Inspired by Karpathy’s 2025 LLM Year In Review, I’ve decided to write some reflections of my own applied to my current area of focus, growth marketing and generative AI

The Floor Is Higher / AI Slop is here

We started to see an influx of AI generated videos across platforms this year and even in TV ads. Text has already hit mainstream over the last several years. It’s clear the age of AI slop is here to stay thanks to generative AI. What this means is that the floor is higher - tools are making it easier and cheaper to write compelling marketing copy, build beautiful emails, design thumb-stopping graphics, and craft campaigns that drive brand awareness or growth. In a world where models continue to improve every 3-6 months, the floor will continue to rise.

Vibe Growth / Vibe Marketing

Karpathy coined the term, “vibe coding” in a tweet earlier this year that defined the paradigm shift within engineering. Vibe only works in the coding sense because you almost forget that the code exists and the underlying logic is abstracted so much that it exceeds your understanding. This does not apply to marketing. While some have tried applying “vibe marketing” to this shift, I think 'velocity marketing' better captures what's happening. You’re getting from idea to execution at a much faster rate with less resources but you should not be letting the LLM drive purely based on “vibes.” Ideas are so cheap now and execution is slowly becoming a commodity.

Taste / What keeps us human

The democratization of execution makes curation more valuable, not less. Taste is what makes us human - we all have preferences and opinions on what looks good and that evolves and grows with each season of life. While AI can now generate photos, AI cannot replace good and opinionated taste. Yes, you can tell it to write a line of copy for a certain segment or create a specific photo with an aesthetic, but you are the one curating and deciding what is good to share to the world.

Agents / Workflows that actually matter

After getting pitched marketing agents and also with some attempts at building my own, agents showed flashes of promise but it was mostly a mirage. It delivered value in specific workflows like sales prospecting, but marketing remains elusive.

In the go-to-market sense, tools like Clay or Unify have unlocked meaningful productivity for sales and revenue teams but we have yet to see the same step changes for marketing. Marketers can utilize LLMs for brainstorming or writing copy but stitching it all together is still quite manual and requires humans in the loop.

The biggest blocker to marketing agents for enterprises (and most agent workflows in general) is verifiability. Code is easily verified through tests and the software fails but marketing does not have that luxury.You can generate 100s of A/B tests an email subject line, but you can't cheaply verify whether a campaign performed well or repositioned your brand in a way that'll pay off six months from now. It’s mostly about taste, delivery, and how the brand shows up within different channels - the time horizon is much longer, the reputation risk is higher, and understanding results is more expensive.

Conclusion

This was a huge year in generative AI which is setting the stage for the foreseeable future. Yes, we'll be drowning in AI slop in 2026. But we're also seeing the floor rise across the board. In a world where everyone has access to the same generative tools, taste and judgment become the only defensible moat. The marketers who thrive won't be the ones who let AI drive, they'll be the ones who use it to move faster while staying opinionated about what's worth creating