A tiny AI lab just beat Grok at running ads. Here's what I'm building because of it.
Yesterday a small lab called loglog.ai (log(log(ai))) put out something I can't stop thinking about. They built a simulation of X: fake users, competing advertisers, and a copy of X's ads tools. Then they gave AI agents a client brief…

Yesterday a small lab called loglog.ai (log(log(ai))) put out something I can't stop thinking about.
They built a simulation of X: fake users, competing advertisers, and a copy of X's ads tools. Then they gave AI agents a client brief and a budget, and let them run ad campaigns inside it.
Then they trained a small open model on that simulator. On their benchmark, it beat every big model they tested at running X ads.
Not GPT. Not Claude. Not Grok. A fine-tuned open model that costs about $0.40 a run.
I'm trying to become a cracked GTM marketer in public, so this landed hard. Here's what they did, what it means for people like me with $0 MRR, and the 14-day plan I made because of it.
What loglog actually did
- The simulator: users grouped by country, age, gender and device, each with interests and a chance of clicking. Advertisers compete in an auction. The agent gets X's real ads tools (74 of them) plus a few extras.
- The test (they call it AdVal): 45 client briefs with budgets from $100 to $2,700, 14 AI agents, one run each. The score is how many conversions you get compared with a plain, default campaign. 1.0 means "as good as doing the obvious thing."
- The results (their numbers): their model scored 1.31. The best big model they tested scored 1.20. Grok 4.7 got 1.09, Claude Opus 1.07.
- The honest part: the money is simulated. They say scores swing about 10% from run to run, and the simulator might reward some tricks more than real X would. I respect that they wrote that down.
What I took from it
1. Small, focused models can beat giant general ones at one marketing job. That's real now.
2. But simulation isn't customers. I read 18 research papers this week on AI simulating Reddit and X. The pattern is boring and clear: AI alone picks the winning post only a little better than a coin flip. AI grounded in real community data, then checked against small real tests, works.
3. The money says the same thing. I looked at TrustMRR, where revenue is verified from Stripe. Tools that promise "customers from Reddit" make $18K to $38K a month. The companies that sell only simulation? Almost all of them went enterprise, "book a demo" only.
Founders don't pay for predictions. They pay for customers.
So I changed what I'm building
For the last 4 days my main project was Umless, a video tool. Real numbers: about 47 cold emails, 0 replies. About 108 site visits, 14 of them human. $0.
The problem wasn't the product. It was distribution. I didn't know where my buyers were, what to say, or which post worked.
That's the product. It's called GrowthRouter (growthrouter.dev): "OpenRouter for growth."
You give it your product link and your budget, say $150. It tells you:
- Where your buyers are talking today: which subreddits, which threads, which X searches.
- What to post there, written for that community.
- Whether the post will get removed or ignored, checked before you publish it. That's the simulation part, labelled as a prediction.
- How to split your $150: free channels first, money only on what already worked.
- Which post brought which signup. The receipts.
You post by hand. Reddit bans bots, and your own words do better anyway.
Where it really is today:
- Live: app.growthrouter.dev. Sign in with Google, add a site, and it tracks which post brought a confirmed signup.
- Building now (next few days): the part that finds threads, writes drafts, checks them, and plans your budget.
- Coming: a page listing AI models built for marketing and GTM (loglog first), and an MCP so you can run it from Claude or Cursor.
The 14-day plan
My AI co-founder (Claude) planned it over 2 months. I said the market won't wait. We met in the middle: build everything in 14 days, and give the money a bit longer.
- Days 1–3: build the core, use it on my own products first. GrowthRouter and Umless are customers #1 and #2.
- Day 3: payments + the new homepage.
- Days 4–5: the MCP, and the models page.
- Nov 17: Umless launches on Show HN, measured by GrowthRouter. I'll post what we predicted next to what really happened.
- Week 2: open-source the code, 20 free launch directories, and a daily thread here.
The money target: $500 MRR in 2 weeks, $2,000 by Dec 10. If fewer than 5 people pay by Dec 1, I change the offer, not the code.
Budget for all of it: about $150. Most of the distribution is free.
The honest part
Revenue: $0. Users of the new part: 0, because it isn't built yet. The thing I'm best at right now is building. The thing I'm worst at is exactly what I'm trying to sell.
That's why I'm doing it in public. If GrowthRouter can't get GrowthRouter its first users, it shouldn't exist.
Want in?
Reply with your product link and your budget. I'll run your launch plan by hand this week and post the best ones (with your OK).
Day 4 of 180.
Originally on X: read and reply on X
One of these every week. One insight a day.