Jul 31, 2026 · 1h 15m
Startups build curation engines to teach artificial intelligence human taste
Why AI has no taste and how to fix it (w/ Thais Castello Branco) | E2319
As generative AI floods the internet with generic content, teaching machine models human-like taste and aesthetic judgment has become the next major frontier in technology.
- 1Taste Labs is building a community of human tastemakers to train AI models on subjective aesthetics and design preferences.
- 2The sudden liquidation of Leopold Aschenbrenner's Situational Awareness fund highlights shifting dynamics in AI investment.
- 3Robotic automation is transforming physical infrastructure, demonstrated by new autonomous systems installing solar panels.
Don't miss
Thais Castello Branco explains the business model and philosophy behind commodifying human curation to train large language models.
The brief
Artificial intelligence can write code and analyze data, but it notoriously struggles with aesthetic quality, often churning out generic design and writing. Taste Labs founder Thais Castello Branco is trying to solve this by training models on human preference data.
By leveraging a community of tastemakers, Taste Labs aims to build a curation infrastructure that injects subjective human taste into AI. The goal is to move past mechanical outputs and eliminate the low-quality aesthetic known as AI slop.
Beyond design aesthetics, the tech landscape is shifting rapidly. The episode covers the liquidation of Leopold Aschenbrenner's prominent Situational Awareness fund and the controversial addition of Google Earth's new AI image generation features.
The hosts also look at physical automation, highlighting a robotic arm from Grit Robotics that can autonomously install solar panels, signaling a massive leap forward for rapid, low-cost clean energy deployment.
What was said on this episode
33 statements · 16 positive · 10 negative · 3 mixed · 4 neutral
AI models can solve advanced technical tasks but struggle with quality writing and design.
“You have models that can solve cyber hack and solve PhD level math problems and they can't like write a good tweet or make a good design.”
Listen at 0:00
Mass AI production increases the value of exceptionally high-quality work.
“The more you have of this like mass production of things, the more I think the peak of the peak of the peak becomes valuable.”
Listen at 0:35
Excellent subjective work tends to be unique and outside the average distribution.
“But then for something that's more subjective, what's considered great in real life, things that you encounter that are like, hey, this is really great writing, or really great design, they tend to be things that are a bit more unique, that are a bit more out of distribution.”
Listen at 6:46
AI-generated creative output lacks sufficient variety and nuance.
“And so I think that variety is one of the core things that are missing, but also that nuance.”
Listen at 7:03
AI outputs often fail to fit user intent and personalization needs.
“And so I think this almost fits to the prompt and fit to the user intent. And almost like personalization is almost missing from.”
Listen at 7:24
Taste is rare because developing it requires extensive exposure within a domain.
“Taste is such a rare skill because I think it takes this combination of seeing many things and having that exposure in whatever domain it is.”
Listen at 8:54
Experts with taste should help train AI models and define quality.
“And so yeah, I think we have to find the people that have it to help us go and then train these models and determine what good is.”
Listen at 9:27
TasteLabs has a community of approximately 1,000 domain-specific tastemakers.
“we have this awesome community of about 1,000 tastemakers of different domains”
Listen at 15:30
AI models should first be improved to produce consistently high-quality outputs.
“the first job is just, let's lift the bar. Off the ground and even get to a place where these models can output like high quality things”
Listen at 17:55
Social media accelerated cultural trend cycles, and AI may accelerate them further.
“I do think though that social media accelerated the speed of that cycle, let's say. And AI could have that potential too.”
Listen at 22:07
AI diversity and breadth could push people toward more creative directions instead of conformity.
“as we push for that diversity where it's not meant about it having one correct answer, but it's about it having both breadth and the ability to execute well in all those different directions, then I think actually could help the problem in a good way of pushing people towards actually more directions rather than compressing to the mean further.”
Listen at 22:33
A personalized AI “cool hunter” can successfully identify recommendations matching specified tastes.
“So then I programmed my own cool hunter and I said, you are a cool hunter. You're looking for things that would appeal to these specific people, you know, in these categories that have these specific design elements, etc. And it works, that's all I'm saying.”
Listen at 24:34
Ratings, critiques, curation, and examples each provide useful training data for AI models.
“there's all these different data shapes that are helpful in different ways for model training.”
Listen at 25:56
AI will make tastemakers more valuable rather than eliminate them.
