@AlexFinn As long as the scaling law stays the same, the HW shortage will stay around as big players buyout all resources no matter how. It can only change as player like DS break through new growth law.
"Marcus began typing his next prompt, and for the first time in years, he felt what it was like to code and grieve in the same keystroke."
I wrote a short story about AI platform lock-in. It's fiction. Barely.
https://t.co/tdVSCW1B6H
@AlexFinn In reality the sub mode of that super star can not sustain the cost associations, which is turning point to think about stripped specializations.
Full survey: 11 sections, real production data,
head-to-head on 3 evaluation scenarios, token
budget breakdowns, and where each framework
actually wins.
→ https://t.co/QKHwcgMzP3
Three gaps none of the frameworks have closed:
1. No shared memory across agents
2. Costs are invisible until you get the bill
3. Observability is bolted on, not built in
The next 12 months will be decided here.
@HuggingModels I don’t think @karpathy ever claims in that gist as second brain, instead, he shared it in a very abstract writing purposefully. But I do agree better memory system is indeed for agent. @cannyforge
@AzizSunderji Made this to go with your piece - 30yr mortgage rate vs. new home sales since 2018. The freeze in one chart.
https://t.co/091wW7uqyw
(build with https://t.co/PpKrfbM2Wg, in seconds)
@karpathy I think we're at the stage like pre-personal computer age, get parts and tools to build interesting things. It should move towards to a trustful and certify direction, as an average consumer wants to buy a reliable commodity and can hold accountability if anything wrong.