Expert Insights on ChatGPT: Technical Challenges, Applications, and Future Directions

After major tech companies such as Microsoft, Baidu, Alibaba, Tencent, and Xiaomi announced their entry into the ChatGPT arena, the model has become a headline topic in the technology sector.

The discussion shifts from pure technical details to commercial prospects and the impact on the internet ecosystem. Xiaohongshu, a lifestyle community with 200 million monthly active users, provides a massive multimodal dataset (images, short videos, text notes, queries, and comments) that is valuable for natural language processing and AI‑assisted content generation.

In a REDtech live broadcast ("REDtech来了" 第六期), NLP expert Li Lei (Assistant Professor, UCSB) and Xiaohongshu technical leaders Kaiqi (Tech Dept Head) and Yuchen (Multimedia Intelligent Algorithm Lead) explored ChatGPT’s technical difficulties and application outlook. Their key observations include:

ChatGPT sometimes gives correct answers but can also produce contradictory or fabricated responses.

Compared with GPT‑3, ChatGPT generates longer, topic‑focused replies and often admits uncertainty (“I don’t know”).

When prompted to act as a Linux terminal, ChatGPT can maintain a long operation history and produce logically consistent outputs, suggesting a form of “thinking”.

Further Q&A sessions covered several themes:

Model Limitations and Safety: ChatGPT may hallucinate, struggle with factual consistency, and lack real‑time internet access, which raises concerns for security engineers.

In‑Context Learning & Few‑Shot Demonstration: The ability to leverage a few examples at inference time (without parameter updates) is a major breakthrough that could be applied to many NLP tasks.

Human Feedback (RLHF): While low‑cost feedback can improve models, the impact on massive models like GPT‑3 is limited; nevertheless, ChatGPT’s success shows RLHF can yield significant gains.

Model Miniaturization: Researchers believe large models can be compressed for specific tasks, but preserving full capability remains an open research problem.

Future AIGC & Multimodal Applications: Combining large language models with vision, audio, and video generation could enable intelligent content creation, smart customer service, and enhanced search in platforms like Xiaohongshu.

Impact on Search Engines: ChatGPT may outperform traditional search for certain Q&A scenarios, but replacing full‑featured search engines is unlikely in the short term.

The experts also discussed the potential of integrating ChatGPT‑style models into Xiaohongshu’s ecosystem, emphasizing the need for careful risk management, the importance of high‑quality feedback loops, and the opportunities presented by multimodal generation.

Overall, the interview highlights the current strengths of large language models, their remaining challenges (hallucination, safety, scalability), and promising research directions such as in‑context learning, RLHF, model compression, and cross‑modal AI generation.