Was asked today “what is ‘Balboa?’” and answered that it was a Caribbean island named after a 16th century Portuguese general etc. before pausing to acknowledge that these assertions were based only in a vague feeling of plausibility rather than fact and this is how I discovered that I am an LLM.
Veritasium has a fascinating video on the subject of the completeness, consistency, and decidability of mathematics.
3. The solvability of problems. From the work of Gödel, John von Neumann, and Alan Turing, we know that there are problems that are true, yet cannot be proven so. This is the ultimate limit of mathematics, and so it is also the ultimate limit of artificial intelligence.
Expanding this ability to be capable of reasoning across the entire human lexicon would require immeasurable effort to encode primitives, concept learning, stories, and analogical reasoning. This isn't exactly impossible, just monumentally time-consuming. This is the field of KBAI.
2. The ability to reason. Modern LLMs are not capable of reasoning in the robust way that we do with propositional logic. There are algorithms that are capable of constructing logical proofs, however they must be provided the logical building blocks.
1. The veracity of information. As the adage goes, "garbage in, garbage out". Without mechanisms to ensure the accuracy of information, AI algorithms will flounder in a cycle of learning inaccurate patterns on bad data, producing inaccurate results, publishing said results, and repeating.
Had a chat with my dad after Thanksgiving dinner about AI. He said "My view of AI is that it is like the universe; it continuously grows and expands as it learns new things. But is there a limit to how much it can grow or what it can do?"
Aside from energy and data storage, 3 limits came to mind.
Found a cool simulation tool for robotics development!
Fibonacci spiral by the artist Anita Chowdry, based on medieval Iranian and Mughal illustrations, 2012.
There are now several benchmarks testing spatial reasoning and agent capabilities of LLMs and VLMs:
- arxiv.org/abs/2410.06468 (does spatial cognition ...)
- arxiv.org/abs/2307.06281 (MMBench)
- arxiv.org/abs/2411.13543 (BALROG) - additional points for the LOTR ref.
e introduce the effective horizon, a property of MDPs that controls how difficult RL is. Our analysis is mo- tivated by Greedy Over Random Policy (GORP), a simple Monte Carlo planning algorithm (left) that exhaustively ex- plores action sequences of length k and then uses m random rollouts to evaluate each leaf node. The effective horizon combines both k and m into a single measure. We prove sample complexity bounds based on the effective horizon that correlate closely with the real performance of PPO, a deep RL algorithm, on our BRIDGE dataset of 155 deterministic MDPs (right).
Kind of a broken record here but proceedings.neurips.cc/paper_files/...
is totally fascinating in that it postulates two underlying, measurable structures that you can use to assess if RL will be easy or hard in an environment
Adding my love letter to
arxiv.org/pdf/2304.01315
Empirical Design in Reinforcement Learning
by
Andrew Patterson, Samuel Neumann, Martha White, Adam White
JMLR 25 (2024) 1-63
#ReinforcementLearning
These aren’t the heroes we deserve, but they are the heroes we need.
You can read all my essays for free on Substack: deconstructivediscovery.substack.com
I aim to continue my analysis of this book. I aim to inform you of the truth about AI and its capabilities. We can avoid the disastrous future that this book hawks. Doing so requires us to be informed and to work for the benefit of others. I hope you’ll join me.
It is a tremendous tool. And like all powerful tools, it can be used for good or for evil. AI can be used to detect cancerous tumors with remarkable precision. It can also hunt down enemy combatants on the frontlines of modern wars.
I find it disheartening that people feel afraid of AI - though I empathize. The truth is that AI isn't anything new. We have had AI for centuries, implemented in other mechanical machines before our moden computers.
I know the relationships are there, but I fear the author's depiction. Such topics deserve the delicate precision of a neurosurgeon's scalpel, not a dull cleaver from the corner market’s butcher.
Still, I find myself lured to read further. The jacket promises an analysis of “the complex relationship between information and truth, bureaucracy and mythology, wisdom and power.”
So if an AI model denies your job application, it’s because a human would’ve denied it too. Thankfully (for the corporate psychopaths), the AI model doesn’t need healthcare and a retirement plan.
AI is anything but alien. Humans design and build every algorithm. Some algorithms call for training (i.e. machine learning). When the algorithm’s aim is to replicate a human’s decisions, humans train the model on human decisions.
I am a machine learning professional. I am also currently enrolled in a master’s program in computer science specifically to study AI. I fumed.
Evidently the author couldn’t withhold their fear-mongering campaign. The anterior flap of the book cover heralds “the age of AI - an alien information network that threatens to annihilate us”.
The New York Times bestselling author bills the book as “a brief history of information networks from the Stone Age to AI”. I expected the neutral, informative voice that is ubiquitous in non-fiction. How wrong was I.
That is why I was so excited to pick up Yuval Noah Harari’s latest book “Nexus”. After reading a few pages, I felt I was witnessing a horrific highway pileup - one tragedy of a paragraph slamming headlong into the next.
I am enthralled by the critical role that truth plays in human society. In the Information Age it has become clear that nations live and die on truth and lies. It is a topic that I rarely see discussed on popular media outlets.
This is the first time that I want to finish a book purely out of spite. It is simultaneously motivating and frightening to know that such baseless drivel can be so widely published. I’ll be writing a critical analysis of this for months to come.
Is it wrong to think that other data structures like linked lists or trees are just variants of graphs? Isn’t it all just a bunch of nodes pointing to each other?
Learning about Graph SLAM in my probabilistic robotics course. The more I learn about CS, the more I’m convinced that anything and everything can be represented by a graph structure.
D&D Combinatorics xkcd.com/3015
