By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
The Twenty Minute VCAidan Gomez: What No One Understands About Foundation Models | E1191
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
At a glance
WHAT IT’S REALLY ABOUT
Aidan Gomez Explains Foundation Models, Data, and AI’s Real Future
- Aidan Gomez, cofounder and CEO of Cohere and coauthor of the Transformer paper, discusses the economics, technical progress, and product landscape of large AI foundation models.
- He argues that while scaling compute reliably improves models, the real frontier is data quality, new methods for reasoning, and efficient smaller models tailored to enterprises.
- Gomez predicts a world of multiple horizontal and vertical models, falling inference costs, tight margins at the model layer, and major value capture at both the chip and application layers.
- He is optimistic about AI’s role in productivity, agents, robotics, and copilots for workers, while dismissing doomsday scenarios and emphasizing trust, privacy, and deployment models for enterprises.
IDEAS WORTH REMEMBERING
5 ideasScaling models with more compute works, but it’s inefficient and economically constrained.
Bigger models almost always perform better, but each incremental gain requires exponentially more compute and cost; this favors tech giants unless startups differentiate via data, algorithms, and efficiency.
High-quality and synthetic data are now the primary drivers of model improvement.
Open-source gains largely come from better data filtering, weighting, and synthetic generation; models are extremely sensitive to data quality, making curation and task-specific datasets a competitive edge.
We’re heading toward a multi-model world combining large general models and small specialized ones.
Teams prototype with powerful general models, then distill or fine-tune down to smaller, cheaper models optimized for specific tasks or verticals, creating an ecosystem rather than a single-model monopoly.
The model API layer is becoming commoditized, with margins squeezed by price cuts and open source.
With OpenAI price dumping and Meta releasing strong open models for free, selling “just models” will be a low-margin business; durable value is more likely at the chip and application/product layers.
Enterprise adoption hinges on trust, privacy, and deployment flexibility—not just raw capability.
Large customers resist training on their data and fear IP leakage, so vendors must support private deployments (e.g., in-VPC, on-prem, multi-cloud) and strong guarantees that customer data isn’t used for training.
WORDS WORTH SAVING
5 quotesThere’s no market for last year’s model.
— Aidan Gomez
It’s definitely true that if you throw more compute at the model, if you make the model bigger, it’ll get better. It’s also the dumbest way to improve models.
— Aidan Gomez
Pretty much all of the major gains that we’ve seen in the open source space have come from data improvements.
— Aidan Gomez
If you’re only selling models, it’s going to be a really tricky game… it’s going to be like a zero-margin business.
— Aidan Gomez
You might want your children to be speaking to an extremely empathetic, extraordinarily intelligent and knowledgeable, safe intelligence that can teach them things and doesn’t get tired of them.
— Aidan Gomez
High quality AI-generated summary created from speaker-labeled transcript.
