RE: https://mastodon.social/@LukaszOlejnik/117205515325366973
We must not see this as an AI problem at all.
This is an obvious attempt by OpenAI to create regulatory lock-in for AI use, make local, "ungoverned" AI look dangerous, and require people to use AI-as-a-service from a licensed and managed provider, justifying the existence of OpenAI in the first place.
Current frontier models require large clusters of graphics cards with fast memory in the Terabyte range in the first place.
If you improve that by a factor of 100, the same frontier-level capability will run on a desktop machine, with 16 GB to 256 GB of memory and a memory bandwidth in the 256-512 GB/s range – things that already exist and can be bought and run by anyone.
The need for all these data centers, AI-as-a-service subscriptions and the trillions of investment will evaporate. Maybe not completely, but 85% of it.
This happens, at the current rate of improvement, in less than 3 years, because a 100x improvement are three to four 4x improvements and we see these arriving at a rate right now faster than once a year.
This entire discussion is not about responsibility or dangerous AI. It is about securing the reason of existence for OpenAI and Anthropic, killing off open weights models as "dangerous unmanaged AI" and making AI a subscription only model, for safety and security.