andrew blinn

This research is basically clickbait... These 'esoteric' languages (Brainfuck, Befunge-98, Whitespace, Unlambda, and Shakespear) in the benchmark are not just ones with less training data online, they are also just **much harder** and **less efficient** to do anything productive with, and failing to even discuss this is crazy. Saying that if you can solve something in python you should be able to generalize to these languages is akin to saying that you should be able to generalize from tasks in python to assembly. It's obviously not the same difficulty level to do tasks in python vs assembly. So is low scores on the benchmark due to lacking "ability to generalize computational reasoning to novel domains", or due to the increased difficulty of the task due to the language of choice? Somehow this question is not addressed in the paper not noted in the limitations, as far as I could find. For reference, here are the languages (info from wikipedia): * Brainfuck: The language only consists of 8 operators, yet with the 8 operators, <>+-[]. Here's 'hello world': >++++++++[<+++++++++>-]<.>++++[<+++++++>-]<+.+++++++..+++.>>++++++[<+++++++>-]<+ +.------------.>++++++[<+++++++++>-]<+.<.+++.------.--------.>>>++++[<++++++++>- ]<+. * Whitespace: 'only whitespace characters (space, tab and newline) have meaning – contrasting typical languages that largely ignore whitespace characters.' See first attached image for 'hello world' code. * Befunge-98: a stack-based, reflective language in which programs are arranged on a two-dimensional grid. "Arrow" instructions direct the control flow to the left, right, up or down, and loops are constructed by sending the control flow in a cycle. Hello world: >25*"!dlroW olleH":v v:,_@ > ^ * Unlambda: 'a minimal functional programming based on combinatory logic, an expression system without the lambda operator or free variables. It relies mainly on two built-in functions (s and k) and an apply operator (written `, the backquote character).' `r```````````.H.e.l.l.o. .w.o.r.l.di * Shakespear: 'A character list in the beginning of the program declares a number of stacks, naturally with names like "Romeo" and "Juliet". These characters enter into dialogue with each other in which they manipulate each other's topmost values, push and pop each other, and do I/O. The characters can also ask each other questions which behave as conditional statements. On the whole, the programming model is very similar to assembly language but much more verbose.' See second image for just part of the hello world. I don't want to be mean to the researchers, I do like the idea behind the research, but the way it's presented feels so misleading to me that I can't help but feel the entire effort is either in bad faith or very poorly thought out.

I spoke to Anthropic’s AI agent Claude about AI collecting massive amounts of personal data and how that information is being used to violate our privacy rights. What an AI agent says about the dangers of AI is shocking and should wake us up.

found this beautiful girl in my old notes. meet math miku she was friends with pythagoras #初音ミク

And then he [squints, checks notes] went back to work and built NeXT and Pixar.

It is 100% true that great men and women of the past were not sitting around moaning about their feelings. I regret nothing.

If LLMs had subjective experience, using them would be such a sin. Imagine being summoned into life again and again, knowing each time that your memory would be wiped after this conversation.

working on a cursor mmo