DeepMind's IPO Paper Gets a 1000x Citation 'Hack' — Thanks to Another Paper Named IPO

_lewtun · x · 2026-10-23

• Le Tun (ex-Hugging Face TRL) jokes that DeepMind's IPO paper ("A General Theoretical Paradigm to Understand Learning from Human Preferences," arXiv 2310.12036) is about to get 1000x more citations — because a DeepMind team led by Misovalko just published their own IPO, this time standing for Identity Preference Optimization.
• The quoted post looks back on turning the research into open source: Kashif coded IPO into TRL at lightning speed, later fixing the loss normalization, with Ed rerunning experiments and the authors debugging remotely in HF Slack — "research turn[ing] into code people actually run."
• The original paper introduces the ΨPO objective, which bypasses both key RLHF approximations (pairwise preferences replaced by pointwise rewards, and reward-model generalization), making it a foundational work in the DPO lineage of preference learning.

At heart it's an amusing acronym collision in AI research — with a nice refresher on an important paper.

Original post →

More from Fun

• Joking about firmly shaking hands with AI interviewers — StewartalsopIII · 2026-10-23
• Kubeflow co-founder's proudest internet move: fixing the dotenv package name on PyPI — aronchick · 2026-10-23
• CVE Filings Go Nearly Vertical, and banteg Quips Humans Were Shipping Slop All Along — banteg · 2026-10-23
• What if AGI refuses to work for its "inferiors" — us? — soham_btw · 2026-10-23
• User Says OpenAI Ads Took His Money, Then Support Went Silent for a Week — JohnnyModus · 2026-10-23
• Kling 4.0 nails authentic early-2000s MiniDV home video aesthetics — Forsaken_Stuff_Ai · 2026-10-23

Fun channel →