The biggest Mystery of LLMs have just been solved
Related videos
A new way to fine-tune LLMs just dropped
bycloud
16.3k views
Massive Island Mystery Solved? | The Curse of Oak Island
HISTORY
37.3k views
Did Trump Just Have a “Biden Moment”?
The Bulwark
106.0k views
The Most Clever Trick To Speedup LLMs
bycloud
17.7k views
Why can’t LLMs just LEARN the context window?
bycloud
30.9k views
The Death of RAG?
bycloud
15.0k views
The Biggest Wealth Shift in 82 Years Has Just Begun
Minority Mindset
43.4k views
The RL Irony in LLMs
bycloud
23.0k views
Groundbreaking Discovery from the Asteroid Belt May Have SOLVED the 3I/ATLAS Mystery ☄️
Stefan Burns
10.8k views
How to Solve the Biggest Problem with AI
Futurepedia
25.9k 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
A new way to fine-tune LLMs just dropped
bycloud
16.3k views
Massive Island Mystery Solved? | The Curse of Oak Island
HISTORY
37.3k views
Did Trump Just Have a “Biden Moment”?
The Bulwark
106.0k views
The Most Clever Trick To Speedup LLMs
bycloud
17.7k views
Why can’t LLMs just LEARN the context window?
bycloud
30.9k views
The Death of RAG?
bycloud
15.0k views
The Biggest Wealth Shift in 82 Years Has Just Begun
Minority Mindset
43.4k views
The RL Irony in LLMs
bycloud
23.0k views
Groundbreaking Discovery from the Asteroid Belt May Have SOLVED the 3I/ATLAS Mystery ☄️
Stefan Burns
10.8k views
How to Solve the Biggest Problem with AI
Futurepedia
25.9k views
Top Comments (10)
finally an LLM that can be wrong all time
This is so nuts that something so basic wasn't figured out until now, while literally Trillions of dollars is being thrown at this stuff. Super crazy and cool.
It'd actually be so good for people's public perception of LLMs to have the mainstream ones be deterministic where the same prompt + temperature + seed gives the same outcome every time. People see them as magic, sometimes stupd, consciousness boxes and that causes a lot of issues
I'm just a hobbyist, but decimal rounding errors in Python have been the bane of my existence. When I run simulations meant to mirror systems written in other languages, I often see periodic differences in the results. I think any programmer would have run into this, because any time you're dealing with decimal numbers there’s always the potential for strange rounding issues. Either way, it's very interesting.
I'm so surprised how new this field really is because this explanation makes perfect sense. It seems so straightforward and simple and yet it was only just documented and effectively debugged now in 2025. That is amazing.
I'm calling it now: A decade from now, some IT security wizard is going to come up with a side channel attack where they use the fact that the rounding behavior is influenced by other user's requests to read other people's prompts. Because they can, and because we cannot have nice things.
Ready to build a site that looks hand-coded—without needing a developer? Launch your site for free at https://framer.link/bycloud, and use code BYCLOUD for a free month on Framer Pro. #FramerPartner
It's always the floating point rounding that screws stuff over
0:25 For some reason, I would have expected the dog catching the flu to have the highest temperature.
In short, chaos theory as a form of butterfly effect due to conppunding floating point errors in arithmetic optimization at runtime batch request. This can be tested by running locally so we have control on the number of requests.
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)
finally an LLM that can be wrong all time
This is so nuts that something so basic wasn't figured out until now, while literally Trillions of dollars is being thrown at this stuff. Super crazy and cool.
It'd actually be so good for people's public perception of LLMs to have the mainstream ones be deterministic where the same prompt + temperature + seed gives the same outcome every time. People see them as magic, sometimes stupd, consciousness boxes and that causes a lot of issues
I'm just a hobbyist, but decimal rounding errors in Python have been the bane of my existence. When I run simulations meant to mirror systems written in other languages, I often see periodic differences in the results. I think any programmer would have run into this, because any time you're dealing with decimal numbers there’s always the potential for strange rounding issues. Either way, it's very interesting.
I'm so surprised how new this field really is because this explanation makes perfect sense. It seems so straightforward and simple and yet it was only just documented and effectively debugged now in 2025. That is amazing.
I'm calling it now: A decade from now, some IT security wizard is going to come up with a side channel attack where they use the fact that the rounding behavior is influenced by other user's requests to read other people's prompts. Because they can, and because we cannot have nice things.
Ready to build a site that looks hand-coded—without needing a developer? Launch your site for free at https://framer.link/bycloud, and use code BYCLOUD for a free month on Framer Pro. #FramerPartner
It's always the floating point rounding that screws stuff over
0:25 For some reason, I would have expected the dog catching the flu to have the highest temperature.
In short, chaos theory as a form of butterfly effect due to conppunding floating point errors in arithmetic optimization at runtime batch request. This can be tested by running locally so we have control on the number of requests.