This reminds me of the time when the report on woke grants found a huge increase (<1% to 17%) from like 2021 to 2022 because the database actually contained the full abstract instead of just the title.

Are we just early? Right now training a near-frontier model isn't that expense. In 1980s making your own OS wasn't too complex and had options.

Models are going get more expense. Lock-in/moats are possible around APIs, tool use, data/history/prompts etc. By 2000 Windows was incredibly dominant.

This day9 description of it and why his mods had to ban words like Mario, Luigi, and Peach might be funny to you.
youtu.be/WbaT7zPskPM?...

Oxygen is an abundant oxidation agent found throughout the atmosphere and enables abundant energy when reacted with similarly abundant organic material. Argon and nitrogen don't have energy releasing reactions with anything found lying around.

This is my feeling about Astronomy. Astronomy constellations/classes of objects is stamp collecting. But there is so much going on. How we know of far stuff is away is interesting. Stars are unstable reactions with 15+th temperature relationships. And then there are things we don't understand yet

Computer analogy would suggest those able to use the models and get some sort of moat would dominate. Though frontier models tend to gobble wrappers so these use cases would need to be resistant to becoming trivial as base models improve

Analogies with computer/internet seem to suggest that frontier models might be commoditized but not necessarily that hardware is going to be where the profit is.

One AI company's model might only be 6 months ahead but they might have the best app and all your data and have good branding.

How do you feel about Java 😅?
stackoverflow.com/questions/88...

What's the best case for immigrant's "poor culture harming institutions" being a significant impact? Lot of these seem like trying too hard to have speculative second order effects dominate