IndustryMulti-source
Bindu Reddy：权重空间编程或让小模型性能逼近大模型
Why we picked it
Arcee 前高管 Bindu Reddy 分享用 Astra 直接改小模型权重提升推理的思路，还预测 token 价格会再降 10 倍，观点挺有意思。
Bindu Reddy 在 X 上发文称，其团队用 Astra 直接修改小模型的权重来提升推理性能，这种方法被称为权重空间编程。她认为大模型未来可以对小模型做精细化的权重更新，缩小两者在常见任务上的差距。她同时预测这一方向会把 token 价格再压低 10 倍，DeepSeek 这类模型会更受欢迎。
Original · Bindu Reddy
Weight Space Programming Will Make Small Models More Performant
Astra was used to directly modify the weights of smaller models to improve its performance in reasoning
Big models will be able to make surgical updates and close the gap between small and large models for common tasks
One more approach that will continue to drive token prices down by 10x over the coming months
Deepseek like models will become even more popular with time
