China, compression and open model competition

• Continuing our series on model distillation techniques.

• This week in AI, we discuss Thinking Machine’s first open-weight model.

• The opinion section explores the idea that Google is the biggest threat to NVIDIA’s dominance.

For years, advances in AI have been a horse race: bigger models, better benchmark scores, more expensive clusters. This week we broke that framework. The most important development was not simply an advance in intelligence. They were competing to answer deeper questions. Who can own it, adapt it, and govern it?

Thinking Machines Lab has entered the modeling field. heartwarminga 975 billion parameter expert mixture model with 41 billion active parameters, native text, image, and audio inputs, a 1 million token context window, open weights, and an Apache 2.0 license. The company’s emphasis on tuning, controllable reasoning effort, and customization is more interesting than any leaderboard position. Inkling is a bet that the next frontier is a model that people can shape based on their own judgment, rather than simply borrowing from a centrally managed oracle.

Moonshot AI went in the opposite direction on scale while converging on the same access policy. Kimi K3 It has 2.8 trillion parameters, activates 16 out of 896 experts, supports vision and context of 1 million tokens, and is intended for long-term coding and knowledge work. Moonshot describes this as the first open model for the 3 trillion parameter class, but full weighting is not expected until later this month. This warning is important, but so is its trajectory. Chinese laboratories no longer compete solely on cost. They use openness itself as a strategic tool.

PrismML Bonsai 27B makes its most radical argument this week, through compression rather than scale. Its ternary model is 5.9GB, while the 1-bit version is 3.9GB, which PrismML says is small enough to fit into the available memory of modern smartphones. The company reports that it maintains most of its full-precision baseline performance across its benchmark suite. Independent testing determines how durable these numbers are. But the direction is clear. Information density can be as strategically important as raw information. Local models change latency, privacy, costs, and even bargaining power between users and cloud providers.

OpenAI also introduced another type of model this week. GPT-Red introduced a different but equally important form of scaling. This is not a consumer model or a new chatbot, but an internal automated red team system trained through self-play to attack other models, observe their defenses, and gradually invent stronger prompt injections. In an unfamiliar test environment, GPT-Red successfully compromised GPT-5.1 in 84% of scenarios (compared to 13% for human red teams). That attack was then used to significantly enhance GPT-5.6. The deeper idea is a new scaling law for safety. As your model’s functionality increases, the system that tests it can scale with it. The future may be defined by machine-speed ecosystems where attackers and defenders continually co-evolve, rather than models that simply improve themselves.

The common denominator is distribution as an engineering goal rather than openness as an ethical abstraction. Open weights allow adaptation. Sparse architecture makes large models economical. Extreme quantization pushes inference to its limits. Each move weakens various bottlenecks that have kept advanced AI concentrated in a few institutions.

This technology story clearly has a geopolitical counterpart. At the World AI Conference in Shanghai, Xi Jinping positioned open source AI as a global public good, promoted China’s assistance to developing countries, and elevated new international AI cooperation organizations as a means to shape global governance. Access rhetoric is not geopolitically neutral. Standards, training programs, and model ecosystems can build dependencies just as effectively as chips and cloud infrastructure.

Taken together, these announcements describe a fragmented and perhaps democratizing frontier. The battle is no longer just about who can build the smartest system. Gone are the days when intelligence will be centralized or portable, proprietary or adaptable, and controlled by businesses, nations, and communities. This week, the AI ​​no longer looks like a single race. It began to look like an emerging world order.

AI Lab: OpenAI

Summary: This article introduces GPT-Red, an automated red team model that uses self-play reinforcement learning to efficiently discover vulnerabilities and facilitate injection attacks. By integrating GPT-Red into their training pipeline, researchers were able to increase the robustness of production models like GPT-5.6 Sol against malicious instructions without reducing core functionality.

AI lab: CSAIL, MIT, Hebrew University of Jerusalem

summary: This paper analytically derives an optimal representation for contrastive learning using natural images and proves that a simple extension optimally yields a partial whitening process computable by a basic CNN with a sinusoidal filter. Experimental results confirm that CNNs trained with contrast loss on various datasets naturally learn these sinusoidal filters and perform partial whitening as predicted by theory.

