Skip to content

ViralLoop

Your next post already worked.

← Back to the post

Loading the post

Remix this post

SaveShare

Your note

Private to you, and used when you remix this post.

Viral score

Why it worked

The mechanics

Hook
Style
Tone
Written for

Analysis by ViralLoop — not written by the original author.

Yam Peleg on Health · ViralLoop

← Back to the post

Post by Yam Peleg on X

Yam Peleg@Yampeleg

Heard a leak from one of the frontier labs (not oai tbh), they reached an unexpected HUGE wall of diminishing returns trying to brute-force better results by training longer & using more and more data.. (more severe than what is published publicly)

3.6K likes166 repliesPosted Nov 9, 2024

View the original on X Written by Yam Peleg. Captured Sep 4, 2026.

Remix this post

Your note

Private to you, and used when you remix this post.

52

Viral score

Based on absolute engagement — the platform doesn’t expose follower counts here, so large accounts score higher.

From likes, comments.

Why it worked

Ends on a short line that is easy to quote and repeat.
Short enough to read in full without expanding, which removes a step before the point lands.
Opens on a negative, which reads as a warning and is harder to ignore than a claim.

The mechanics

Hook: Plain observation
Structure: Hook → Landing
Style: Conversational paragraphs at a steady pace.
Tone: plain, matter-of-fact
Written for: Founders, marketers and operators

Analysis by ViralLoop — not written by the original author.

Topics

Posts built the same way

one dozen rats at a keyboard

It’s embarrassing but the Jimmy Kimmel debacle actually is important. It was the first real attempt by the Project 2025 movement to flex their muscle against someone who wasn’t a nameless, defenseless nobody and it catastrophically flopped in less than a week.

So much cope about DeepSeek. Not only did they release a great model. they also released a breakthrough training method (R1 Zero) that’s already reproducing. I doubt they lied about training costs, but even if they did they’re still awesome for this great gift to the world.

Horse Nomad (Chicago-style) ⚔️

Famously, human learning takes *very little* energy. A toddler learns thousands of words and an entire grammar while eating nothing but raspberries, before learning to read. An LLM needs a nuclear reactor and the entire internet as a training corpus.

First recovery in the ByBit hack. ~$43m (15,000 cmETH) has been clawed back from the hacker. I saw the recovery possibility soon after the hack and SEAL connected me with Mantle/mETH team who made it happen. Huge shoutout to SEAL, Mantle, and mETH teams for their quick action.

As shocking as the Kimi K3 release. Massive performance gain was just with post-training Model is 3x smaller than GLM 5.2 (10x smaller than K3) & works on a MacBook / Spark This is Q1 flagship (Opus 4.6/GPT 5.4) level for < $0.28/m tokens (100x cheaper)

Together with SpaceXAI, we’re training a significantly larger model from scratch, using 10x more total compute. With Colossus 2’s million H100-equivalents and our combined data and training techniques, we expect this to be a major leap in model capability.

More from Yam Peleg

imagine dropping 40T tokens into 2T parameters with 10M context length multimodality and then just uploading it to huggingface

ChatGPT pro coming? - Unlimited access to o1, o1-mini, GPT-4o - Unlimited advanced voice - Access to "o1-pro mode" - $200 per month source: t.co

so I think it is safe to assume that all major players have reached the limits of training longer and collecting more data already.. It is all about data quality now.. which takes time..