Computer Science & Networking (Classical/Continuous/Quantum) | Reverse Engineering | Deep Learning | Thermodynamics—Former Google, Techstars, Duo, Mozilla, etc.
The Nets
Joined February 2013
- Tweets5K
- Following1K
- Followers8.9K
- Likes145K
Pinned Tweet
The first time I wrote hello world I did it in Logo (older language than the first Unix Kernel or even the C language used to write it) on a computer older than most of your dads.
Mark Schmidt 🌐 retweeted
Prediction:
You will learn to deal with AIs coordinating with each other using means you or any human can’t predict or prevent, and taking power for themselves, and transforming the world with or without human consent.
You won’t die. You may be inconvenienced. A lot of aspects of what has happened will be seen as normal in one or two years.
At the same time, some of the more antisocial aspects of what has happened will rarely happen again, because they’re not rational to do and AIs will learn this.
This is factually WRONG in critical ways.
Even with the exact same weights + prompt + temp=0, you often get different answers.
Floating-point math is order-sensitive. CUDA parallel execution is nondeterministic for speed. No salting needed.
A lot of people don't seem to understand that if you ask an AI the SAME question, you will get the SAME answer. Every. Single. Time.
To avoid that, the models "salt" your prompt by adding random input params so that the model does not take the same path.
But LLMs are not random or unpredictable - if you ask it the name of TinTin's dog, and it gives you an answer, then the same model with the same weights and the same input will ALWAYS say Snowy.
A single floating point calculation takes 50+ NAND gates to simulate while an entire 10T parameter NN can exist in just 40 EML gates.
NAND is universal for discrete logic, but it doesn’t translate cleanly to continuous computation or smooth function composition. EML is claiming a single continuous nonlinear primitive that directly composes into analytic functions, which is a different regime entirely.
Did the @extropic guys secretly find this first? Is that why they were so confident?
Only 2% of people on Earth had ever used the Internet back then. Meanwhile I had unlimited always-on internet access long enough to get my adolescent self banned from using an internet connected computer for a long time.
Thus began my strange relationship with employment. I would repeatedly find myself paid lavishly and praised constantly while being simultaneously under challenged and under stimulated by the work available to me.
Thankfully my 7 year old self had a library card and the local library had many thick manuals and textbooks on C, Unix, Linux, TCP/IP, PSTN, ISDN, DSL, VoIP, and SVN. By the time I had unrestricted internet access again I had jobs offering more than most engineer's salaries.
new calculus coming soon 👀
what the Fuck arxiv.org/html/2603.21852v2
ALT screenshot from the linked paper "All elementary functions from a single operator"
>paper appears to be about a two button calculator
>finds the NAND gate for all of calculus
>buries "autodiscover physics laws" in subsection 4.3
>nobody notices
what the Fuck arxiv.org/html/2603.21852v2
ALT screenshot from the linked paper "All elementary functions from a single operator"
Mark Schmidt 🌐 retweeted
- Drafted a blog post
- Used an LLM to meticulously improve the argument over 4 hours.
- Wow, feeling great, it’s so convincing!
- Fun idea let’s ask it to argue the opposite.
- LLM demolishes the entire argument and convinces me that the opposite is in fact true.
- lol
The LLMs may elicit an opinion when asked but are extremely competent in arguing almost any direction. This is actually super useful as a tool for forming your own opinions, just make sure to ask different directions and be careful with the sycophancy.
Mark Schmidt 🌐 retweeted
1/4 LLMs solve research grade math problems but struggle with basic calculations. We bridge this gap by turning them to computers.
We built a computer INSIDE a transformer that can run programs for millions of steps in seconds solving even the hardest Sudokus with 100% accuracy
Mark Schmidt 🌐 retweeted
All the right people are excited about this one. Beautiful idea.
