Joe Salvatore on AI, Creative Direction and the Taste Behind a Brand

Joe Salvatore, a visual artist who creates brand content, joins Reid Hoffman and Parth Patil to discuss using AI in creative work. Salvatore says the tools can take on production, but the creator still has to supply taste, articulate direction and decide what is worth making.

The episode introduces the token grantee program, which gives selected creators $1,000 a week in model tokens, freedom to choose their tools and access to custom agents. The stated goal is to shorten the distance between an idea and a finished work.

Patil met Salvatore at the Lighthouse, a creator studio in Venice, Los Angeles. Coming from visual storytelling and brand assets, Salvatore was already deeply engaged with image and video models. Patil began asking him about models, workflows and how separate images could support a coherent brand idea.

Patil was curious what would happen if a visual creator without an engineering background had access to coding agents. Salvatore says he had spent roughly four years experimenting with creative tools after seeing an AI-made trailer that mixed *The Lord of the Rings* with a Wes Anderson aesthetic.

Salvatore says Patil encouraged him to step back from making one piece at a time and orchestrate a broader process. That shift let him explore parts of his own thinking that had previously been hard to access. He connects the experience to Hoffman’s idea of augmented intelligence.

His first experiments were with Midjourney and image generation, driven by curiosity rather than a business need. Over time, he explored video tools and found ways to apply them to his work making brand content; he now uses Codex to automate static advertising production.

The skill that became apparent was the ability to articulate a creative idea. Salvatore says the output exposes the quality of his direction: a half-hearted brief produces poor work. He has moved from controlling one image, to groups of images, to systems that can produce hundreds.

For a larger run, he specifies the creative process itself: a brief, the brand elements to extract and the steps the system should follow. He compares learning to building a house, starting with one image, then two, then scaling the structure.

Patil describes the shift as moving toward a cohesive aesthetic or universe across many generations. Salvatore says node-based visual workflows, such as Weavy, can help creative professionals see how inputs and intermediate steps connect to a large set of outputs.

Once the path becomes familiar, Salvatore says, he can describe the workflow to Codex, Claude or another agent without relying on the visual nodes. Patil emphasizes that the result comes from chaining models and feedback rounds, rather than asking for one generation and stopping there.

Asked for practical direction, Salvatore recommends constraints. Give the system too much latitude, he says, and frustration can grow downstream. He starts by spelling out what makes a good image for him, so the tool has a clearer target.

Hoffman adds that models can take on roles. When working on *Superagency*, he might ask an AI to act as a historian of technology, identify what is right or wrong in an argument and propose elaborations or objections. The goal is a more useful exchange than a generic request.

Salvatore agrees that he gives a model a role, references and boundaries, much like a creative brief: who it is, what it should do, what it must avoid and what context the product needs. Specific language at the start narrows the range of possible results downstream.

Patil contrasts creative evaluation with an app feature that either runs or does not. Aesthetic work is subjective, and judging it requires experience. He asks what is counterintuitive about agents when the definition of a good result belongs to the creator.

Salvatore says an agent may announce that a generated ad looks great whether it succeeded or produced poor work. He can outsource production, but not his taste: a creative professional still needs to intervene and judge the output.

Taste develops through creative repetitions, Salvatore says. He spends more time in pre-production looking at designers, magazines, old print advertising, Cosmos and Pinterest. The work is to identify a coherent theme in those references and communicate it precisely.

Hoffman asks for a concrete workflow. Salvatore describes a campaign for Ketone-IQ, a beverage brand. He uses Hermes for research and Codex or Claude for creative work; in this example, Codex handles the production workflow.

First he asks Hermes to research 20 static or print advertising styles across decades, from the 1940s to contemporary social media. He wants case studies and templates, along with an explanation of why each approach might work, why it might fail and what to change.

He takes the research into Codex, connected to creative tools through fal.ai, and supplies the product image and campaign copy. He says he uses GPT Image 2 for this work. The inputs are the selected template, the bottle and variations of the promotional text.

Rather than run every template immediately, he tests two or three first. His existing creative-director skill pulls the components into the image workflow, and he grades the result when it returns. Once a few runs are working, he asks for ten variations of each of the 20 templates.

He lets the larger batch run overnight and finds it finishing when he wakes. Patil sees a system built around the creator’s direction. Salvatore says he came to the work through creative curiosity and enjoyment, while acknowledging the pushback that AI tools receive in creative fields.

Patil describes a shared Hermes agent that he and Salvatore use through Telegram. It has helped with graphic novels and acquired material about Salvatore’s advertising process. Patil sees the agent as reflecting the people who work with it, with skills manually packaged and passed to other agents.

Salvatore gives another example: a daily newsletter for him and Patil delivered at 8 a.m. He asks it to cover AI and technology developments from the previous 24 to 48 hours, then consider what the news might suggest about coming releases and trends.

Patil says the brief gives him a starting point for thinking about what developments mean. Salvatore says his most valuable skill is clear articulation. Repeatedly explaining ideas to tools has, in his experience, also improved his communication with people on photo shoots and other projects.

Hoffman asks which workflow took effort to learn and how it works with other people. Salvatore says automation clicked only after visual experimentation: he needed to see the intermediate steps before he could confidently ask an agent to execute the whole process.

In consulting with brands, he now uses visual Weavy workflows as a teaching device. Seeing how the process goes from input to output helps clients understand an otherwise opaque automation. Patil calls this showing the system before asking people to trust it.

Salvatore asks what professionals should practice as language becomes a way to bring ideas into the world, invoking Wittgenstein’s thought about the limits of language. Hoffman says both imagination and the expression of imagination can be developed through repeated work.

Hoffman suggests asking an AI for exercises and feedback on how to express an idea better. Instead of requesting a generic script, a creator might explore an artistic perspective, reject the result if it is poor, then try another reference. The iterations can enlarge the creator’s expressive palette.

Specific references, Hoffman says, can communicate a richer context quickly. Salvatore likens looking back at advertising from earlier decades to a kind of time travel: creators can draw on work made by people who are no longer present to address a new problem.

Asked what creative and technical work might look like in two or three years, Patil offers a prediction, not a settled outcome. He expects software to become cheaper and quicker to create and modify, producing interfaces and tools tailored to a particular task.

As an example, Patil says he had Codex build a version of a visual workflow interface for his video-model pipelines. He imagines much faster inference making software more fluid: an agent could create or modify a small tool when a project needs it.

Patil argues that system design and storytelling grow more valuable when models can generate code or video. Image and video tools may give many people access to production capabilities, he says, but that does not automatically give them the ability to tell a story someone wants to revisit.

He says he now studies films and graphic novels more closely, paying attention to dialogue and individual frames. Salvatore suggests that creative professionals need to learn to direct; Patil extends the idea to directing many kinds of work, not just film.

Salvatore returns to Wittgenstein’s lion and asks whether an AI saying “I understand” means the same thing a person does, or whether it is merely mirroring language. The question opens a discussion of meaning rather than a claim that the issue has been settled.

Hoffman recounts discussing Wittgenstein with philosopher Bill Child. He describes questions about truth, language and forms of life, and asks how they apply to conversation with AI systems. In his view, the practical exchange suggests there is something worth examining in the notion of understanding.

As a metaphor, Hoffman describes AI models as something like alien forms trained intensively on human material. He says their mixture of familiarity and difference raises interesting questions about language and knowledge. Patil thanks Salvatore for joining the conversation.