Why AI has no taste and how to fix it (w/ Thais Castello Branco) | E2319
Summary
This episode centers on the idea that AI is excellent at objective problems like math and coding, but still produces bland, average, or aesthetically weak results in subjective domains like design, writing, and content creation. Thais Castello Branco explains how Taste Labs is trying to fix this by using expert human evaluators, or "TasteMakers," to provide preference data and improve model outputs. The conversation also explores whether AI is accelerating the "half-life of cool" by flooding culture with derivative content, while simultaneously increasing the value of taste and curation. Later segments broaden the discussion to platform responses to AI slop, including anti-AI moderation on LinkedIn and Substack, and the growing impact of autonomy and robotics on labor, especially driving-related jobs.
Key Takeaways
- 1Frontier AI models are powerful at objective tasks, but still weak at subjective ones like aesthetics, tone, and originality.
- 2Taste Labs is building tools to improve both foundation models and application-layer products using human preference data.
- 3A network of 'TasteMakers' is being used to curate, critique, and rate outputs so models can learn what high-quality work looks like.
- 4As AI makes mediocre content abundant, taste and curation may become more valuable rather than less.
- 5Platforms are starting to treat AI-generated content as a feed quality problem, not just a productivity feature.
- 6Autonomy is becoming easier to prototype, but large-scale deployment will still depend on regulation, licensing, and labor dynamics.
Notable Quotes
""You have models that can solve cyber hack and solve peace, your level of math problems and they can't write a good tweet or make a good design.""
""The average is that thing that you're going to kind of not even take a second glance at that you've seen a lot.""
""We pay them, of course, for their time, but I think we've also been exploring some more interesting ways to collaborate with them, both like evolving them more deeply in the company.""
""I think taste makers are going to go away by any means; I actually think it's the opposite.""
""LinkedIn added a report a i slot button and yes and they've gotten rid of the let a i help you write your post button that they put in that originally that's gone.""
""if anything scores over 90% over 50% on a i they should grade out in the feed not put it in the feed they should um what was the term they should shadow banish shadow ban posts""
""the first wave of people losing their jobs to AI is going to be drivers""
""we hit 51% of energy coming from solar and batteries""
Episode questions
Why do AI models that excel at coding and math still produce bland creative outputs?
Because the training objective often pushes models toward the most likely answer, which is useful in objective domains but produces average results in subjective ones. Thais says creativity requires uniqueness, variety, and judgment that are not well captured by simple likelihood-based training.
How does Taste Labs try to improve model output quality?
It uses expert reviewers and 'TasteMakers' to generate ratings, critiques, curation, and examples of good and bad work. Those signals are used both to benchmark foundation models and to improve products built on top of them.
Will AI eliminate taste-makers and cultural curators?
Thais says no; she believes they become more valuable as the volume of mediocre output increases. Jason agrees that curation has historically been important and may become even more valuable in a world of mass production.
What is the main risk of AI-assisted personalization for taste?
If personalization is done poorly, it can still converge on the same generic output for everyone. Thais says the immediate challenge is not perfect personalization, but simply raising overall quality and reducing slop.