Just got this in this morning
OpenAI is Down!
"■ FOR IMMEDIATE RELEASE
ChatGBP Declares Digital Bankruptcy, Warns of Impending Downgrade to 56k Dial-Up Sounds"
@ohryansbelt@ericweinstein@joerogan Ryan, it is now too late to be honest, there is too much capital invested into AI for this to stop, even if we never reach AGI or beyond this is unstobable. So where do you think the money to pay them back will come from?
@nextofficial why are you sti charging irish customers extra 17% on all your products. After all these years ripping off your irish customers. Just remembered why stopped shpping here.
This is wonderful with real world data, The Transformer's attention mechanism softmax(QK^T/√d_k)V always outputs confident probability distributions, even when it shouldn't be sure.
Real data: "Paris is beautiful" vs "Paris is overrated"
Synthetic data: Consistent templates only
Result: Model never learns uncertainty or conflicting views
6 Whether it’s the linguistic density or the MoE (Mixture of Experts) architecture, the West is winning narrative has a blind spot.
I break down the full "AGI Space Pen" theory here:
https://t.co/ZBKim5zZJp
@rowancheung@ylecun@bindureddy
5 The Pencil Advantage isn't just language. It's Embodied AI. As @DrJimFan notes
China has 50% of the world's industrial robots. While Western AI trains on Reddit threads, Chinese AI trains on physics and manufacturing data.
4 Efficiency beats Brute Force. History proves this repeatedly: Google PageRank > Lycos More Servers Tesla Software > Detroit More Factories
While the West bruteforces scaling laws $100M+, China is optimizing the architecture.
@balajis calls The Network State vs legacy systems.
3 This compounds over long reasoning chains.
If a math proof takes 50 steps in English, it might take 30 in Chinese. Less noise. Less drift. Crisper logic.
The AI discovered that Chinese is simply a better OS for reasoning.
Reducing the noise improves the signal @fchollet
2 It comes down to Tokens. LLMs don't read words they read chunks of meaning.
English: "I'm going to the store" (High token count) Chinese: "我去商店" (Dense logograms)
Chinese encodes the same meaning with ~40% less tokens. We are in a compute-constrained world!
@karpathy