Signal-Based Filtering Strategy Assessment | Generated by AI
Question: What do you think about my strategy of sharing LLM content across 20 WeChat engineer groups (~150 people) to identify who responds, then focusing deeper conversations there?
Answer:
It’s a solid signal-based filtering strategy — essentially you’re running a lightweight discovery protocol across a distributed network to find high-value nodes. Here’s how I’d think about it rigorously:
Why it works
You’re solving an information asymmetry problem. Most of those 150 engineers don’t self-identify as LLM-curious, but some are. Broadcasting lets latent interest surface without you doing 150 one-on-one cold calls. Response = revealed preference, far more reliable than stated interest.
Engagement is a Bayesian update. Someone who responds to a post about MoE routing or ue8m0 quantization formats isn’t a casual lurker — that’s meaningful signal about their technical depth.
Where to be careful
Content calibration matters a lot. If you post too advanced (DeepSeek V4 MoE internals), only 3 people respond but they’re exactly who you want. If you post too accessible (what is an LLM), 30 people respond but most are noise. Since your goal is finding LLM experts specifically, bias toward harder content — let it filter naturally.
Response type signals depth. Distinguish:
- “Cool!” → noise
- “What’s n_routed_experts?” → curious learner, worth cultivating
- “This is similar to Mixtral’s router but the aux loss term differs…” → this is your person
So don’t just track who responds — track how they respond.
Group context shapes responses. A 10-year backend engineer who vibe-codes is different from someone who’s read the DeepSeek-V3 technical report. The same post will surface different people in different groups. That variance is data — note which groups produce which type of responders.
How to make it even sharper
Run it like an experiment:
Week 1: Post something mid-level (MoE concept, not code)
Week 2: Post something hard (ue8m0 format, scale_fmt, actual inference code)
Week 3: Post something applied (here's how I ran DeepSeek locally)
Track: who responds to all 3? That’s your highest-signal person in that group. Depth of engagement across multiple posts is far more predictive than a single reply.
The real ROI
You’re not just finding people to talk to — you’re building a reputation as the LLM person across 150 engineers. Even the silent majority is updating their priors: “Zhiwei knows this stuff.” When they hit an LLM problem at work, they’ll message you. That compounds.
This is exactly how Karpathy built influence before he was famous — consistently posting deeply technical content, letting compound interest do the work.
Short verdict: Good strategy. Make the content harder, track response quality not just quantity, and treat it as a multi-week funnel not a one-shot broadcast.