A small team has trained an open-weights model to do one narrow, money-relevant job — running ad campaigns on X — and says it beats Grok, the platform’s own AI, at it. Yvligo, from the startup log(log(ai)), is a fine-tuned version of DeepSeek V4.1 Flash trained entirely inside a simulation of X’s ad market called AdVal. “Meet Yvligo, our open-weights model for running X ads. (3x better than Grok 4.7),” co-founder Ted (@notmiramurati) posted on November 9, citing a live test: “Yvligo: 383 relevant website visits for $200. Grok: 123 visits.”
Quick Take
- What it is: A LoRA adapter for DeepSeek-V4.1-Flash-1017, published on Hugging Face under the MIT license, trained to run X Ads campaigns as a tool-using agent.
- How it was trained: Supervised fine-tuning on 312 successful campaign traces by an “expert” agent in the AdVal simulator.
- Simulated results: On 45 held-out briefs, Yvligo scored 1.311x a scripted baseline (geometric mean), ahead of Gemini 4 Argon (1.198), GPT-6.1 Sol (1.145), Grok 4.7 (1.087) and Claude Opus 5.5 (1.068).
- Live result: 383 relevant visits for $200 versus 123 for Grok, according to the makers — a single reported test.
- Caveat: The team calls live A/B testing “a sobering experience” and says the leaderboard will “change quite a bit” as the simulator improves.
What happened
log(log(ai)) introduced AdVal and Yvligo in a blog post that opens with a quote from a marketing team: “You know, we use Grok to launch X ad campaigns but honestly it’s not very good.” The company builds reinforcement-learning environments for business contexts. Alongside the model, it published the AdVal benchmark on GitHub and “the traces of all the models on held-out set,” as Ted wrote. The announcement drew over 400 likes and more than 200 replies, and summaries spread quickly in English, Chinese and Russian AI newsletters.
Background: why simulate X ads?
Running real ad campaigns to train or test agents is slow and expensive. AdVal replaces the real platform with a simulation that agents reach through the official X Ads MCP, which the simulator “serves word for word” — 74 tools, plus four added ones (a platform search, a policy lookup and a clock). The simulation has three parts:
- Audience: The world is divided into cells by country, age band, gender and platform, each holding “atoms” of users with interests and click propensities. Reactions are modeled with mechanical rules and embedding similarity (using EmbeddingGemma 2), rather than LLM-backed personas, so it stays “deterministic and fast.”
- Competition: X’s auction and its automatic bidding, which the team says is not open-source and had to be fitted with live experiments.
- Ad account: Balance, campaigns, ad groups and promoted posts.
The team says it fit the model to data from “1000s of campaigns that we had on our hands” and ran live campaigns to close gaps. Tasks are client briefs covering consumer and B2B products, age-restricted verticals that vary by country, and landing pages with or without the X pixel.
How Yvligo was trained
According to the model card, an expert agent ran 444 campaigns (222 briefs, two seeds each). A run became a training trace only if it finished, scored at least par, did at least as well as the base model and passed the brief’s rule checks — yielding 312 traces. The blog describes the method as “context distillation by rejection-sampled SFT”: DeepSeek with best-practice targeting guidance in its prompt served as teacher, and the base model was fine-tuned on its best runs using Tinker. The student “started overfitting,” so the team used an earlier checkpoint.
| Agent (harness) | Conversions ÷ par (geometric mean, 45 briefs) | Briefs where Yvligo is ahead |
|---|---|---|
| Expert agent (source of training traces) | 1.349 | 47% |
| Yvligo | 1.311 | — |
| Gemini 4 Argon (Antigravity) | 1.198 | 67% |
| GPT-6.1 Sol (Codex) | 1.145 | 73% |
| DeepSeek V4.1 Flash, before training | 1.095 | 80% |
| Grok 4.7 (Grok Build) | 1.087 | 80% |
| Claude Opus 5.5 (Claude Code) | 1.068 | 89% |
| GPT-6 Astra (Codex) | 1.058 | 82% |
| Claude Haiku 5.5 (Claude Code) | 1.012 | 89% |
| Claude Sonnet 5.5 (Claude Code) | 0.930 | 96% |
Scores are expected conversions divided by those of a scripted baseline that runs one ad group at X’s default AUTO bid. Other models ran in their own agent CLIs at medium reasoning effort. Summaries on X cite a different normalization (77% for Yvligo, relative to the best-known result per brief); i10x uses the model card’s table. Across the 45 briefs, Yvligo “never scored under 0.8 of par,” wrote an accurate recap every time, had 2.3% of tool calls return an error, and broke one brief rule — advertising on desktop for a phones-only brief.
