Kinda sure that Google is paying some people to talk up their models. Too much of a coincidence that anyone with 10K+ followers and AI in their bio is “blown away” by Neon.
The biggest risk in your social strategy is the slow and invariable drift into sounding like Claude or ChatGPT.
You know what I’m talking about.
“It’s this, not that.”
“The thing worth sitting with…”
“That’s the real kicker.”
It’s words that mean nothing. People subconsciously gloss over that stuff because their brain files it under “not real”.
But look, if for whatever reason, you’ve decided that writing is not for you… and you want to use AI, you can smart about it.
There are physical limits that govern how much a mass market model can sound like you. Yes, you can give it good grounding, writing samples, etc.—it’s better than nothing, but it still fails the authenticity test.
A better approach today is fine-tuning a model on your writing corpus. You needed to absorb big infra cost if you wanted to do this a couple years ago.
Now, you can train a 3.5-5B parameter model under a dollar and in 10 mins or so. @lettercompany just released this as one of their features, so go check it out!
How is this different? Well, fine-tuning physically shifts the weights of a model, so instead of getting a close approximation of someone’s style… you get a true mimicry of it from the training corpus being encoded right into the model’s neural network.
I’m going to be training one using my published works soon and will share the results and side-by-side comparisons. Stay tuned.
I used GrokBot to fine-tune an AI model that predicts whether or not a given website uses an adblock recovery vendor... for $0.82.
The task: given a publisher's category, organic traffic, ad servers, supply paths, consent tool and detected vendors, predict whether it runs an adblock recovery tool. I stripped the adblock vendors out of every profile, so the model can't read the answer.
The data comes from ~50k publisher sites MediaContext has scanned. I asked Grok Bot to build a balanced set of 1,000 training rows and 200 test rows, then fine-tune Qwen3.5-0.8B through Overmind.
On the 200 held-out sites, the base model got 51% right, which is a coin flip on a 50/50 set. After about nine minutes of training on one H100, the fine-tuned model got 89% right. The run cost $0.82.
For an adblock recovery vendor, the output is a ranked list of publishers that look like existing customers but don't run a vendor yet. The same configuration works for any vendor category MediaContext detects, from consent tools to identity providers.
Go train your model on @OvermindLab 🚀🚀🚀
Turn Grok Bot into your own AI lab, right from your phone.
• Connect @bot to the Overmind MCP
• Upload a dataset of examples to learn from
• Tell it what to learn and let it fine-tune the model
Over to you: https://t.co/RM90VWzVaY
VJ’s take: this could work for publishers, brands, and people looking for useful information. The bigger question is whether it helps make media more sustainable.