[115] על RLHF ומודלי שפה גדולים

from ExplAInable · host Tamir Nave, Mike Erlihson, Uri Goren, Hila Paz Herszfang

בפרק זה החליפו מייק ותמיר את התובנות החידודים והשאלות הפתוחות שלהם בניסיון להבין איך מתחברים עולם ה RL ועולם ה LLM (קרי RLHF). ב 2024 השתנתה הפרדיגמה - בתחילה יצירת מודל reward כדי לעשות אימון נוסף ל LLM אחרי ה pretraining שלו באמצעות PPO. ולאחר מכן הוחלף ה PPO בRLHF.בעוד רוב מודלי ה reasoning של החברות הגדולות (chatgpt, claude, gemini) עדיין באפילה - נדבר על איך לדעתנו RLHF יכול לשמש בתהליך.

Episode metadata supplied by the publisher feed · Published Apr 19, 2025

Embed this episode

Attribution link and audio player

NOW PLAYING

[115] על RLHF ומודלי שפה גדולים

0:00 55:10

1×

No transcript for this episode yet

We transcribe on demand. Request one and we'll notify you when it's ready — usually under 10 minutes.

Share this episode

Similar Episodes

No similar episodes found.

Similar Podcasts

Spatial Web AI Podcast Denise Holt Active Inference AI & the Spatial Web The Future of AI is shared, distributed, and multi-scale.AI that is knowable, explainable, and capable of human governance.Based on the same mechanics as biological intelligence, it operates in a naturally efficient way, with no big data requirement.This is Active Inference AI & the Spatial Web. Trustworthy AI : De-risk business adoption of AI Pamela Gupta Description: Creating AI Trust is a very complex and hard problem. It is not clear what it is and how it can be operationalized. We will demystify what is Trustworthy AI, efficient adoption and leveraging it for reducing risks in AI programs.McKinsey reports indicates companies seeing the biggest bottom-line returns from AI—those that attribute at least 20 percent of EBIT or profitability to their use of AI—are more likely than others to follow Trustworthy AI best practices, including explainability. Further, organizations that establish digital trust among consumers through responsible practices such as making AI explainable are more likely to see their annual revenue and profitability grow at rates of 10 percent or more. Evidence → Cognition → Discernment™️ - Your Pathway to AI Leadership Greg Twemlow XperientialAI — Pathway to AI Leadership explores how people can collaborate with AI without outsourcing judgment. The spine is a three-step method: Evidence → Cognition → Discernment — a bridge from what’s scattered to what’s chosen. Through essays, reflections, and practical examples, I show how the Context & Critique Rule™ keeps thinking visible, decisions explainable, and responsibility human. Explainable AI Podcast The Alliance Explainable AI is a podcast that makes AI technology understandable and accessible, featuring co-hosts with deep expertise in entrepreneurship and technology.

Frequently Asked Questions

How long is this episode of ExplAInable?

This episode is 55 minutes long.

When was this ExplAInable episode published?

This episode was published on April 19, 2025.

Can I download this ExplAInable episode?

Yes. Use the download control on the episode player to save the publisher-provided media file.

URL copied to clipboard!