A roundup of AI news: Thinking Machines unveils Inkling, a new AI model; GPT-Red focuses on security; and Perplexity launches sandbox environments for safe AI experimentation.
SynthePulse Insight · AI deep reading
Version 2 · 1 source
Thinking Machines releases a massive parameter model, Anthropic's valuation surpasses OpenAI's, but autonomous recursive improvement experiments lasted only 8 days, and safety test results are based on specific benchmarks.
Thinking Machines has released its first open-source model, Inkling. Utilizing a Mixture of Experts (MoE) architecture, the model boasts a total of 975 billion parameters, with 41 billion active parameters. This configuration aims to balance computational efficiency with reasoning capabilities.
Inkling supports multimodal reasoning and features a context window of up to one million tokens. This long-context capability enables it to handle extremely complex document or codebase analysis tasks. The model offers customized services through the Tinker platform and includes a smaller preview version for developer testing.
Anthropic is accelerating its Initial Public Offering (IPO) plans. Investment banks such as Goldman Sachs, Morgan Stanley, and JPMorgan Chase are arranging investor meetings for the company. Recently, Anthropic completed a $65 billion funding round, resulting in a post-money valuation of $965 billion.
This valuation level exceeds competitor OpenAI's $852 billion. Anthropic's high valuation reflects market recognition of its safety-first approach, but it has also sparked discussions about whether there is a valuation bubble in the AI industry. This IPO will serve as a key node for testing public market judgment on the true value of AI companies.
Researchers conducted a pioneering experiment using a self-developed autoresearch agent to achieve recursive self-improvement within 8 days. The system included inner-loop code optimization and outer-loop agent framework optimization, ultimately designing a novel search algorithm.
The results showed that the autonomous system outperformed baseline systems that had undergone two years of human tuning, while reducing prompt size by 16 times. However, this experiment lasted only 8 days, and its long-term stability and generalization capabilities remain unclear. Additionally, preventing reward hacking remains a security issue that the system needs to address.
OpenAI launched the GPT-Red tool to iteratively generate adversarial prompts to expose model vulnerabilities. It is claimed that incorporating GPT-Red attacks into training reduced the failure rate of GPT-5.6 Sol on difficult prompt injection benchmarks by 6 times.
It should be noted that this '6x reduction' effect is a source claim based on specific benchmark tests, and its generalizability has not been independently verified. Meanwhile, Perplexity launched the SPACE sandbox platform, which ensures the security of AI agents handling sensitive tasks through measures such as temporary sandboxes, credential isolation, and encrypted storage.
At the application level, OpenAI's GPT-5.6 Sol ranked first in the Design Arena web design competition, an 18-position improvement over its predecessor, GPT-5.5. The model can identify and compress common AI design anti-patterns, combining strong templates with highly personalized outputs.
These developments indicate that AI is not only making breakthroughs in underlying model parameters and security mechanisms but also demonstrating stronger practicality and aesthetic capabilities in real-world application scenarios such as design and coding.
This article is strictly written based on the provided TLDR AI news summary. Regarding Anthropic's valuation, the magnitude of GPT-Red's performance improvement, and the results of the recursive self-improvement experiment, all are marked as source claims or limited experimental evidence and should be interpreted with caution.
The AI industry is at the intersection of technological breakthroughs and capital frenzy. While model capabilities and security mechanisms have significantly improved, the long-term impact of frontier technologies like recursive improvement remains uncertain, and investors need to be wary of valuation bubble risks.
Primary report
Primary source