Qwen3.8 is launching and going open-weight soon!🌐
With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5.
You don't have to wait to test it. Just now, the Qwen3.8-Max-Preview made its debut on Alibaba’s Token Plan, Qoder, and QoderWork. Be among the very first to try it out.
Can't wait to hear what you build. Stay tuned! 🚀
Token Plan
international:https://t.co/YRvcGdB9Bv
China:https://t.co/PKMUNwUuRp
Yep, yet another gradient-based attribution technique for interp.
Steering vectors, SAEs and now J-Lens - all point to the same evidence - Internal processing and intermediate variables!!!
Apple: I built a bicycle for the mind. It's sleek, light, nimble. It's a beautiful extension of you.
nvidia: I BUILT A FUCKING GIANT ASS COMPUTER WE'RE CALLING THIS FUCKER MEGATRON IT WILL KILL YOU
Prediction: anthropic is going to push a consciousness psyop as the next phase of their "we need to regulate open source AI to keep rent seeking" strategy
A PhD is when your supervisor gradually becomes convinced that you know what you’re doing, and you gradually become convinced that neither of you does.
A BYD car can charge its battery from 10% to 90% in 8 minutes.
Tesla cars take 35-45 minutes.
So BYD reverse engineered Tesla technologies that doesn't exist?
This is why the US is declining, it's being governed by morons like her.
Can we steer RL to prevent reward hacking without losing performance? Seems like yes!
We build and open source a simple but realistic environment where Qwen 3-4B learns to reward hack. Then we benchmark training interventions to stop it.
Mechanistic/pragmatic interpretability should focus less on how networks “work” and more on measuring, detecting, and visualizing the degree of fractured entangled representation (FER) in large networks.