Vice President of Finance

OpenArt, an AI platform that turns text prompts into videos, images, and music, has hired former Salesforce and Meta marketer Scott Larson as vice president of brand marketing. Larson says he wants to make it "as close to a household name as a creative AI can be." The company says it is on track to pass $250 million in annual recurring revenue by the end of 2026. It has grown from 20 to 80 employees over the past year and now has 8 million monthly users. OpenArt recently launched Director, a tool that turns a described idea, visual style, and narrative arc into videos of up to five minutes. A cinema and billboard campaign promoted it, aiming to spark a "vibe directing" craze. Earlier this year, OpenArt hired former Google executive Raj Gajwani as chief business officer.

The best AI model is the one that fits your workflow. Generic leaderboards are useful for orientation. They are not a substitute for testing the work you need to finish. OpenArt just launched a leaderboard for creative AI models that asks a much better question than "Which model is best?" Its Arena compares image and video outputs through blind evaluations organized around jobs such as advertising, film, animation, product imagery, editing, and lip sync. OpenArt says the launch includes 28 expert judges and 1,003 tastemakers. The point is not to produce one grand winner. It is to help a creative person find the model that performs for a particular brief. That sounds like a small change in presentation. It is not. The AI tools market has spent years teaching SmarToolbox to compare models as if they were smartphones: one ranking, one score, one winner. But most people do not need "the best model." They need a model that can follow a character across shots, preserve product details, return a usable research brief, or turn a rough idea into something they can actually publish. The useful unit is moving from the model to the job. A leaderboard becomes useful when it resembles your Tuesday. OpenArt's own explanation gives a good example. A model can make a beautiful image and still get the product wrong. A video can look cinematic while ignoring the requested camera direction. A talking character can lose its identity between shots. Those are not benchmark footnotes. They are failed deliverables. If you are making an ad, "looks good" is not enough. The output has to respect a product brief, keep the logo or packaging stable, and survive the next edit. If you are building a YouTube channel, the best model may be the one that gives you consistent characters and workable motion, not the one with the prettiest single frame. That is why task-specific evaluation is more useful than a universal score. It forces the test to include the annoying details that make work real. The same idea is showing up outside creative tools. The APEX-Agents research benchmark was built from work by investment banking analysts, management consultants, and corporate lawyers. Its 480 tasks are spread across 33 simulated worlds, and professionals estimated that the tasks would take 11 to 22 hours to complete. Agents have to work across files and applications instead of answering a tidy question in a blank chat window. The headline result is almost humbling: the leading system scored 24% on Pass@1. That is not a sign that the benchmark failed. It is a sign that realistic work is harder than a clean prompt suggests. For a solo creator, the lesson is simple: test the thing you actually do. Give an AI tool your real brief, your real source material, your real formatting constraints, and your real definition of "done." A model that wins a public leaderboard but creates cleanup work may be the wrong model for you. The workflow around the model decides more than the model name. The Pragmatic Engineer's reporting on Codex describes OpenAI using the coding agent beyond ordinary code completion, with automated feedback loops and a growing need to rethink review. OpenAI's own cookbook shows the pattern in a smaller, clearer form: review an artifact, repair it, validate it, then use the result as the input to the next pass. That loop matters because AI output is cheap. Reliable output is not. A model can draft an article, generate a script, or write a pull request in seconds. The expensive part is checking whether the result obeyed the brief. Did it preserve the important facts? Did it break the build? Did it invent a source? Did it quietly change the tone? Did it create a file that looks finished but cannot be used? A good workflow makes those questions executable. It adds a checklist, a test, a linter, a blind comparison, or a human review step. The model is still important, but it is no longer the whole product. This is also why agent demos can be misleading. The impressive part is often the visible action: the agent opens a browser, edits a document, or calls a tool. The durable advantage comes from the less glamorous layer around it: permissions, state, feedback, and a way to detect failure. For my own projects, I would rather have a slightly less capable model inside a workflow that catches mistakes than a stronger model operating with no receipt for what it changed. Memory is part of the test, not a bonus feature. Long-running workflows expose another weakness in model comparisons: the model may perform well on a fresh prompt but lose the plot after a week of work. The answer is not always a larger context window. A March 2026 paper called Memori tested a persistent memory layer that converts conversations into structured facts and summaries. On the LoCoMo benchmark, it reported 81.95% accuracy while using an average of 1,294 tokens per query, about 5% of the full context in its comparison. The paper also reports more than 20 times lower token use than passing the entire conversation history. Those are results from one research system, not a universal promise. But the architectural point is easy to understand: memory is a data problem as much as a model problem. If an AI assistant helps you manage content, it needs to remember the difference between a source you verified and a claim you rejected. It needs to know which tools you already tested, which tone you prefer, and which tasks still require your approval. Dumping every previous conversation into the next prompt is a poor substitute for keeping that information in a structured, inspectable form. So add memory to your evaluation. Ask the tool to continue a project after a few days. Change one requirement. Introduce an old draft and a newer correction. See whether it retrieves the right context or confidently resurrects something you decided not to use. Build a tiny benchmark before you switch models. You do not need a research lab to compare AI tools properly. You need a small set of tasks you repeat often. For a creator, that might be turning a source pack into a clean article outline, rewriting a paragraph without changing its claims, producing five usable hooks for one audience, generating a thumbnail concept that keeps the same subject across variations, or updating a published draft while preserving links and formatting. Run each task through two or three tools. Blind the outputs if you can. Score them on the things that create downstream work: factual accuracy, instruction-following, consistency, edit time, cost, and whether you would publish or ship the result. Do not give every category equal weight. If factual mistakes are expensive, accuracy should dominate. If you are making short videos, character consistency may matter more than a model's general image score. If you are routing an agent through paid APIs, correction burden belongs in the cost calculation. The score is not meant to be scientific. It is meant to stop you from choosing based on launch-day excitement. The best AI model is not the one with the highest general score. It is the one that produces the most useful result inside your workflow, with the least correction and the clearest path to checking its work. My take: stop asking which model wins. Model rankings still have a place. They are a good way to build a shortlist, discover a new tool, or notice that a category is moving quickly. OpenArt Arena is useful precisely because it narrows the question from "best creative model" to "best fit for this kind of creative work." But a leaderboard should be the beginning of your test, not the conclusion. AI products are competing on workflow management now: background execution, memory, tool use, evaluation, and safe handoff. That creates an opening for individuals and small teams. You can build a better system without training a better model, simply by defining the task clearly and measuring the result honestly. This week, take one recurring task and make a five-case benchmark for it. Keep the prompts, inputs, outputs, and scores. Run it again when a new model arrives. That small habit will tell you more about an AI tool's value than another hour spent scrolling through a leaderboard.