DeepSeek 'Bad at Post-Training'? Community Pushes Back Citing GRPO and Coding Models
teortaxesTex · x · 2026-08-13
Pushing back against rumors that DeepSeek struggles with post-training, a developer pointed out that the lab released the best open-weights coding model by October 2023, and that half the industry relies on their GRPO algorithm.
The original thread mentioned rumors that founder Liang Wenfeng deprioritized AI agents, allegedly leading to the departure of a key member. Commenters remained skeptical about their software engineering (SWE) agent capabilities, noting severe context-loss failures in the recent 0813 release.
Related event: Community Refutes Claims of DeepSeek's Weak Post-Training(2 posts)→
More from Companies & People
- Creator says xAI left 20+ creators unpaid amid new X video bonus program — koltregaskes · 2026-10-16
- F{AI}R: a free AI unconference in Liebefeld, Bern, builds its schedule on-site — nicodotdev · 2026-10-16
- Sakana AI scientist to keynote on self-improving AI at Seoul's DEEP DIVE INTO 2026 — Stefania_druga · 2026-10-16
- Open Logic Textbook forall x: Calgary Joins ChapterPal's Free Textbook Collection — burkov · 2026-10-16
- Resurfaced 2021 Anthropic roadmap tied AI progress to personal-life decisions — akbirthko · 2026-10-16
- Supabase CTO on how AI agents will reshape software buying and open source — David Ondrej · 2026-10-16