Overall, there haven't been many changes in my mind in 2023, but there have been bigger changes in my life.

"I've kept the summary in my notebook, and now I've publicly published the predictions on my blog, which you're welcome to subscribe to. Having spent a significant amount of time on Large Language Models (LLMs) in 2023, I've used my intuition to write the following predictions for the next year, or possibly for the next five years.

In my predictions for 2022, about 50% of the predictions regarding blockchain and Defi were correct, about 80% of the predictions regarding China's epidemic prevention policies were correct, and about 40% of the predictions regarding information security were correct.

Impact of large language models

• Better open-source frameworks than Langchain will emerge: Langchain currently has many bugs, chaotic management, and unsatisfactory code abstractions. When OpenAI iterates with new features and APIs, there may be more competitive frameworks emerging.
• AWS starts offering affordable APIs like OpenAI (Azure): Currently, Sagemaker provides deployment of large language models like Llama2, serving small and medium-sized enterprises. However, what more users may need is not fine-tuning but cheaper APIs. Currently, 80% of startups use RAG, and what they need the most is APIs.
• Integration of compilers and large language models: People will gradually realize that it is time to update the entire toolchain, not just code completion at the application level like Copilot. Large language models will handle compiler errors and provide better suggestions, eliminating the need to copy and paste error messages to ChatGPT.
• Refactoring of creative tools: Soon, everyone will agree that simply adding an AI chat feature to an existing tool is not enough. It will be necessary to reconstruct the entire product's interaction based on AI logic. The most crucial interactions in creative interaction, as I see it, are: creation, observation, using tools for further creation, further observation, and further adjustments. This kind of interaction cannot be accomplished through simple conversation. The combination of simple NUI and GUI may not be effective; it is necessary to delve deeper into human habits and create better tools.
• More accurate search engine results: Large language models can annotate videos, images, audio files, and other types of files. They also have a better understanding of semantics, which may lead to more accurate search results.
• Mature web crawlers based on large language models emerge: A web crawling tool at the level of Scrapy, with a large language model at its core, will appear for information retrieval, and corresponding startup companies will emerge.

Information Security and Privacy

• Privacy and copyright issues intensify: The extensive use of artificial intelligence will inevitably make companies realize the importance of data. Once data is collected, it is easy to leak and abuse.
• Prompt injection attack frameworks emerge and mature: With the combination of LLM and various services, prompt injection attacks based on principles similar to XSS and SQL injection will become popular.
• Frequency of major incidents in cloud servers continues to increase: In the current centralized internet structure, cloud services are becoming increasingly concentrated among a few vendors. The product lines of these vendors are also becoming more complex, and as the situation becomes more precarious, the core services become more complex. In the coming years, it is only a matter of time before we hear news of large-scale outages from AWS.
• LLM will be applied to security solutions such as symbol injection: no matter how many rules the XSS filter writes, it is better to send it to the openai API for review of injection strings. The openai API may have a slow response time, but we can train a model with fast response and fewer model parameters.
• LLM-based penetration testing tools have emerged: penetration testing has a certain process sequence and is highly related to handling content and strings, making it very suitable for building a set of tools using LLM.

China's domestic policies.

• Large language models applied to content moderation: in the past few years, government security orders have supported many companies engaged in facial recognition. In the future, government orders will be important for social media moderation.
• The number of military and security orders is far less than the previous round of chip and security orders: after the investigation and cleanup of the military personnel, there is a corresponding time for personnel adjustments. Personnel adjustments and policy adjustments occur at the same frequency. After personnel adjustments are completed, policies will be adjusted. This is a consistent pattern of behavior.
• There will be no groundbreaking work in areas such as chips and artificial intelligence: in the past sixty years, from scientific research to the business world, from Chen Bodai to Xiao Yaqing, almost everyone has to follow a cycle of policy changes every three years, resulting in relatively weak accumulation in any technological direction. Therefore, the possibility of groundbreaking work is very small.
• Speculative investments have increased significantly: due to the instability and incoherence of policies, the investment calculation period will be shortened. Long-term investments will either withdraw from China or transform into short-term speculation.
• Increase in mass incidents: due to the current and expected economic downturn caused by deglobalization, companies are expected to make further decisions to reduce labor costs. The problems accumulated in the past are gradually spreading from the border to the core, leading to an increase in mass incidents and increasing tension between various components.

Heidegger said that poetry is the other side of thinking (poetry and thinking are the way of speaking, and it is a prominent way of speaking). I will vulgarly write a vulgar poem for my vulgar life. Goodbye, 2023!