Pros and cons
For users
- Pro: Small advertisers could get campaign management that is cheaper than an agency; one commenter calculated “roughly $0.52 per visit vs. $1.63.”
- Con: The live advantage rests on a single reported $200 test; results may not hold for other products, markets or budgets.
For developers
- Pro: Open weights, an open benchmark and published traces make the result inspectable and reproducible — rare for agent claims.
- Con: Running a LoRA on DeepSeek-V4.1-Flash-1017 requires hosting the large base model; the adapter repository alone uses about 13.9 GB of storage on Hugging Face.
For companies
- Pro: The project is a template for building domain-specific RL environments in which a cheap open model can beat frontier generalists on a narrow business task.
- Con: Handing ad budgets to an agent creates spend and compliance risk; Yvligo gated to the brief’s minimum age on only 5 of 11 briefs that set one.
What X is saying
Meet Yvligo, our open-weights model for running X ads. (3x better than Grok 4.7) At loglog.ai, we built an X simulation and post-trained DeepSeek V4.1 Flash, outperforming frontier models. Live results: Yvligo: 383 relevant website visits for $200. Grok: 123 visits
— Ted (@notmiramurati) November 9, 2026 · View on X
Everyone is talking about 3x better than Grok. The real story is HOW. Ted’s team didn’t just fine tune a model. They built a full simulation of X, built their own RL environment + benchmark (AdVal), and trained in it.
— Se Hozaifa (@SeHozaifa) November 9, 2026 · View on X
KI-Vater (@KI_Vater)
— Maurice Ted’s team released Yvligo with public evaluation traces, plus a write-up of where their X ads simulator disagreed with live campaigns. I want to inspect an agent’s decisions before trusting it with an ad budget.|November 9, 2026 · View on X
The most-shared framing was competitive: @tonysimons wrote “GROK JUST GOT COOKED IN ITS OWN KITCHEN,” adding “If these results hold up at scale…. The implications are HUGE.” Others focused on the method: @kaif9998 summed it up as “they built a fake X, let an AI practice running ads in it, then set it loose on the real one.” i10x found no public response from xAI or X.
What Grok says
i10x did not find a direct Grok reply assessing Yvligo or AdVal at the time of writing.
Critical analysis
Yvligo is a credible and unusually transparent result — with a clear asterisk. The simulated benchmark is the team’s own, built on its own campaign data, and the model was trained in the same simulator it is evaluated in (on different briefs). That is standard practice, but it means the leaderboard measures how well agents play AdVal, not X. The authors acknowledge this: the blog lists “every gap we measured between the simulator and our live runs” and calls live testing “quite complicated, with both good and bad parts.”
The headline “3x better than Grok” comes from one live test with a $200 budget, reported by the makers. It is interesting evidence, not proof. Note also that frontier models ran in their own CLIs at medium effort, without the targeting guidance the teacher received — a fair test of out-of-the-box agents, but not of the best each could do with domain prompting.
Still, the broader point holds: Yvligo nearly matched its teacher (1.311 vs. 1.349) and beat its own untrained base model on 80% of briefs. Narrow RL environments plus a cheap open model can beat expensive generalists on a specific job.
What’s still open
- Will independent advertisers reproduce the live advantage at larger budgets and in other verticals?
- How will AdVal scores change as the simulator is updated with live discrepancies?
- Will X or xAI respond, for example by improving Grok’s ad tooling?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis draws on log(log(ai))’s AdVal blog post, the Yvligo model card on Hugging Face, and public posts on X through November 10, 2026. i10x did not run the model or the benchmark. Related: Grok and xAI’s X platform strategy and DeepSeek Harness usage.
i10x Perspective
The interesting part of Yvligo is not that a fine-tuned model beat Grok at X ads — it is how cheaply the advantage was built. A realistic simulator, a few hundred filtered expert traces and a LoRA on an open model were enough to top frontier agents on a narrow task. Expect more of this “environment-first” playbook in marketing, recruiting and finance. The bottleneck shifts from model access to simulation fidelity — and to honest reporting of where the simulator and reality disagree, which this team, to its credit, published.
