Keep pulling the thread on Aidan Gomez.
Aidan Gomez suggests that Google's current AI models may have surpassed the capabilities of OpenAI's models.
Despite Google's technological catch-up in AI, Aidan Gomez believes it is uncertain whether the company can effectively compete from a product perspective.
Aidan Gomez predicts that large-scale AI model training, which currently uses tens to hundreds of thousands of GPUs, could scale to millions of GPUs within a couple of years.
According to rumors cited by Aidan Gomez, OpenAI's upcoming GPT-5 model is smaller than previous models because their larger "Orion" series failed to yield material improvements.
Aidan Gomez believes the performance gains from purely scaling up the size of LLMs, data, and compute are beginning to saturate.
Aidan Gomez predicts the next major capability for LLMs will be the ability to learn from experience and user interactions, a feature he expects in the next model generation (e.g., GPT-6).
The majority of training data used by all major AI labs is now synthetic data generated by other models.
The top five to seven large language models have converged in capability and are now largely interchangeable with minimal differentiation.
Cohere strategically constrains its model development to what can run efficiently on a maximum of two GPUs, believing the enterprise market cannot support larger models.
Enterprises are now moving from broad, small-scale AI proof-of-concepts to deploying a narrow set of use cases at massive scale across tens or hundreds of thousands of employees.
Cohere has raised approximately $1.6 to $1.7 billion in funding.
The European Union's primary technology strategy has been to regulate foreign tech companies rather than fostering the growth of its own competitive tech industry.