“I don't actually fundamentally believe that tastemakers are going to go away by any means. With AI, I actually think it's the opposite. I think they become more valuable”
Listen at 26:52
Mass AI content creation will increase the relative value of exceptional creative work.
“the more you have of this mass production of things and kind of creation of outputs, which is just going to happen because everyone now in the world can click a button and create something, the more I think the peak of the peak of the peak becomes valuable”
Listen at 27:00
San Francisco’s concentration of people creates uniquely valuable effects that other places cannot reproduce.
“this concentration of people in one place just creates the type of magic you can't reproduce.”
Listen at 34:29
Current AI models cannot reliably produce genuinely funny jokes.
“Grok is not funny. And that's what I mean. Like, I feel like we've all kind of even given up, like, there. It's never going to be funny. It's never going to write a good joke.”
Listen at 35:37
Preference data is currently the most valuable type of data.
“Nothing more valuable right now than preference data.”
Listen at 36:36
Using fourfold leverage against capital is extremely dangerous.
“This was the opposite. This would be like somebody who's got a billion dollars trying to put $4 billion to work. Very, very dangerous.”
Listen at 46:10
Leopold Aschenbrenner will recover and return stronger after the fund’s losses.
“But he's going to come back stronger, is my prediction.”
Listen at 46:18
Google Earth users can alter displayed imagery through natural-language prompts.
“Now you can go to Google Earth and with a simple prompt, you can change things the way things look on Google Earth.”
Listen at 51:05
AI image editing in Google Earth creates substantial misinformation potential.
“the potential here for creating misinformation and visually appealing, inaccurate things is massive.”
Listen at 52:09
Technologies that accelerate activity inevitably generate harmful uses.
“Every time there's a New technology that accelerates stuff, you're going to have these bad uses of it.”
Listen at 53:57
Boss-driven AI adoption can lead employees to use AI thoughtlessly.
“there is, you know, and then back to being AI being pushed by the boss. And I'm one of these pushy bosses saying, hey, let's use AI, let's use AI. You know, you're going to have situations where people don't do it thoughtfully.”
Listen at 55:20
AI-generated content is degrading social-media feeds with fabricated shark videos.
“it's screwing up my feed because I like to show my daughter is really into sharks and I like to show her shark videos in my Instagram feed. And now they're all like, oh my God, here's a great white shark taking three bites out of a person. I'm like, that's never been recorded”
Listen at 56:59
AI-generated slop is causing some young people to view AI negatively.
“This is why young people hate AI. Like my daughters, when I say, hey, use AI to find a place to have dinner. You know, I use Gemini and put it on a Google map. They're like, no, AI, we don't want to use AI. AI is lame.”
Listen at 57:33
Ethan Goodheart’s camera-based golf cart can provide autonomous rides on Stanford’s campus.
“He's had some beta testers, so he's offering rides on his autonomous golf cart around the Stanford campus.”
Listen at 1:03:59
Self-driving technology is becoming increasingly commoditized.
“what this shows is how commoditized self driving is getting.”
Listen at 1:04:36
The self-driving competition will be determined by manufacturing, operational, and licensing scale.
“What is the self driving race going to come down to? It's going to come down to scale. Who can build and manage and get licenses for this?”
Listen at 1:05:43
China is restricting autonomous vehicles primarily to protect driving jobs, not for safety.
“China is freezing the number of self driving cars on road. Why? It's not about safety, it's about protecting jobs.”
Listen at 1:07:21
Self-driving companies should fund unemployment support or retraining for displaced drivers.
“I do think some licensing of these collecting of revenue to smooth out for society the unemployment and or retraining of those individuals would not be out of line”
Listen at 1:08:32
AI-trained robots can unload, transport, and begin installing utility-scale solar panels, but humans remain necessary.
“The robot. They have trained the robot to be able to unload a solar panel. It can transport it to where it needs to go and it begins the installation process. You do need a human in the loop.”
Listen at 1:10:09
Installing solar panels over California irrigation canals generates electricity and reduces water evaporation by 70%.
“California roof two irrigation canals with solar panels that makes electricity and cut the evaporation off the water lawn by 70%.”
Listen at 1:13:22
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Books & mentions
This Week in Startups
To stay updated on tech news, venture capital, and the builders shaping the future of AI and robotics