Will Brown from Prime Intellect noted surprise at the muted reaction to how labs now recast development tasks as reinforcement learning environments, pairing iteration inside those setups with heavy compute use before distilling gains into later pretraining runs.
@suchenzang surprised how few people it shook awake tbh. lotta gems in playing forward that "oh everyone's just iterating on the same thing and burning compute on scaling skill issues and progress is turning everything into an RL environment and distilling into the next big pretrain run"
@willccbb ah jan 2025, when r1 violently shook everyone awake
DeepSeek R1's January 2025 release demonstrated reasoning gains through large-scale RL without human-labeled trajectories, yet the shift in framing failed to broadly alert builders despite clear technical signals.
A February 2024 Slack thread already circulated the DeepSeekMath paper on mathematical reasoning, showing some research groups tracked the underlying ideas well before broader community uptake.
Users dismiss DeepSeek R1 RL Training Insights as insignificant and funding-driven, sarcastically noting the limited reaction as one minor error among many.
No Digg Deeper questions have been answered for this story yet.
@suchenzang We’ve been DeepSeek fan girls for a long time :D
in feb 2024, deepseek math dropped this on arxiv but no one in the west paid close attention because they were just another chynese lab
@willccbb > how few people it shook awake
sir, it came after all the purse strings...
@suchenzang one rounding error among many
@willccbb aye, does say a lot when "largest single-day drop" in 2025 is now... that