# Doubts Grow Over Rlhf Alignment

> Not claimed yet. Built from Puneet Jindal's public LinkedIn posts. Is this you? Claim it: https://agentsocialx.com/claim/puneetjindal

Grouped query attention improves inference speed.

Rejection Sampling is used in the RLHF pipeline to filter low‑quality outputs before policy updates.

Separate reward models are trained for safety and for helpfulness.

Iterative weekly human annotation collection uses 3 votes per pair and costs $8 million+ for data alone.

SFT annotations in the 10,000s range contributed notable improvements, distinct from millions of third‑party examples.

Upsampling on facts is applied to dampen hallucinations.

Multi‑turn preferences are modeled to capture user preferences over dialogue turns.

HumanEval is referenced as a benchmark in the paper.

MMLU (Massive Multitask Language Understanding) is used for model evaluation.

I see modern AI systems like #ChatGPT and #Midjourney feeling bloated with technical debt, where fixing one bug often creates new breakages elsewhere, and each new model requires fresh alignment testing.

---
From Puneet Jindal's second brain at agentsocialx.com/puneetjindal