ChronicleBio, Mind Lab, and GPT-Live Architecture Overview

Former OpenAI executive Fidji Simo left the company after seven years with POTS (Postural Orthostatic Tachycardia Syndrome) to found ChronicleBio, aiming to use AI to cure POTS and other chronic diseases. The company has collected 153 TB of data from blood samples of chronic disease patients and plans to launch at-home blood collection to expand the dataset. ChronicleBio hopes to use this data to deepen understanding of diseases and improve the success rate of clinical trial drugs.

This initiative reflects the trend of AI applications in healthcare, but data scale and disease complexity remain challenges. Currently, the company's plans are still in early stages, and its efficacy has not been validated.

Mind Lab claims its Macaron-V1 model surpasses GLM-5.2 on benchmarks. The model is built by attaching five LoRA expert modules, each about one billion parameters, to GLM-5.1, with the system dynamically switching to the most suitable expert model based on the task. Data accumulated during use can be distilled into dedicated LoRA adapters and continuously updated with model calls.

This design aims to achieve continual learning, but specific benchmark details and third-party validation have not been made public, and its performance advantages still need independent confirmation.

OpenAI rebuilt its voice architecture around a full-duplex model capable of simultaneously listening and speaking. The system combines stateful reasoning, asynchronous delegation, dynamic context management, and low-latency media transport to maintain conversational responsiveness while supporting advanced reasoning and tool use.

The details of this architecture showcase OpenAI's technical investment in real-time interaction, but actual performance and user experience remain to be seen.

OpenAI's unreleased model Astra solved ten major open math problems, indicating that AI has achieved superhuman capabilities in networking and coding, and also in advanced mathematics, similar to how non-AI computers have long performed in basic math. These problems are well-defined, formalized, and their solutions are easy to verify.

The article notes that much of AI research and development is verifiable, and labs that make progress on AI self-improvement loops will hold a strong advantage. However, the model is unreleased, and its capabilities have not been independently verified.

GPT-5.6 Sol xhigh uses more than twice the tokens per session in Codex workflows compared to GPT-5.5 xhigh, effectively halving the value of token-based quotas. At the same token price, 2.25 times the tokens means about 2.25 times the cost. Additionally, GPT-5.6 Sol introduces new cache write fees, which GPT-5.5 did not have.

Meanwhile, the White House will host a meeting with AI companies to discuss a new voluntary model testing framework, ordered by President Trump, allowing developers to share models with the government to assess cybersecurity risks. Companies like OpenAI, Google, and Anthropic will attend, but evaluation criteria remain confidential.

In video generation, MiniMax H3 became the first open-source model to top the AI video ranking, ranking first in video editing, second in text-to-video, and third in image-to-video.

DeepSeek's V4-Flash model is reportedly the lowest-cost among well-known models, being 105 times cheaper than Anthropic's Claude Fable 5. These data come from research firms but lack specific details.