Supervised fine-tuning and Reinforcement Learning from Human Feedback (RLHF) can reduce the output diversity of language models

Overview Debates Related claims Topics

Reducto trained its R1 model using reinforcement learning loops around mid-training and post-training on edge cases

Natural language processing models operate on compressed, lower-dimensional manifold representations of high-dimensional physical phenomena

Sequin does not use generative AI or large language models within its ranking pipeline

Google developed Gemini 4 Argon from a new pre-training run after terminating the planned Gemini 3.5 Pro release

Google positioned Gemini 4 Argon as an ultra-large artificial intelligence model in the same operational class as Astra and Fable

Google announced the Gemini 4 Argon frontier artificial intelligence model at the end of September 2026

Google DeepMind hired professional writers during Gemini 4 Argon's development to eliminate low-quality machine-generated phrasing and improve prose structure

Leash Bio fine-tuned open-source language models, including Qwen and GLM, to describe chemical structures and molecular differences in natural English

Prompts are a key unit of post-training for AI models

Historically, AI model post-training was heavily reliant on supervised fine-tuning using the outputs of a very strong model

AI model post-training has evolved from supervised fine-tuning to being heavily reliant on Reinforcement Learning (RL) in complex environments

A paper from Anthropic demonstrated that LLMs can de-anonymize individuals from anonymized data

A study based on NeurIPS papers found that LLMs were better than human reviewers at predicting which research ideas would be empirically successful

While at Nvidia, Jason Ma co-developed the Eureka project, which used GPT-4 to auto-generate reward functions for simulated robotics tasks

OpenAI's documentation suggests a theoretical maximum speed of 280 tokens per second for GPT-6.1 Sol in its fast mode

Jason Ma worked at Google DeepMind, exploring applications of the Gemini model for robotics

The Eureka paper, co-authored by Jason Ma, was the first to use a large language model (GPT-4) to self-generate simulation environments and reward functions

The iterative process of collecting data, deploying, and refining a model is a robotics version of Reinforcement Learning from Human Feedback (RLHF)

In 2022, early AI systems were known for "hallucinating" and could not count the number of 'r's in "strawberry."

Carly Taylor suggests that if LLMs developed their own communication protocol, it would likely be closer to machine code