a hilariously dramatic video on a fascinating corner of control theory: Insects are probably aerodynamically unstable and the obvious implication is that they're actively making adjustments but vibration can stabilize without any feedback
youtu.be/Tnv186IFovQ
I'm not in a place where I can evaluate the accuracy of this but the writing is quite good and entertaining and the story is convincing.
I would love to read more articles of this tone and quality.
I hit a bug in the Attention formula that’s been overlooked for 8+ years. All Transformer models (GPT, LLaMA, etc) are affected.
Researchers isolated the bug last month – but they missed a simple solution…
Why LLM designers should stop using Softmax 👇
evanmiller.org/attention-is-…
AI diary: it gives you a phone call every evening, asks you how your day was, asks thoughtful follow-up questions. Your answers are transcribed and emailed to you w summary and comparisons to previous entries.
pretty funny that the original mechanical turk was a human in guise of a machine while the modern iteration is a machine in guise of humans
One of our key sources of human data is no longer fully “human"!
We estimate that 33-46% of crowd workers on MTurk used large language models (LLMs) in a text production task - which may increase as ChatGPT and the like become more popular and powerful.
arxiv.org/abs/2306.07899
The existence of undetectable backdoors seems to suggest some kind of impossibility result for alignment: you have no idea what your optimizer might have planted in there: arxiv.org/abs/2204.06974
Some of programming is like writing, an evergreen skill of synthesizing solutions out of whatever tools are available. However, a lot of programming is like scraping parchment and sharpening and tempering quills; you’ll want to find a mentor who can tell you the difference!
GPT-4 is still held back by the tokenizer: it still cannot reliably fulfill requests such as “don’t use the letter ‘e’” or “no words with more than eight letters”. It does much better than GPT-3 though!
AI alignment seems like too low of a bar.
Nuclear weapons are very well aligned; so far they have only committed mass-murder when it was asked of them.
It should be easy to train a model which exceeds human performance at quantun soccer: gregegan.net/BORDER/Soccer/S…
Many problems can be solved given just a little intuition and models can pick up that intuition in a wide variety of domains, not just the ones we're good at.
Could the success of this be partially explained by compute-optimal/chinchilla (Hoffmann et al., 2022)? By that paper OPT-175B seems wildly undertrained, BLOOM-176B is less clear, but they note that larger models are easier to sparsify: larger models are more undertrained!
Another fanciful connection: to the degree that smaller networks can use superposition to emulate larger (sparse!) networks ( transformer-circuits.pub/202… ), one might guess the post-pruning weights are implementing superposition, this would allow 50% pruning w little loss
