➡ Chinese AI startup, DeepSeek, recently faced a major security breach, exposing sensitive user data and operational details. Despite quickly addressing these vulnerabilities, it’s unclear if any harmful actors accessed the data before the issue was resolved. Meanwhile, DeepSeek’s AI performance has been underwhelming, with a high failure rate and accusations of false claims. Amid these challenges, Onex acquired Kind Robotics, aiming to advance the development of intelligent, adaptable robots for everyday use.

Onex just acquired Kind Robotics to build new humanoids, but first, Chinese AI startup DeepSeek just suffered a ridiculous security failure. The company’s database was somehow left completely exposed on the internet, allowing literally anyone to potentially steal sensitive user information or launch several other attacks. But that’s just the beginning. The exposed Clickhouse database linked to DeepSeek’s infrastructure granted full control over operations, including access to internal data. Even more shocking was the breach’s inclusion of over a million lines of log data, chat histories, API secrets, backend information, and other operational metadata. To make matters worse, the database’s configuration allowed privilege escalation and enabled unauthenticated users to execute SQL queries directly via a web browser, opening the door to account takeovers, ransomware attacks, data manipulation, remote code execution, and more.

And while DeepSeek quickly addressed these vulnerabilities, it remains unclear whether any malicious actors already accessed or downloaded sensitive data before the issue was fixed. If so, the potential consequences may include further data leaks, security exploits, or unauthorized use of DeepSeek’s AI. In terms of performance, DeepSeek’s R1 has reportedly underperformed in a NewsGuard audit, achieving only 17% accuracy in delivering news while ranking 10th out of 11 compared to Western competitors like OpenAI’s ChatGPT and Google Gemini. As for results, DeepSeek’s R1 repeated false claims 30% of the time and provided vague or unhelpful answers 53% of the time, resulting in an 83% fail rate, which is significantly worse than the 62% average fail rate of its Western rivals, raising doubts about DeepSeek’s claims of outperformance.

All the while, DeepSeek faced large-scale cyberattacks that shut down its web service and prompted a temporary pause in new user registrations. Plus, scrutiny over its privacy practices has intensified, particularly as its Chinese origins have led to the US government banning its employees from using DeepSeek altogether. Then in Italy, DeepSeek’s apps became unavailable shortly after the country’s data protection regulator requested information about its sources of training data. And to add to DeepSeek’s troubles, OpenAI and Microsoft are currently investigating whether the startup improperly used OpenAI’s API to train its own models, with OpenAI reporting abnormal spikes in usage from China.

This practice, known as distillation, involves training a model on the outputs of another model, which OpenAI claims is intellectual property infringement. Despite these setbacks, though, DeepSeek is still gaining mainstream traction as users excitedly download the open-source model to run locally on their own computers. And while it’s uncertain whether the rough start is due to ignorance, state-sponsored agendas, or something else entirely, another worldwide AI moment may be about to occur with humanoids, as one X-technology’s just-acquired kind humanoid, combining two teams with a shared vision for advancing general-purpose robots designed to assist in daily life. Looking forward, the acquisition aims to accelerate the development of intelligent humanoids that are finally capable of learning, adapting and working alongside humans in order to bring practical household robots closer to reality.

But what makes Kind Robotics unique is their integration of large language models to enable natural communication and adaptability. Unlike robots trained for specific tasks, the company’s prototypes were designed as general-purpose machines capable of assisting people in a wide variety of everyday situations. This approach reflected a belief that humanoid robots should evolve and improve by living and learning among humans, rather than being confined to isolated environments or rigid programming. In fact, Kind Humanoid began its journey as a small team focused on creating bio-inspired, human-centric robots, with its bipedal humanoid named Mona being designed not only for home use, but also for applications in industries like healthcare.

Meanwhile, Alibaba just introduced its newest QEN 2.5 MAX language model claiming that it was trained on a record-breaking 20 trillion tokens of data, which if verified would position the company to begin taking on industry-leading competitors like OpenAI, Anthropic, Meta and Google Next. QEN 2.5 MAX stands out for its sheer scale, with Alibaba asserting that the 20 trillion tokens used in its training surpass the data size employed by many leading models, including DeepSeek V3 and Llama 3.1405B, both of which are estimated to have been trained on 15 trillion tokens. And while some rivals like GPT40 and Claude 3.5 Sonnet have not disclosed their exact training data sizes, Alibaba’s latest claim of a new benchmark in data volume has caught the attention of the AI community.

Crucially, the model leverages a mixture of experts’ architecture, a system designed to allocate computational resources more efficiently by activating only relevant parts of the model for specific tasks. This approach enables QEN 2.5 MAX to scale effectively while maintaining strong performance across diverse use cases. In benchmark tests, QEN 2.5 MAX outperforms DeepSeek V3, GPT40, Claude 3.5 Sonnet, and Llama 3.1405B in specific areas, with the most notable results coming from Arena Hard and Livebench, which measure complex reasoning and real-world task performance. However, its advantage over competitors remains modest in many categories, reflecting broader industry trends where gains from larger datasets are increasingly incremental.

Additionally, Alibaba developed QEN 2.5 MAX using a combination of supervised fine-tuning and reinforcement learning from human feedback, which are training techniques widely adopted in the development of advanced AI models. And currently, QEN 2.5 MAX is accessible through Alibaba Cloud’s API and via QEN Chat, which is Alibaba’s chatbot platform that supports features such as real-time web search and content generation. Meanwhile, other models such as QEN 2.5-VL are available as open-source tools, with Alibaba having opted to keep QEN 2.5 MAX as API only for now. This decision appears to target enterprise developers, with Alibaba using OpenAI-compatible APIs and competitive pricing to attract users to its cloud platform.

But Alibaba has not disclosed the exact sources of its training data, as experts suggest that synthetic data that was generated by other AI models has likely played a significant role in achieving its unprecedented token count. This aligns with a growing trend in AI development, as companies increasingly rely on synthetic data to scale training efforts. However, the reliance on vast training datasets has reignited debates within the AI community. Some experts argue that test-time computing power, the amount of computational resources used when a model generates responses, may now play a more significant role in advancing language model performance than training data size alone.

As with other Chinese AI models, QEN 2.5 MAX operates under government-mandated content restrictions, limiting the range of topics it can discuss. With QEN 2.5 MAX, Alibaba is positioning itself as a serious contender in the growing AI market. By combining a record-breaking dataset, competitive benchmark performance, and developer-friendly tools, the model reflects Alibaba’s ambition to challenge Western AI leaders like OpenAI and Anthropic. And while QEN 2.5 MAX’s performance gains are incremental rather than groundbreaking, its strategic focus on enterprise adoption and accessible APIs could make it a valuable tool for developers and businesses worldwide. [tr:trw].