LLM that loops instead of Doing Chain-of-Thought
Related videos
Alex Hormozi’s Warning: Stop Chasing AI, Build This Instead!
The Diary Of A CEO
243.2k views
This Shrinks Fat Cells 41% While You Sleep (instead of burning carbs)
Thomas DeLauer
26.0k views
The Death of RAG?
bycloud
15.0k views
Kimi K2.5 Brought Us 3 brand NEW LLM Frontier!?
bycloud
22.6k views
LLM’s Billion Dollar Problem
bycloud
44.6k views
Chinese DoorDash Is Making Better LLMs Than Meta
bycloud
22.8k views
The RL Irony in LLMs
bycloud
23.0k views
Why is Trump Fighting NATO Instead of Putin? | Command Post
The Bulwark
79.2k views
The End of These Health Trends (what to do instead)
Thomas DeLauer
23.1k views
Everyone Thinks 2026 Will Boom… Here’s What I’m Doing Instead
Minority Mindset
27.3k views
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Related videos
Alex Hormozi’s Warning: Stop Chasing AI, Build This Instead!
The Diary Of A CEO
243.2k views
This Shrinks Fat Cells 41% While You Sleep (instead of burning carbs)
Thomas DeLauer
26.0k views
The Death of RAG?
bycloud
15.0k views
Kimi K2.5 Brought Us 3 brand NEW LLM Frontier!?
bycloud
22.6k views
LLM’s Billion Dollar Problem
bycloud
44.6k views
Chinese DoorDash Is Making Better LLMs Than Meta
bycloud
22.8k views
The RL Irony in LLMs
bycloud
23.0k views
Why is Trump Fighting NATO Instead of Putin? | Command Post
The Bulwark
79.2k views
The End of These Health Trends (what to do instead)
Thomas DeLauer
23.1k views
Everyone Thinks 2026 Will Boom… Here’s What I’m Doing Instead
Minority Mindset
27.3k views
Top Comments (10)
close enough, welcome back RNNs
Chain-of-Thought is just the LLM 'yapping' to find the answer. works but sucks
You need to cover LFM2.5 8b's way of stopping doom loops.
dude imagine every step of thinking you do, you would have to write down what you just thought about, then get amnesia and then have to read what you are supposed to do completely again and what you have already come up with and trust that what you wrote down conveys all the information and all of the subtleties to solve the task correctly. And then you do that 100 times. Like playing telephone with yourself. Actually honestly that's exactly how my ADHD ass does things
I'm surprised that the idea of telling the model to "think step-by-step" back in 2022, is still one of the largest advancements since transformers themselves. If I recall correctly CNNs were developed by observing that neural networks slowly extracted features. So understanding and improving the mechanisms is the right direction. But other than scaling and training data curation, have they just hit a wall?
We've already dragged transformer to death. Maybe we should revive RNNs or architectures similar to that.
Need to fine-tune a model without the hassle? Try out Crusoe's serverless fine-tuning today! https://www.crusoe.ai/contact-sales/serverless-preview?utm_source=bycloud&utm_medium=influencer&utm_campaign=serverlessfinetuning
This is already a year old idea. See "Scaling Latent Reasoning via Looped Language Models", published October 2025.
I was about to comment how choosing the number of loops at the get-go seemed entirely backwards. Glad to see that they thought of that and tried deferring that choice to the end of the inference step, and saw better performance when they did. Neat!
This video provides the excitement of something mathematically elegant with a slap of reality at the end.
Unlock the Data Inside
Turn Videos into Knowledge
- Get FREE 10/day: transcripts, summaries, chats
- Chat with videos, export text & PDF
- $1 free API credit for RAG, chatbots & research
Free forever plan • All features unlocked
Top Comments (10)
close enough, welcome back RNNs
Chain-of-Thought is just the LLM 'yapping' to find the answer. works but sucks
You need to cover LFM2.5 8b's way of stopping doom loops.
dude imagine every step of thinking you do, you would have to write down what you just thought about, then get amnesia and then have to read what you are supposed to do completely again and what you have already come up with and trust that what you wrote down conveys all the information and all of the subtleties to solve the task correctly. And then you do that 100 times. Like playing telephone with yourself. Actually honestly that's exactly how my ADHD ass does things
I'm surprised that the idea of telling the model to "think step-by-step" back in 2022, is still one of the largest advancements since transformers themselves. If I recall correctly CNNs were developed by observing that neural networks slowly extracted features. So understanding and improving the mechanisms is the right direction. But other than scaling and training data curation, have they just hit a wall?
We've already dragged transformer to death. Maybe we should revive RNNs or architectures similar to that.
Need to fine-tune a model without the hassle? Try out Crusoe's serverless fine-tuning today! https://www.crusoe.ai/contact-sales/serverless-preview?utm_source=bycloud&utm_medium=influencer&utm_campaign=serverlessfinetuning
This is already a year old idea. See "Scaling Latent Reasoning via Looped Language Models", published October 2025.
I was about to comment how choosing the number of loops at the get-go seemed entirely backwards. Glad to see that they thought of that and tried deferring that choice to the end of the inference step, and saw better performance when they did. Neat!
This video provides the excitement of something mathematically elegant with a slap of reality at the end.