đ Engineers â want to meet these four founders and more? Apply to attend our September 16th SF Startup Showcase.
These three companies are tackling a slightly different version of the same question: how do we make AI systems better at learning from the world around them? Core Automation is working on models that can keep learning after deployment, Design Arena uses human taste as an eval layer for subjective AI outputs, and Simile is modeling human behavior itself. The common thread is a move away from static models toward systems that learn, adapt, and respond to real people.
Jerry Tworek, Founder of Core Automation
Todayâs frontier models are limited in design: once they leave the lab, they stop learning. Theyâre trained on enormous datasets, deployed in the real world, and then remain static until theyâre trained again. Jerry Tworek believes the next major leap in AI will come from models that can continue learning through interactions with users and the problems they encounter.
Jerry came to that conviction after leading reasoning and reinforcement learning work at OpenAI. At Core Automation, the team is doing more than just optimizing the current process for better performance. Theyâre researching new learning algorithms that could move beyond large-scale pretraining and reinforcement learning, alongside architectures that scale better than transformers. Core is building what it calls âthe worldâs most automated AI lab,â using agents to automate its own experiments, systems work, and research so their small team can attempt problems that would traditionally require a much larger organization.
And that small team is stacked â including Avery Lamp, longtime engineer and LP at FYSK. When we spoke with Jerry, Core had grown to 20-something people drawn largely from OpenAI and Google DeepMind, along with researchers from Anthropic, Meta, xAI, and Amazonâs AGI lab. They plan to reach roughly 40 people by year-end and are hiring researchers across inference systems, synthetic data, large-scale reinforcement learning, neural network architectures, and agentic systems. Core had already raised $100M at a $1B valuation from investors including Nvidia, Accel, and Spark Capital.
If you want to help figure out what comes after todayâs static models, come meet Jerry at the showcase.
Grace Li, Founder of Design Arena
AI models are now great at generating websites, images, videos, and slides. The newest problem is deciding which models have the best outputs for a certain task. Itâs one thing to benchmark whether a coding model got the right answer. It is much harder to rank an output when human taste is involved â how do you concretely decide which generated image is most fun or which design is more aesthetic?
Grace Li found this problem shortly before graduating from Harvard. She and a group of friends were building an AI game engine. The models could generate games that worked but werenât actually fun. There was no automated benchmark that could explain why. They realized that for subjective domains like design, the missing ingredient wasnât another AI judge. It was human taste at scale. Users ask multiple frontier models to create the same thing and choose which result they prefer. Those comparisons rank the models while generating valuable evaluation data in domains where there often isnât one objectively âcorrectâ answer.
The companyâs growth so far is ridiculous for its size. On August 3, Tech Crunch reported that Design Arena had reached 5.3M users, roughly $60M in ARR, and raised a $7.9M seed round led by Index Ventures. Speaking with Grace two weeks later, she told us they had already reached 5.8M users across 192 countries. The company is still only 10 people but is looking to scale dramatically over the next two months.
If youâre interested in figuring out how we measure what âbetterâ means as AI becomes capable of creating almost anything, come meet Grace at the showcase.
Joon Sung Park, Founder of Simile
Companies spend enormous amounts of time and money trying to answer questions that ultimately boil down to the same thing: what are people going to do? Will customers buy a new product? Which advertisement will resonate? Today, answering those questions means surveys, months of research, and the time-honored tradition of paying consultants seven figures for predictions. Joon Sung Park thinks many of those decisions can be simulated.
Simile grew directly out of Joonâs Stanford research on generative agents and human behavior. Companies can tell Simile which population they want to understand and ask its simulated people questions about products, pricing, purchasing behavior, and more. They arenât generic LLM personas: Simile trains and validates its models against real human data. The company works with ~40 data and panel partners, including Gallup, giving it access to tens of millions of real people for data collection and validation.
The company has scaled just as quickly financially. Revenue grew 5x in the five months following launch. Index Ventures led a $100M Series A in February, and five months later Simile announced a $200M Series B at a $2B valuation. The team is now around 60 people and Joon expects they could grow to 100â120 by year-end.
If building the models that could eventually simulate human behavior at planetary scale sounds interesting, come meet Joon at the showcase.
Apply to attend our next SF Startup Showcase and meet these founders in person.