Retired in my 20s. My new hobby is building useful apps. Interviewed by MIT Tech Review CN ‘26|EPYMT No. 1 Graduate
travelling worldwide
Joined July 2022
- Tweets288
- Following77
- Followers123
- Likes400
I run 1,000 computers at once.
I've decided to turn the remote desktop app into a standalone product.
My MacBook pretty much stays home now, running ChatGPT and Claude 24/7, so I barely take it out anymore.
I redesigned the touch gestures and app scaling so each Mac app is properly sized and optimized for iPad, iPhone, and Android screens.
Keyboard input and voice dictation go straight to the Mac. The whole experience feels a lot smoother.
macOS actually works surprisingly well in portrait mode, especially for reviewing code and using a full desktop browser.
I can keep different accounts logged in on separate computers, each with its own workspace, without having to sign into everything on my personal phone.
1
1
152
Running hundreds of AI responses at the same time on one Nvidia GPU.
I’ve been pushing local inference with Qwen3.8-27B for fun.
Squeezing everything out of the GPU by exploring different paths, then distilling them into one answer.(1/4)
1
1
75
Reusing the prompt cache lets me generate multiple answers in parallel. Combine them. Repeat across rounds.
On specific problems, potentially with task-specific fine-tuning, this could help a model compete with models 4–10 times its size. (3/4)
1
3
If Strata keeps repeating and never stops, use temperature=1, top_p=0.95, top_k=20. Thank you me later.
53
The math is correct if you look at it from another angle. It has taken roughly 400,000 years for Homo sapiens to evolve into what we are today.
If any genius understands what I’m saying, today’s AI is still in its pre-training stage if you look back from the future. A four-year-old child can learn incredibly fast because the child is the result of countless generations of evolution; they are not starting from a random state.
Yann LeCun's (@ylecun ) latest talk at ETH Zürich
Scaling LLMs to reach AGI is "impossible"
A large language model is trained on about 30 trillion tokens, which is roughly 10^14 bytes of text and would take a person about 400,000 years to read.
A 4-year-old child receives about the same amount of data, 10^14 bytes, through vision alone in about 1 year and 10 months.
In his view, intelligence is the ability to learn new tasks quickly or perform them without prior training, as a teenager learns to drive in about 20 hours. Scaling increases stored knowledge, but it does not produce this ability to adapt.
----
From "Perfology Clips" YouTube channel, (link in comment)
2
63
I’m giving away $1M in AI token credits through WishRig to help more people start building with AI.
There has never been a better time in history for an individual to leverage technology, build something useful, and create real value for the world.
Build with AI like a pro.
60
Turn your iPhone Duo into a 24/7 computer for GPT, Claude, Gemini, and more, without 20 different subscriptions.
1
134
Turn your phone into a full computer.
Introducing my new product.
It’s time to give your GPT and Claude a 24/7 home.
1
51
As the author of the looped neural model, I want to say that these kinds of “recurrent depth” tricks work extremely well — not only on large models, but on small ones too. One day you’ll see a Qwen3.8 27B perform as well as a 100B model. It’ll just run at a lower TPS. That day will come.
OpenAI’s Astra AI uses a new reasoning approach called “recurrent depth.” Though it can help model costs and performance, researchers are concerned bc it obscures a model’s thinking process, making it more difficult to monitor.
w/ @amir @rocketalignment
theinformation.com/articles/…
2
130
The smartest people see the future before it arrives, move before everyone else notices, and do what’s right—even when it’s difficult.
I wrote a theory many years ago. The true measure of an economy's health is not reflected by conventional economic output —but by the logarithmic rate of change in its resources and its accumulated reserves. Once the strait is closed, the damage dealt to markets, especially China's, will be enormous. Like a tsunami, the impact will come. Just later than you expect — even as people keep buying right now.
137
