There's a delicate balance here - between bringing proprietary learning in-house versus outsourcing improvement loops to generalist, more powerful models
On one hand, the case that Satya makes is clear and probably consensus - private companies should have private control over their training, airgapped from the outside
But taking the other extreme, maximum atomization results in the amplification of differences between otherwise similar companies
e.g. a car repair shop in China and a car repair shop in the US each build their own in-house AI, which results in vastly different, custom-made operational architectures, even though they'd likely save resources if they combined their overlapping knowledge
The middle ground is that the more public pools of synthesized knowledge exist at different layers (AI across industries, countries), while the core, idiosyncratic knowledge stays private (AI for Microsoft specifically)
Imagine a one-way valve where public domain knowledge is allowed for private use, but private IP doesn't leak out
Abstractly speaking this follows a natural power law, and it reminds me of how we communicate and form text: words like "the", "and", and "you" are commonly used but lack specificity - but combined with more niche words, they provide full value
AI is about balancing common knowledge with proprietary knowledge - and together providing full value