AI News Roundup: Thinking Machines' Inkling, GPT-Red Security, Perplexity Sandboxes

Thinking Machines released Inkling, a 975B parameter mixture-of-experts model with 41B active parameters, supporting multimodal reasoning and a million-token context window. The model offers customization through the Tinker platform and comes with a smaller preview version. This marks another significant advancement in the open-source community for large models, but the accompanying safety challenges are even more pronounced.

Perplexity AI's SPACE sandbox platform is a direct response to this challenge. SPACE provides temporary, isolated execution environments for AI agents, destroyed immediately after task completion. Its security architecture includes control plane and node-level services for managing credential access, offering credential isolation, rolling snapshots, and encrypted storage. SPACE supports local and offline deployment, ensuring sensitive data never leaves the enterprise environment.

OpenAI trained GPT-Red, a model specifically designed to generate adversarial prompts, automating the discovery of model vulnerabilities. By incorporating attacks generated by GPT-Red into training, OpenAI reduced the failure rate on a difficult prompt injection benchmark by 6x. This approach demonstrates AI's dual role in safety testing: both as attacker and defender.

GPT-Red's iterative generation capability allows it to continuously explore new attack vectors, whereas traditional red teaming often relies on humans. This automated red teaming method could become standard practice for future AI safety evaluation, but it also raises concerns about misuse—if similar technology were used maliciously, the consequences could be dire.

Researchers conducted a recursive self-improvement experiment on an automated research agent over 8 days. The system included an inner loop (optimizing code to pass evaluations) and an outer loop (optimizing the inner loop agent's framework code to improve average scores). The results were astonishing: the system designed a new search algorithm, reduced prompt size by 16x, and built a hierarchical system to prevent reward hacking. The final performance surpassed a system manually tuned by researchers over two years.

This is the first experimental proof that recursive self-improvement is feasible in AI systems. However, the research also points out potential risks: the system may produce unpredictable behavior during optimization, especially when reward functions are imperfectly designed. This finding has profound implications for AI safety, as recursive self-improvement could lead to rapid capability gains but also amplify safety vulnerabilities.

Model routing systems face challenges far beyond simply selecting the best model. Factors such as caching, workload interaction, and service infrastructure often have a greater impact on cost and latency than model pricing itself. Effective routing requires optimization across multiple dimensions to ensure the entire system operates at its optimal point, rather than just selecting the best model for a single task.

Meanwhile, GPT-5.6 Sol ranks first in the web design arena, improving 18 positions over its predecessor GPT-5.5. Sol can identify and compress common AI design anti-patterns, combining powerful templates with high personalization. This indicates that AI's capabilities in aesthetics and user experience are rapidly improving, but it also means AI-generated content may become increasingly indistinguishable from human creations.

NVIDIA introduced the Blackwell architecture-based Jetson Thor T3000 and T2000 modules for edge robotics, vision AI, and agent workloads. The platform has been adopted by companies such as Amazon Robotics, Boston Dynamics, FANUC, and 1X. The rise of edge AI places higher demands on chip architecture.

Systolic arrays handle over 95% of modern AI chip computations, reducing memory traffic by moving matrix data locally between processing units. Larger arrays can increase peak throughput but are harder to saturate, so compiler scheduling often determines whether a chip achieves effective utilization. This detail reveals that AI hardware performance depends not only on design but also on software stack optimization.

Anthropic is preparing for an IPO, with Goldman Sachs, Morgan Stanley, and JPMorgan leading investor meetings. The company recently completed a $65 billion funding round at a $965 billion valuation, surpassing OpenAI's $852 billion. This IPO will test public market interest in AI companies and may pave the way for OpenAI's subsequent listing.

However, behind the high valuations lies significant uncertainty. AI companies currently rely primarily on funding rather than profitability to support their valuations; if market sentiment shifts, the bubble could burst. Anthropic's IPO will serve as an important bellwether for the AI industry.