In an AI-driven future, organizational charts will evolve into "work charts," where tasks and throughput are more important than hierarchical reporting structures.
The new intellectual property for companies will be AI products that can think, live, and learn from user interactions.
According to a study by Nathan Lambert, it is no longer economically sensible to pre-train a foundation model from scratch once it reaches 30 billion parameters, as post-training optimization becomes more efficient.
Microsoft's Dragon AI product for physicians improved its character acceptance rate from a 30-60% range to 83% after being fine-tuned on a specialized dataset.
The current strategic "season" in the AI industry is the "advent of agents."
Over 15,000 customers have used Microsoft's Azure AI service to produce agents.
The total number of AI agents running on Microsoft's Azure platform is in the millions.
Asha Sharma predicts that in the future, more money will be spent on post-training AI models (fine-tuning, RL) than on pre-training them.
The marginal cost of high-quality output from AI is approaching zero, which will lead to an exponential increase in demand for productivity.
Asha Sharma predicts that all future software products will be "model-forward," with an AI model being a core component alongside traditional software.
Approximately 70,000 enterprise AI tools were launched in the last year.
The performance of Microsoft's Dragon AI product was significantly improved by fine-tuning its model on 600,000 expert-annotated patient-physician interactions.