Review the real performance of LLM in 2023 from the perspective of an application developer
Feb 17, 2024 · Jah Guo
Recently, OpenAI released Sora, adding to the growing interest in Large Language Models (LLMs). As a product manager who has been deeply involved with LLMs for a year, I will share my practical work experience in the broader context of LLMs.
1. LLM has become the standard of political correctness in tech giants
After SNS, mobile internet, blockchain, and metaverse, 2023 is once again hailed as the first year of the fourth industrial revolution - the AI era. Major tech companies have discovered new opportunities beyond the traditional concepts and have commenced fierce competition. OpenAI, Microsoft, Google, and Meta are rapidly advancing with clear strategies. Different from before, large Internet companies in China are now cautious and uncertain about their next steps due to limited computing power and unclear application scenarios.
2. Bosses experience love, fear, and helplessness
Bosses and CIOs are also being crazily brainwashed by self-media. Despite most product decision-makers not understanding how large models operate, they still decide to take a step ahead of others in implementing large models, both internally and externally. After branding ChatGPT as their own assistant, they found that internal corporate information was not only sent outside the company but even abroad, thus triggering legal risks. And the sluggish interaction is almost unusable.
3. The biggest tech giants around the world are copying ChatGPT
Since ChatGPT started the dialogue + Prompt interaction mode, tech giants have mirrored its intelligent assistant almost identically. APIs, multimodal applications, and a store follow this. While the next groundbreaking feature from OpenAI remains unknown, one thing is certain: everyone is eagerly waiting to follow suit.
4. It might be "The Emperor's New Clothes"
Whether it's on social media, industry forums, or product launches, many people claim that large models have revitalized their products and even incubated industry-specific models for thousands of businesses. However, no one dares to admit that these large models in professional fields often lack understanding and there are many tasks they can't perform. The acquisition of professional or specialized knowledge by large models often comes with a significant cost.
5. While the media is excited, the PMs, coders, and business experts are all feeling down
A significant portion of time, around 90%, is consumed by data-related tasks such as collection, writing, cleaning, formatting, slicing, training, and labeling. This process continues day in, day out, for weeks and months. Often, the production research team is uncertain about the accuracy of the information. Business experts are unsure of how to train the model. Product managers are contemplating ways to facilitate direct communication between business experts and the LLM. The end of this process remains unclear to everyone.
6. Rapidly become a new growth point for the cloud platform
Traditional cloud platform providers have introduced training platforms for LLMs based on ML. Following MLOps, the concept of LLMOps was introduced. Despite technical and computing power challenges, the cloud platform transforms LLMs into next-generation infrastructure. However, its design is often perceived as crude and difficult for business experts to use directly. It appears that cloud service product managers have yet to fully understand how users will utilize the capabilities of LLMs.
7. Gradually transitioning from the forefront to the background, focusing on business as the core
After six months of intense effort, I've come to realize that large models alone can't accomplish everything. Building business processes with large models as the primary method falls short of delivering timely, cost-effective solutions that meet business needs and provide a solid return on investment. Many product managers now use large models offline to aid in asynchronous knowledge and data processing. To allow these models to effectively learn and combine new forms of knowledge, we've given this business-logic-containing code snippet a new name — Agent.
8. The industry is beginning to define a new form of next-generation applications
Imagine a future where people no longer have to decide which website or app to use. Instead, they simply tell the AI their intentions, and the AI directly finds the answer or performs an operation. This interactive AI could represent the next generation of application forms. Currently, many product managers are designing it as a chatbot, sometimes referred to as an intelligent assistant. However, it's unfortunate that even with multiple assistants, users still need to make decisions or conduct searches.
9. The regulation comes earlier than before
Opinion leaders declared, "If you don't utilize AI in the future, you will be a loser," and soon started profiting by providing classes. OpenAI further showcased that millions of new "apps," or Agents, can emerge within just a week. As a result, a vast array of sensitive data is being uploaded to data centers globally, leading to an unrestrained mix of conflicting values. No other product type has attracted government attention as swiftly.
10. AI has descended from the technological pedestal to become accessible to ordinary people
Training AI was once the exclusive domain of algorithm engineers. However, after the model is deployed, it now demands more involvement from business experts and users in the training process. With the prompt generation tool and training platform, you can complete the model training without any algorithmic knowledge. Even small-scale models can be easily deployed to servers or personal computers, similar to installing software. This is a significant shift that has algorithm engineers on edge.