I fed an AI 4,580 of my social media posts. Roughly 650,000 words of my life, spanning five years of thoughts, projects, and late-night ramblings. Then I asked it for career advice.
Its first sentence hit me like a bucket of ice water: “You’re not short on opportunities, you’re spreading your leverage too thin.”
I didn’t get a generic list of SMART goals. I didn’t get a LinkedIn-flavored pep talk. I got a brutal, hyper-specific reality check based entirely on my own digital footprint.
You’ve probably noticed the AI arms race lately. Everyone from OpenAI to the Chinese labs is obsessed with making models bigger and smarter. We are drowning in SOTA benchmarks, parameter counts, and monolithic architectures that can write a symphony but still give you completely generic life advice.
The AI industry is obsessed with making models smarter. But nobody is asking how to make them actually know you.
Enter Macaron V1. Instead of chasing the top of the leaderboard, they took a completely different path. They took a massive 744B parameter base model (GLM-5.2) and froze it completely. Then, they attached four tiny 1B parameter “experts” to it. They call this Mixture of LoRA (MoL).
Traditionally, LoRA—Low-Rank Adaptation—has been treated as a cheap patch job. You use it to teach a model a new trick without paying to retrain the whole thing. But Macaron flipped the script. They elevated LoRA from a temporary hack to a permanent architecture.
Why? Because the base model already has all the general intelligence it needs. What it lacks is *you*. Your preferences, your habits, your history. By keeping the base frozen and letting the small experts specialize (one for chat, one for life tasks, one for code, one for UI), the model stops being a generic calculator and starts becoming a personal companion.
When I tested this, I didn’t run coding benchmarks. I ran my life.
I tried to trick it. I asked it to find a post where I claimed to have stayed up all night hand-coding an app in Swift. It’s a false memory—I don’t code. Instead of hallucinating to please me, it searched the 650,000 words, told me my memory was flawed, and pulled up the exact post where I actually used an AI tool to build it in an hour.
Then came the advice. It pointed to specific projects I had abandoned. It called out my scattered focus using my own data. It didn’t stroke my ego; it held up a mirror.
When a model hasn’t read your history, its advice is a generic template. When it has read 650,000 words of your life, the answers actually start to cost you something.
This is the tension the industry is missing. We are racing toward Artificial General Intelligence (AGI) while ignoring Personal General Intelligence. A model that can pass the bar exam is useless to me if it doesn’t know how I work, what I’ve built, and where I’m wasting my time.
Macaron V1 isn’t perfect. Its complex reasoning is slower, and it’s built for agentic workflows rather than quick chat replies. But the direction is undeniable. They are building systems that can accumulate context over weeks, remember the tools they built yesterday, and evolve alongside the user.
General intelligence is a leaderboard. Personal intelligence is a mirror.
We’ve spent two years asking AI to be smarter. It’s time to ask it to be honest, to be specific, and to actually know who we are.
The era of the one-size-fits-all chatbot is ending. Give it a few years, and everyone will want an AI that has read them.
FAQ
Q: Isn't this just a fancy context window?
A: No, it's an architecture shift. Macaron turns LoRA from a cost-saving hack into permanent memory, accumulating your preferences over time, not just reading a long prompt.
Q: What can I actually do with this?
A: Stop asking generic chatbots for life or career advice. Feed this model your notes, journals, or work files, and ask it to roast your strategy. The answers are tailored, not templated.
Q: Are SOTA benchmarks completely useless then?
A: For building general infrastructure, no. But for consumer-facing personal assistants, they are a massive distraction. A model that scores 5% lower but knows your entire digital history will beat a SOTA model every single time.