AI lab: Google Research

summary: This paper explores the generative creativity of diffusion models and proposes that their neural network backbone naturally learns a smoothed version of the empirical scoring function rather than memorizing the exact training data. Through theoretical analysis and numerical experiments, the authors demonstrate that this score smoothing guides denoising dynamics and generates new samples that smoothly interpolate between training points even across complex nonlinear manifolds.

AI lab: Google Deep Mind

summary: The authors propose GenCeption, an integrated generalist vision model that leverages a pre-trained text-to-video diffusion backbone to perform various dense and sparse visual tasks via text instructions. By repurposing iterative diffusion into an efficient feedforward architecture, GenCeption achieves state-of-the-art performance across a variety of perceptual tasks while demonstrating emergent behavior such as sim-to-real transfer and zero-shot generalization.

AI lab: Yale University and University of California, Irvine

summary: This paper presents a comprehensive taxonomy and review of metacognition in large-scale language models, and details how the model’s ability to monitor and adjust its own cognitive processes is measured, induced, and improved. It synthesizes current findings on metacognitive functions such as confidence adjustment, self-reflection, and knowledge boundary detection, and highlights the potential of these mechanisms to enhance confidence, reasoning, and human-AI collaboration in LLM.

AI lab:Shanghai AI Research Institute

summary: ADVANCED MATHBENCH introduces a rigorous evaluation suite focused on the generation and process-level verification of advanced natural language mathematical proofs at the undergraduate and doctoral exam levels. Benchmarks reveal that Frontier LLM still has considerable difficulty constructing and verifying complex mathematical proofs, highlighting a significant bottleneck in its ability to accurately detect subtle logical errors.

thinking machine Open source Inklinga large MoE model with a 1M context window.

Moonshot AI Kimi K3 releaseda large-scale 2.8T parameter model optimized for long-term coding, reasoning, and knowledge tasks.

Mistral release First model optimized for embodied AI.

Prism ML Bonsai 27B releaseda model that can be run entirely on the phone.

1. Xi Jinping makes his debut at the World AI Conference in Shanghai He touted China’s low-cost AI and called for an open and cooperative technological order, saying AI development should be a symphony of international cooperation rather than a solo performance by one country.

2. Reuters reported The Cyberspace Administration of China approved the launch of Apple Intelligence on the back of a deal to integrate Alibaba’s Qwen model into iOS, iPadOS, macOS, and visionOS, and Baidu also confirmed that it is working with Apple on features for Chinese users.

3. Emergency announcement The $130 million Series C led by Creaegis has quintupled in a matter of months at a valuation of $1.5 billion on the back of a $120 million revenue run rate and over 200,000 paying customers for the AI ​​software creation platform.

4. Demis Hassabis publishes proposal on X We want a FINRA-style independent standards body funded by industry and supported by the U.S. government. The organization will review Frontier models up to 30 days before launch, ultimately restricting their deployment in the U.S. market. (Replaced TechCrunch with Hassabis’ original post.)

5. Reflection AI signs over $1 billion computing deal with Nebius The plan will run until 2029 and will give open model labs access to Nvidia’s GB300 chips. This is the company’s second major capacity acquisition after last month’s SpaceX deal.

6. reported by bloomberg Google says development of Gemini 3.5 Pro is several months behind schedule because the model hasn’t met internal goals, particularly in coding, and has frustrated employees who worry that Anthropic and OpenAI are ahead of the curve.

7. Greylock announces Greylock 18 at Xa $1.5 billion early-stage fund whose 18th is aimed at betting heavily on AI-native founders, managing partner Aseem Chandna argued that the next $1 trillion company is still a long way off.

8. TSMC June earnings report In the same month, sales increased by 67.9% year-on-year to NT$442.68 billion, and sales in the June quarter increased by 36% to approximately NT$39.6 billion, confirming that demand for AI is maintained. (Replaced Bloomberg with TSMC’s own monthly earnings release.)

9. Walden Robotics launches from stealthis a spinout of Toyota Research Institute, led by MIT’s Russ Tedreke, co-led by Toyota and Deviation Capital, with a valuation of $1.1 billion and about $300 million in seed funding, and its general-purpose robots are already shifting production at Toyota factories in North America.

10. SK Hynix raises $26.5 billion Just as Secretary of Commerce Lutnick pressured memory makers to build new factories in the U.S., the largest U.S. listing ever by a foreign company sold 177.9 million ADRs at $149 each, surpassing Alibaba’s 2014 record.