Riding the AI Storm: My Startup’s Wild 2-Year Journey Through Innovation and Chaos! (II/IV)
OpenAI’s Ambition and the Endgame of the AI Race
Given OpenAI’s recent, rather dramatic internal shenanigans, discussing their ambition comes with a caveat: we’re talking about the OpenAI under Sam Altman’s reign. Again, I’m not here to lecture on the tech details — that’s been done to death by countless newsletters. This is all about my take and speculations.
After the early-year hiccup with Plugins, OpenAI dropped a bombshell in June with Lilian Weng’s paper on the Agent framework, opening a new chapter in AI.
If we liken large models to the brain’s frontal lobe, handling computation, then Agents are more like the whole brain — they’ve got memory, planning, action, tool-using abilities, including search, calendar management, and more. You can feed them knowledge and documents to better represent you, or train them to autonomously write code and tackle complex tasks. Compared to large models, Agents are more like a fully-fledged app. Imagine prepping for a gym session. Normally, you’d fiddle with your fitness app, selecting duration, body parts to work on, equipment, difficulty level… and before you know it, minutes have vanished, especially if you’re indecisive or unfamiliar with the app. In the Agent era, your virtual fitness coach plans your workouts, reminds you to exercise, demonstrates in real-time, corrects your form, and even orders your post-workout meal, guiding your diet. Human-machine interaction in the Agent era is as natural, immersive, and personalized as human-to-human interaction.
Bill Gates’ recent comments on AI agents — preference plus automation as the basic closed-loop elements — underscore this future.
With this background, OpenAI’s launch of the GPT Store a few weeks ago clearly revealed their ambition: to build a super empire of API + OS + Hardware + Chips.
A super empire of API + OS + Hardware + Chips
In the current mobile world, every app can still embed GPT as a chatbot, kind of like Copilot. But let’s be honest, Copilot isn’t the bee’s knees because you still have to dive into the app. Plus, each Copilot is stuck in its own little walled garden — Taobao’s got shopping data, Feizhu’s got travel data. As I mentioned at the start, Copilot is just the awkward teenage phase of AI.
Then comes Phase Three, the arrival of agents, which is going to shake up the app ecosystem like a snow globe.
- First off, the supply is going to balloon like it’s on a strict diet of helium. Technology’s job is to pump up the supply, just like the number of apps in the smartphone era made the PC era look like a garage sale. In the AI era, the number of agents compared to apps is going to skyrocket.
- But the more the supply, the more we lean on centralized distribution. So, the future of distribution rights is going to be a bigger deal than iOS.
At the same time, the AI era is going to spin a never-before-seen flywheel — Data to Intelligence. Businesses and developers, in their quest to create AI employees, will upload their services and data; users, in chatting with AI, will also help it level up.
In summary, the OS of the AI era is going to have stronger barriers than iOS. It’s like iOS is a garden fence, and AI OS is the Great Wall of China.
The Paradox of GPTStore
To reach this utopian stage, GPTStore has a long road ahead, littered with the unavoidable potholes of any platform — economic mobility. Is the Agent market going to be decentralized or centralized?
Today, GPT’s abilities have clear limits: 8000 tokens — and while that’ll improve, it’s not going to be infinite. We found earlier this year that feeding more data into GPT creation can actually make it less expressive, like overfeeding a goldfish. This means an Agent can’t be a jack-of-all-trades. You’ll need multiple agents, just like the world has many financial advisors, legal consultants, each with their own specialty. From a supply perspective, we’ll need a whole army of agents, each focused on a specific area, even within niches like financial advising.
But a person’s attention can’t be spread across an army of agents. In the mobile internet era, 90% of apps were opened just once and then forgotten, like a bad first date. And compared to the straightforwardness of apps, users have to chat with agents to discover their capabilities — a higher barrier to entry. So, from a demand perspective, agents must be more centralized.
This means a lot of agents will be left in the cold, unnoticed. Unless agents can automatically collaborate. Imagine a fitness coach agent automatically teaming up with a dietary specialist agent to develop and guide a user’s health plan.
But this brings two problems:
- Costs will skyrocket, as the market will be flooded with low-quality UGC agents, making the fitness agent cast a wide net in GPTStore, issuing the same prompt to filter out suitable dietary agents;
- Monetization will be challenging. A subscription model won’t work well with a collaborative approach (many agents are one-off uses). A more viable model might be service commissions or pay-per-conversation. Commissions are tough in the short term, as AI services can’t yet close the loop. Pay-per-conversation is possible, but ROI is low, and revenue potential limited (imagine Apple not charging tolls but by traffic).
Moreover, such a supply-demand imbalance means opportunities for other platforms, just like Amazon coexists with Shopify, different distribution mechanisms supporting different platforms.
So, what’s the endgame for AI going to look like? Continue to read Part III.
Please read through part II — IV here:
More about me:
- Product homepage: https://mindos.com
- YouTube: https://www.youtube.com/@Mind-OS
- Twitter: @KissonL
- LinkedIn: https://www.linkedin.com/in/kisson-songqi-lin-49309516/
- Email: k@mindverse.ai