Opinion: DeepSeek should shift focus to larger pre-training instead of 4.1
zephyr_z9 · x · 2026-08-18
The author argues that DeepSeek (DS) should not waste energy on version 4.1 and instead focus on a larger pre-training run. Citing Moonshot's smart move to launch a 3T-tier model first, the author suggests this allows a company to focus on post-training while developing the next model (K4) in the background, aiming for a Q1 2027 launch. Zhipu and Whale are noted as having significant work ahead. A quoted tweet discusses self-verification in Luna and Fable as indicators of how far DS can push RL for 4.1.
Related event: Opinion: DeepSeek Should Skip 4.1 and Go Bigger on Pretraining(2 posts)→
More from Companies & People
- Researcher: ML labs' authorship practices violate Cambridge, Harvard and Nature rules — skoularidou · 2026-10-16
- Tokyo's bilingual dev show-and-tell localboast returns Nov 5, speaker slots full — 4310sy · 2026-10-16
- Slack iOS buries 'Save' behind More Actions to make room for Slackbot, echoing MS Copilot — mobileraj · 2026-10-16
- Robotics founder buried in inbound seeks a personal assistant in SF — cixliv · 2026-10-16
- Musk bought Bond's Wet Nellie sub car for ~$1M in 2013 — and wanted to make it really transform — XFreeze · 2026-10-16
- ML researcher: by journal authorship rules, many prolific PIs should be on <10% of their papers — skoularidou · 2026-10-16