Thinking Machines Lab Releases "Inkling", Its First Open Model

Thinking Machines Lab, led by former OpenAI executive Mira Murati, has released "Inkling", the company's first public model. It is a multimodal model with 975B parameters that accepts text, images, and audio as input and produces text as output. It is available as open weights, and developers are free to use it.

Inkling is designed for a wide range of use cases, including agents, tool use, coding assistants, chatbots, and RAG systems. It has also ranked among the top trending models on Hugging Face, reflecting strong interest from the community.

The company has been building its AI infrastructure privately for the past year and a half, and Inkling is positioned as its "first public deliverable." As a new rival to competitors such as Anthropic and OpenAI, its future moves will be worth watching.

OpenAI Releases "GPT-Red" — Automated Red Teaming That Uses LLMs to Attack LLMs

OpenAI has released "GPT-Red", an automated red-teaming LLM it built to strengthen the safety of its own models. GPT-Red automates red teaming — a type of cyberattack — and serves as a "sparring partner" that uncovers and addresses vulnerabilities in other models.

According to OpenAI, GPT-5.6, released last week, became "the most robust release to date" through training with GPT-Red, with particular improvements in prompt injection defenses. GPT-Red features a self-improvement mechanism based on self-play, allowing it to continuously learn and evolve new attack patterns.

GPT-5.6 Sol Refutes a 30-Year-Old Statistics Conjecture in 90 Minutes

A statistics professor at the University of Pennsylvania reportedly used OpenAI's GPT-5.6 Sol Pro to refute a conjecture about the Benjamini-Hochberg procedure that had remained unsolved for 30 years, in about 90 minutes. The predecessor model, GPT-5.5, reportedly failed to solve it even after 20 hours.

This case suggests that LLMs are reaching a practical level on tasks that require creative reasoning and mathematical intuition. Rigorous verification of the results, however, remains a task for the future.

US Startups Choose Chinese AI Models for Cost Savings

According to NPR, some US startups have begun migrating from US-made models to cheaper Chinese AI models in order to cut costs. Development and operating costs for major US AI models have soared, placing a heavy burden especially on early-stage startups.

Chinese models are rapidly catching up in performance as well, making them an increasingly attractive option from a cost-performance perspective. This trend is drawing attention as a new angle on the geopolitical competition over AI infrastructure.

Independent Developer Unveils O(1) Attention "Wave Field"

An independent researcher has unveiled a new architecture called "Wave Field LLM" in the Reddit LocalLLaMA community that replaces the standard attention mechanism. Using wave convolution based on FFT (Fast Fourier Transform), it reportedly keeps inference compute at O(1) per token — constant regardless of context length.

It is reported to run in environments where standard attention runs out of memory (OOM) even at 128K context, and to achieve 80+ tok/s on a laptop CPU. At this stage it is a base completion model (not a chat model) and still requires independent verification, but it has generated significant buzz in the community.

---