AGI and ASI
Published:
If you ask me about the timelines for AGI, artificial general intelligence, and ASI, artificial superintelligence, I would say:
Posts by Tags
AGI
AI
The Business Model of LLMs
Published:
Ever since ChatGPT was released, I’ve been thinking about the business logic and competitive dynamics of large language models. I’ve struggled to find a good framework for it. A few days ago, DeepSeek V4 came out. I wrote a long post critiquing some of V4’s design decisions. And suddenly, three years of unresolved business questions clicked into place.
DeepSeek V4 Technical Report: First Impressions
Published:
DeepSeek V4’s technical report had just been released, and I opened it immediately. The amount and breadth of information in it, and the technical depth, are vast.
Simplicity: On Paper Writing Styles
Published:
I think the best writing style for a paper must be extremely simple. So simple that reviewers reject it for not being innovative enough. Only then can the paper be considered fully revised.
How to Train 100B+ Large Models on a Single GPU
Published:
Small GPU, huge model. The GPU-poor are about to rise up!
Load Balancing Is All You Need
Published:
I have been on Zhihu for a few years, but this is the first time I am actually promoting my own paper.
Maybe We Shouldn’t Implement Continual Learning?
Published:
The problem with biomimetics is that the way carbon-based organisms implement something is not necessarily the optimal design for silicon-based organisms.
AGI and ASI
Published:
If you ask me about the timelines for AGI, artificial general intelligence, and ASI, artificial superintelligence, I would say:
DeepSeek R1: First Impressions and Reactions
Published:
While reading the DeepSeek R1 paper during the day, all I could think about was R1-Zero.
Pseudoscience
Published:
I generally call artificial intelligence pseudo-science, pseudo-math, and pseudo-engineering: a trinity.
A Kind of Idea I Kill Off Subconsciously
Published:
I often make one kind of mistake: when an idea lacks mathematical beauty, I subconsciously kill it off.
In 2025, Has Post-Training Become More Important Than Pretraining for Large Models?
Published:
Hard to say; the more I look at post-training, the more unnatural it feels.
If You Had o4-Level Models, What Previously Impossible Research Would Become Possible?
Published:
Over the past couple of days I went to UT Austin to visit a friend, and I sat in on a few physics classes with him.
From Muon to Gradient Clipping: Some Thoughts on QK Stability
Published:
Prelude: an elegant theory and a stubborn problem
How Do You Come Up With a Workable Idea?
Published:
1.
In Deep Learning, Which Paper Has the Most Astonishing Idea?
Published:
I once dreamed of a paper like this.
ASI
AGI and ASI
Published:
If you ask me about the timelines for AGI, artificial general intelligence, and ASI, artificial superintelligence, I would say:
DeepSeek
DeepSeek V4 Technical Report: First Impressions
Published:
DeepSeek V4’s technical report had just been released, and I opened it immediately. The amount and breadth of information in it, and the technical depth, are vast.
LLM
The Business Model of LLMs
Published:
Ever since ChatGPT was released, I’ve been thinking about the business logic and competitive dynamics of large language models. I’ve struggled to find a good framework for it. A few days ago, DeepSeek V4 came out. I wrote a long post critiquing some of V4’s design decisions. And suddenly, three years of unresolved business questions clicked into place.
DeepSeek V4 Technical Report: First Impressions
Published:
DeepSeek V4’s technical report had just been released, and I opened it immediately. The amount and breadth of information in it, and the technical depth, are vast.
How to Train 100B+ Large Models on a Single GPU
Published:
Small GPU, huge model. The GPU-poor are about to rise up!
Load Balancing Is All You Need
Published:
I have been on Zhihu for a few years, but this is the first time I am actually promoting my own paper.
Maybe We Shouldn’t Implement Continual Learning?
Published:
The problem with biomimetics is that the way carbon-based organisms implement something is not necessarily the optimal design for silicon-based organisms.
DeepSeek R1: First Impressions and Reactions
Published:
While reading the DeepSeek R1 paper during the day, all I could think about was R1-Zero.
In 2025, Has Post-Training Become More Important Than Pretraining for Large Models?
