Kaggle Co-founder Goldbloom: Competitions Shape the Current AI Wave
Kaggle co-founder Anthony Goldbloom stated that he is most proud of the individuals emerging from the competition community who are shaping the current wave of artificial intelligence, specifically naming figures associated with OpenAI, Anthropic, ARC-AGI, and OpenRouter.
He mentioned that Greg Brockman practiced deep learning using Kaggle before co-founding OpenAI. Tom Brown worked on Kaggle projects while self-learning machine learning, later becoming the lead author of the GPT-3 paper and co-founding Anthropic. Francois Chollet introduced Keras to early competitors and later hosted the first ARC-AGI competition on Kaggle; this benchmark tests whether models can learn new rules and solve unseen problems with very few examples, becoming a common measure for cutting-edge reasoning.
Goldbloom specifically noted that Chris Clark, who was responsible for Kaggle's products and engineering, later co-founded the model routing company OpenRouter. The post was inspired by his listening to the a16z podcast, where a guest discussed how winning a Kaggle competition led them into artificial intelligence. You.com founder Richard Socher shared this summary.
Kaggle was founded by Goldbloom and others and was acquired by Google in 2017, long focusing on prize competitions, public datasets, and leaderboards to train data science and machine learning practitioners. Competitions compress reproducible experiments, feature engineering, and model iteration into public contests, allowing many to transition from tutorials to deployable systems. ARC-AGI is designed to counter rote memorization: the problem rules are given on the spot, and models must generalize immediately, unlike traditional benchmark tests.
OpenRouter consolidates multiple model interfaces into a routing layer, allowing applications to switch vendors based on price, latency, and capability, with the product team coming directly from the competition platform's engineering leads. This path illustrates that Kaggle outputs not only modeling skills but also individuals who can deploy models into production.
In market mechanisms, this is a talent supply narrative, not about model releases or computing power transactions. Buyers are cutting-edge labs and infrastructure companies competing for engineers who have completed full competition cycles; sellers are individuals visible on public leaderboards. The flow of funds is towards lab salaries and startup equity, rather than competition prizes themselves. Benefiting are platforms that can still publicly validate work, while closed internal training faces pressure: candidates without external competitive records find it harder to be seen. The event itself does not change model weights but reinforces the hiring premium that "competition experience equals transferable skills."
ABAB AI Insight
Kaggle's original business was to outsource corporate data science problems to amateur and professional competitors worldwide. In the 2010s, Netflix, insurance, and advertising companies exchanged prizes for models; after Google's acquisition, the platform became a customer acquisition funnel for TensorFlow and cloud services. Brockman came from Stripe infrastructure, Brown from Google Brain, and Chollet from the Keras toolchain, all treating competitions as low-cost experimental grounds: data is public, metrics are strict, and failures can be reviewed. This is in stark contrast to closed benchmarks within labs.
The capital path is for the platform to first cultivate competitors, then be acquired by large firms. Google's acquisition of Kaggle was for community and dataset distribution rights; OpenAI and Anthropic acquired individuals who could push models to the top under public pressure. OpenRouter has turned multi-model switching into a paid gateway, effectively productizing the competition habit of "switching models and resubmitting." The motivation is not nostalgia but the need for cutting-edge labs to have verifiable hands-on records, as papers on resumes are insufficient.
This is analogous to how the ImageNet competition brought forth the AlexNet team and the DAC competition supplied chip designers: leaderboards price capabilities faster than degrees. The current phase is shifting from selecting candidates based on competitions to defining capabilities through competitions—ARC-AGI has transitioned from a Kaggle activity to an external reasoning transcript for labs. Whoever controls the benchmarks partially controls the narrative for the next generation of models.
Structural judgments indicate a transfer of pricing power. Talent pricing power has shifted from schools and internal recommendations to reproducible public contests. The mechanism is that when model capabilities are similar and papers are hard to authenticate, the market uses time-limited, data-limited, reproducible competition problems as filters. After Kaggle has proven this point, new benchmarks like ARC will continue to seize this power until the next paradigm renders old rankings obsolete.
ABAB News · Cognitive Laws
- Leaderboards price capabilities faster than resumes.
- Those who publicly fail are better hires than those who score perfectly internally.
- Tool inventors will eventually become benchmark setters.