AI Gold Rush: To Dig for Gold Oneself or Sell Shovels?

Our team's AI projects have all come to a halt.

For example, the core project framework Botastic stopped receiving updates five months ago. Another closed-source commercial application project has also been paused.

Just a few days ago, a friend came to me for advice. He has an AI application project but is uncertain about its future prospects. So, I shared some thoughts, which are essentially why I paused all AI projects.

The information released at the OpenAI Dev Day essentially confirmed my thoughts.

The Grind

If we focus solely on pure AI projects, we must describe the phenomenon as "the grind" — "intense competition" — the gold mine has attracted many gold diggers, and there's no shortage of merchants selling shovels.

In this gold rush, although the chip sector is still dominated by Nvidia (at least for now), other AI fields are overcrowded.

The other side of the grind is "narrowness." Such a vast application market has only unearthed a handful of viable commercial scenarios, making the space particularly narrow.

First, what comes to mind is Prompt engineering. Although initially, most Prompt products merely listed useful Prompts, Prompt engineering will undoubtedly become part of the workflow in the future, and it will have a strong correlation with the LLM model used. While there are barriers, it's not so technical... It's not so technical that it becomes a subject of ridicule.

Then, the most straightforward and brute-force application is AI replacements for classical functions. For example, the emergence of numerous foreign language learning Apps, which replace difficult dialogue scenarios with AI characters. As typical shell applications, converting OpenAI or other suppliers' wholesale APIs into retail lacks technical barriers and innovation.

Next is RAG, using LLMs to interpret services based on existing content, such as customer service robots (including our PAL9000), which also lack technical barriers.

Tool products, like ChatPDF, or entertainment products, such as the MiaoYa Camera, or those offering emotional value, like janitor.ai. When foundational service providers like OpenAI become product companies and start making products for end consumers, their situation will probably become very delicate — there's only so much time users can spend, and everyone wants to be the entry point.

Agents or assistants, I think, have a good shot, but I believe the challenge is not AI but user experience, which needs to be polished slowly.

In short, either you need to put more effort into other aspects, like improving user experience or protecting customer relationships — AI is a silver bullet, but it's not our silver bullet; or you must display a unique competitive advantage in AI technology itself, competing in computing power and data, embarking on a thorny path.

The Poverty

Take the development of an English learning application as an example, making the App's cost per token lower than OpenAI's investment in Speak might be a daunting task. Not only does the cost need to be lower, but whether the LLM's response speed can surpass Speak seems even more challenging.

Or, if planning to make a Code Assistant, then can it surpass the commit messages generated by Github Copilot in speed? More importantly, can it ensure that the generated code is of higher quality than it?

If you're just using these AI suppliers' APIs and services, you'll inevitably find yourself competing with them on price. If you plan to cultivate your LLM, then competition in computing power and datasets with them is also inevitable.

So, What to Do?

The first major decision is whether to go down the consumer product route or to target the enterprise market.

coverage, potential income fluctuation | High value per customer, stable income |
| Customer Loyalty | Low, easily influenced by new products and trends | High, difficult to switch suppliers after establishment |
| Initial Investment | Relatively low, quick to market | Higher, needs targeted development and adjustment |

What to do specifically has no universal solution and depends on each company's resources and characteristics. However, in general, you can consider the following conditions:

Teams with a consumer product DNA should focus on consumer products

At this stage of AI consumer products, with low barriers and no pain points, it's hard to win widespread favor from users due to any unique features, leading back to a competition of hard product power.

On one hand, consumers' pursuit of experience is limitless. On the other hand, a large portion of consumers' purchasing decisions contain irrational elements. Only teams with a consumer product DNA can do well in amplifying humanity's base instincts.

If you can seamlessly integrate AI into existing user experiences, consider making a consumer product

There was a recent discussion about whether chat/conversation-based interfaces would replace GUIs in the AI era — these discussions always end the same and are somewhat meaningless: some can be replaced, others cannot. What matters is which can and which cannot.

If we view software as a collection of functions, then the operations of these functions can be seen as a process tree, where the current GUI represents a series of slices on this tree.

The advantage of GUI is its stability. A good GUI can reassure users: what they want is right there.

The disadvantage of GUI is the flip side of "stability," meaning its information architecture is rigid, making it difficult to provide the optimal process based on context.

Before the AI era, some scenarios were "conversation-based," but they were either not smart enough — unable to make good inferences from context and lacking knowledge to provide good answers — like AI customer service; or they were small-scale, short processes — like search and recommendation, which are used and then left behind.

Looking back, although a portion of GUIs will be replaced by conversation-based interactions in the future, most GUIs cannot be replaced. However, this doesn't mean AI can't play a role. With AI, it's possible to perceive user needs that were previously hard to reach and optimize processes that were costly to optimize before.

If you're already in a regulated market, consider making an enterprise product

Well, of course.