Published:
Hard to say; the more I look at post-training, the more unnatural it feels.
From Muon to Gradient Clipping: Some Thoughts on QK Stability
Published:
Prelude: an elegant theory and a stubborn problem
MoE
Load Balancing Is All You Need
Published:
I have been on Zhihu for a few years, but this is the first time I am actually promoting my own paper.
academia
What Counts as Paper-Padding in Academic Research Today?
Published:
One day, an academic heavyweight wrote a paper titled Tomatoes Are All Eggs Need, which for the first time proposed the method of making tomato fried eggs.
business
The Business Model of LLMs
Published:
Ever since ChatGPT was released, I’ve been thinking about the business logic and competitive dynamics of large language models. I’ve struggled to find a good framework for it. A few days ago, DeepSeek V4 came out. I wrote a long post critiquing some of V4’s design decisions. And suddenly, three years of unresolved business questions clicked into place.
continual learning
Maybe We Shouldn’t Implement Continual Learning?
Published:
The problem with biomimetics is that the way carbon-based organisms implement something is not necessarily the optimal design for silicon-based organisms.
education
When Did Math Start Becoming Counterintuitive?
Published:
Math seems to have been counterintuitive all along.
ideas
A Kind of Idea I Kill Off Subconsciously
Published:
I often make one kind of mistake: when an idea lacks mathematical beauty, I subconsciously kill it off.
How Do You Come Up With a Workable Idea?
Published:
1.
math
Pseudoscience
Published:
I generally call artificial intelligence pseudo-science, pseudo-math, and pseudo-engineering: a trinity.
When Did Math Start Becoming Counterintuitive?
Published:
Math seems to have been counterintuitive all along.
optimization
Pseudoscience
Published:
I generally call artificial intelligence pseudo-science, pseudo-math, and pseudo-engineering: a trinity.
From Muon to Gradient Clipping: Some Thoughts on QK Stability
Published:
Prelude: an elegant theory and a stubborn problem
reasoning
If You Had o4-Level Models, What Previously Impossible Research Would Become Possible?
Published:
Over the past couple of days I went to UT Austin to visit a friend, and I sat in on a few physics classes with him.
reflection
When Did Math Start Becoming Counterintuitive?
Published:
Math seems to have been counterintuitive all along.
In Deep Learning, Which Paper Has the Most Astonishing Idea?
Published:
I once dreamed of a paper like this.
reinforcement learning
DeepSeek R1: First Impressions and Reactions
Published:
While reading the DeepSeek R1 paper during the day, all I could think about was R1-Zero.
research
Simplicity: On Paper Writing Styles
Published:
I think the best writing style for a paper must be extremely simple. So simple that reviewers reject it for not being innovative enough. Only then can the paper be considered fully revised.
A Kind of Idea I Kill Off Subconsciously
Published:
I often make one kind of mistake: when an idea lacks mathematical beauty, I subconsciously kill it off.
If You Had o4-Level Models, What Previously Impossible Research Would Become Possible?
Published:
Over the past couple of days I went to UT Austin to visit a friend, and I sat in on a few physics classes with him.
How Do You Come Up With a Workable Idea?
Published:
1.
What Counts as Paper-Padding in Academic Research Today?
Published:
One day, an academic heavyweight wrote a paper titled Tomatoes Are All Eggs Need, which for the first time proposed the method of making tomato fried eggs.
In Deep Learning, Which Paper Has the Most Astonishing Idea?
Published:
I once dreamed of a paper like this.
satire
What Counts as Paper-Padding in Academic Research Today?
Published:
One day, an academic heavyweight wrote a paper titled Tomatoes Are All Eggs Need, which for the first time proposed the method of making tomato fried eggs.
training
How to Train 100B+ Large Models on a Single GPU
Published:
Small GPU, huge model. The GPU-poor are about to rise up!
In 2025, Has Post-Training Become More Important Than Pretraining for Large Models?
Published:
Hard to say; the more I look at post-training, the more unnatural it feels.
writing
Simplicity: On Paper Writing Styles
Published:
I think the best writing style for a paper must be extremely simple. So simple that reviewers reject it for not being innovative enough. Only then can the paper be considered fully revised.