Keep pulling the thread on Aidan Gomez.
The overwhelming majority of the training data Cohere currently generates for its new models is synthetic.
The introduction of reasoning capabilities has been a step-change improvement for AI models, enabling them to succeed at complex tasks where pre-reasoning models would almost always fail.
Cohere positions itself as a cloud-agnostic AI provider that can deploy anywhere, including in a customer's Virtual Private Cloud (VPC), a strategy supported by its non-commercial release of model weights.
AI agents that can perform a month's worth of research in about an hour are a high-demand enterprise use case, particularly in financial services for roles like wealth managers.
Current foundation models lack the crucial capability of learning from experience and user feedback over time, a key feature of human intelligence that needs to be developed.
The "scale is all you need" hypothesis for improving AI models is breaking down, as the industry is now experiencing heavy diminishing returns on capital and compute investment.
Aidan Gomez argues that vertically integrated companies that build both foundation models and applications have a significant advantage in product quality because they can optimize the underlying model for specific customer use cases.
Aidan Gomez predicts the enterprise AI application market will initially be fragmented with best-of-breed solutions for different departments, but will eventually consolidate towards a single, integrated platform.
The AYA project by Cohere4AI was the largest data collection effort in machine learning history, involving thousands of native speakers of over 100 languages, with the resulting dataset being open-sourced.
OpenAI is currently focused on the consumer AI market.
Google's Gemini and Meta are planning to compete directly with OpenAI in the consumer AI market.
Anthropic's primary strength and strategic focus is on code generation.