Podcast · 156 episodes on record

A recipe for frontier model post-training

Published

August 7, 2024

Length

10 min

Position

41 of 156

August 7, 2024

A recipe for frontier model post-training

0:00 / 10:23

0:00Choose a match or chapter to light up its moment10:23

Listen to the full episode

Excerpt playback isn’t available for this episode — use the full-episode link.

Search the Record

Moment search isn’t open for this episode yet

When it opens, you’ll be able to type any phrase and jump straight to the second it was said.

Nathan Lambert Episodes Around August 7, 2024

Episode 41 of 156 — walk the feed in the order it was published.

View all episodes

14:05

OpenAI's Model (behavior) Spec, RLHF transparency, and personalization questions

May 13, 2024

12:28

OpenAI chases Her

May 16, 2024

9:03

Name, image, and AI's likeness

May 22, 2024

8:29

We aren't running out of training data, we are running out of open training data

May 29, 2024

7:49

RLHF Roundup: Trying to get good at PPO, charting RLHF's impact, RewardBench retrospective, and a reward model competition

Jun 26, 2024

56:30

Interviewing Dean Ball on AI policy: CA SB 1047, upcoming AI disaster response, Llama 3 405B, Chinese open-source AI, and scaling laws

Jun 27, 2024

6:39

Switched to Claude 3.5

Jul 3, 2024

14:19

SB 1047, AI regulation, and unlikely allies for open models

Jul 17, 2024

15:22

Llama 3.1 405b, Meta's AI strategy, and the new open frontier model ecosystem

Jul 23, 2024

7:55

GPT-4o-mini changed ChatBotArena

Jul 31, 2024

1:03:42

Interviewing Sebastian Raschka on the state of open LLMs, Llama 3.1, and AI education

Aug 1, 2024

10:23This episode

A recipe for frontier model post-training

Aug 7, 2024

1:02:22

Interviewing Ross Taylor on LLM reasoning, Llama fine-tuning, Galactica, agents

Aug 8, 2024

8:32

Nous Hermes 3 and exploiting underspecified evaluations

Aug 16, 2024

8:00

On the current definitions of open-source AI and the state of the data commons

Aug 28, 2024

10:31

OLMoE and the hidden simplicity in training better foundation models

Sep 4, 2024

10:40

OpenAI's Strawberry, LM self-talk, inference scaling laws, and spending more on inference

Sep 5, 2024

5:38

A post-training approach to AI regulation with Model Specs

Sep 10, 2024

11:31

Futures of the data foundry business model

Sep 11, 2024

18:51

[Article Voiceover] Reverse engineering OpenAI's o1

Sep 17, 2024

14:04

[Article Voiceover] Llama 3.2 Vision and Molmo: Foundations for the multimodal open-source ecosystem

Sep 27, 2024

1:08:39

Interviewing Riley Goodside on the science of prompting

Sep 30, 2024

10:29

[Article Voiceover] AI Safety's Crux: Culture vs. Capitalism

Oct 2, 2024

11:47

How scaling changes model behavior

Oct 9, 2024

1:00:12

Interviewing Andrew Trask on how language models should store (and access) information

Oct 10, 2024

16:36

(Voiceover) Building on evaluation quicksand

Oct 16, 2024

54:21

Interviewing Arvind Narayanan on making sense of AI hype

Oct 17, 2024

View all episodes

Summary

Apple, Meta, and Nvidia all agree -- synthetic data, iterative training, human preference labels, and lots of filtering. This is AI generated audio with Python and 11Labs. Source code: <a class="linkified" href="https://github.com/natolambert/interconnects-tools" target="_blank">https://github.com/natolambert/interconnects-tools</a> Original post: <a class="linkified" href="https://www.interconnects.ai/p/frontier-model-post-training" target="_blank">https://www.interconnects.ai/p/frontier-model-post-training</a>00:00 Llama 3.1 post-training and the new normal for RLHF 01:18 A new standard pipeline 01:45 Human preference data 02:59 Scaling RLHF 05:03 Synthetic data 06:10 The new normal 06:51 Data quality is king 07:18 Apple confirms the new normalFig 1: <a class="linkified" href="https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_018.png" target="_blank">https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_018.png</a> Fig 2: <a class="linkified" href="https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_020.png" target="_blank">https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_020.png</a> Fig 3: <a class="linkified" href="https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_031.png" target="_blank">https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_031.png</a> Fig 4: <a class="linkified" href="https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_033.png" target="_blank">https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_033.png</a> Fig 5: <a class="linkified" href="https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_035.png" target="_blank">https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_035.png</a> This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit <a href="https://www.interconnects.ai/subscribe?utm_medium=podcast&utm_campaign=CTA_2">www.interconnects.ai/subscribe</a>

Episode details and artwork are published by the show’s own feed. Sentinel plays short cited excerpts and links every episode back to its publisher.

← All episodes from Nathan Lambert

A recipe for frontier model post-training

Search the Record

Nathan Lambert Episodes Around August 7, 2024

OpenAI's Model (behavior) Spec, RLHF transparency, and personalization questions

OpenAI chases Her

Name, image, and AI's likeness

We aren't running out of training data, we are running out of open training data

A realistic path to robotic foundation models

AI for the rest of us

Text-to-video AI is already abundant

Frontiers in synthetic data

RLHF Roundup: Trying to get good at PPO, charting RLHF's impact, RewardBench retrospective, and a reward model competition

Interviewing Dean Ball on AI policy: CA SB 1047, upcoming AI disaster response, Llama 3 405B, Chinese open-source AI, and scaling laws

Switched to Claude 3.5

SB 1047, AI regulation, and unlikely allies for open models

Llama 3.1 405b, Meta's AI strategy, and the new open frontier model ecosystem

GPT-4o-mini changed ChatBotArena

Interviewing Sebastian Raschka on the state of open LLMs, Llama 3.1, and AI education

A recipe for frontier model post-training

Interviewing Ross Taylor on LLM reasoning, Llama fine-tuning, Galactica, agents

Nous Hermes 3 and exploiting underspecified evaluations

On the current definitions of open-source AI and the state of the data commons

OLMoE and the hidden simplicity in training better foundation models

OpenAI's Strawberry, LM self-talk, inference scaling laws, and spending more on inference

A post-training approach to AI regulation with Model Specs

Futures of the data foundry business model

[Article Voiceover] Reverse engineering OpenAI's o1

[Article Voiceover] Llama 3.2 Vision and Molmo: Foundations for the multimodal open-source ecosystem

Interviewing Riley Goodside on the science of prompting

[Article Voiceover] AI Safety's Crux: Culture vs. Capitalism

How scaling changes model behavior

Interviewing Andrew Trask on how language models should store (and access) information

(Voiceover) Building on evaluation quicksand

Interviewing Arvind Narayanan on making sense of AI hype

Summary