YouBet said:
Windy City Ag said:
Quote:
China is cheating their ass off as usual. They are feeding their open weight models chat transcripts from Claude to train them. So if you asked them to do stuff Claude had already figured out it worked great. If you asked it something novel, not so much.
I don't disagree, and it begs the larger existential question of all models open or closed "cheating their ass off" by repurposing other people's intellectual property and work output for commercial profit.
But in China's case they have shown the world the OpenAI and Anthropic approach of incredibly capital intensive closed model systems may not make much sense if OpenSource versions can fast follow at a fraction of the cost.
If China can do it, then anyone can do it and corporations are probably going to build their own versions that they can control and monitor that in turn lean on a wide variety of specialist, narrow ai programs like Legal AI or Medical AI etc.
This is basically what the software stack looks like now.
Agree with your thoughts. Instead of corporations though, think country level. There is an entire sovereign movement underway right now because everyone else but us and China have suddenly realized two countries are going to be a global duopoly unless they build their own capability.
Don't think it goes down like that. Scale provides efficiency. If it's true that Chinese models are fast followers due to copying and some increments in model efficiency techniques then those same efficiencies will be built back into the closed-weight models. The proportion of compute dedicated to training over time will fall relative to inference and the cost to deliver any individual task completion will fall while the revenue derived from compute in total will rise (50:50 training:inference into 10:90 training:inference compute allocation).
Enterprise will rely on the closed-weight models to provide security against agents' unintended actions and expertise. Creating competitiveness against lower cost to train Chinese models with vastly superior capabilities in task completion at a cost lower than what individual sectors could achieve on their own due to greater compute capacity.
There's a reason AWS is a good business.
The game is compute capacity. As it stands the likeliest winners in the long run are SpaceXAI/Tesla (Merger in the next year), Meta/Google (they have AI independent revenue sources), Apple (end-user device buy-in), OpenAI and Anthropic in roughly that order. Anthropic trails OpenAI due to a lack of aggressiveness on compute, trails SpaceX/Meta/Google due to a lack of proprietary data (X, Instagram/Facebook, YouTube), and a moat that is not durable (model quality). There's likely a diminishing return on human useful intelligence (is a 500 IQ really twice as useful as a 250 IQ model with 10x compute?). SpaceX is in the long run pole position due to their ability to deploy compute at the rate of profitability. Everyone else is constrained by resources (energy, land) and social acceptance (NIMBYism). Space is in a minority of individual's backyards and the utility will drown out the astronomers' protests especially when SpaceX is launching telescopes to the far side of the moon at a fraction of the cost of building them on Earth.
Risk is capability exceeding capacity for containment. OpenAI seems to be the primary point of risk in that regard currently (Anthropic secondary to them), because of their inferior strategic position it is a real risk. They have to chase ever increasing capability to stay ahead of the better positioned competitors.
Just a guess, but next 18-36 months determine if p(doom) will rise towards an inevitability (internet/cyber/financial market catastrophe) or fall into a quirk of history in the path to universal abundance.
.
