An AI diary

This page serves as a collection of most relevant ideas and technical advancements I have been exposed to and which I believe will remain relevant in time. Essentially this is what I believe to be worth retaining from AI news feeds.

Ideas

Mental Model

“I am the thinker, let the agents find the answers to my questions”

From Software Exploits to Tax Loopholes

Even more interesting are the broader implications. The same searching, pattern-matching and reasoning capabilities that make these models so good at analyzing software almost certainly apply to similar systems. The tax code isn’t computer code, but it’s a series of algorithms with inputs and outputs. It has vulnerabilities; we call them tax loopholes. It has exploits; we call them tax avoidance strategies. And it has black hat hackers: attorneys and accountants.

Just as these models are finding hundreds of vulnerabilities in complex software systems, we should expect them to be equally effective at finding many new and undiscovered tax loopholes.

Bias

LLMs are trained on human produced content, which is biased (cultural, ideological, political, racial, etc.). China, India, and every other country in the world has to train its own AI models not just to not be economically dependent on the West, but also to avoid being exposed to Western biases daily and, potentially, being manipulated by them. Conversely, what happens if a Chinese biased model is offered for free and consumed by the West?

Access to the less educated

A plugin for ChatGPT corrects text with any kind of mistakes in it. It could allow to type faster but on the other hand it could open opportunities to people with very low education.

Technical

Deterministic LLM Inference

It is possible to have deterministic inference.

On vulnerability discovery and patching

Finding vulnerabilities can be stochastic because it has a very clear and effective verification function: either the exploit works or it does not. Expolit code can be messy and convoluted, as it is not going to be maintained after the vulnerability is fixed.

Vide-coding fixes does not have that kind of verification function: the fix must not only close the specific vulnerability, but also not introduce new ones or re-introduce old ones, and it has to be maintanable in the future.

The Genie Coefficient

In folklore, genies—and other magical beings—grant wishes literally, not how the wisher intended. King Midas asked that everything he touched turn to gold, and starved. The sorcerer’s apprentice wanted the broom to fill the cistern, and it performed its task so well that it flooded the house.

We now have machines that do this. Ask a modern AI agent to save money on your phone plan and it might simply cancel the plan. Tell it to book a flight, and it might hack the airline website to override restrictions. Each time, it recognizably completed the task you set, but it didn’t do what you would have wanted.

The gap is between the words we use and what we mean by them. We call that gap the Genie coefficient.

Wow moments

Wishlist