User Scolds DeepSeek AI for Solving Post-Cutoff AIME Problem

T(

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTexTECH

It can't. StepFun 3.5, the only open model that I see solving this, also can't. But it can brute force to completion what DeepSeek abandons (takes it 180K tokens). Neither of them use elegant insights demonstrated by top models. https://x.com/teortaxesTex/status/2074654263804915750/photo/1

T(

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTexTECH

In sum, if this is V4-Official, I conclude it's a failed pretrain, a failed non-scalable architecture, or DeepSeek operates on a misguided RL paradigm that didn't progress since V3.2 Speciale, and I am worried for their future.

T(

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTexTECH

Infuriating but whatever works I guess https://x.com/teortaxesTex/status/2074639123881623587/photo/1

T(

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTexTECH

devastating https://x.com/teortaxesTex/status/2074655286254555465/photo/1