to a level of industry impact comparable to recent models from Zhipu and Kimi. The report also said ByteDance accelerated personnel changes and reorganized the Seed foundation model unit in mid-August into four first-level departments, including pre-training data and reinforcement learning, with an emphasis on coding post-training and Agentic capabilities for office scenarios. The model’s training did not use data distillation, while claims that its parameter count could be increased to 5T or even 10T were identified as earlier rumors.
