gtoubassi

Its really clever and minimal approach to getting (for my use case at least) the moral equivalent of a cloudflare tunnel without actually depending on CF. Super cool.

bitbang-webcam fires up your camera and gives you a link for anyone to view. Super cool as a quick way to relay vid. WebRTC is CRAZY!

This is super cool and totally works on my crappy low end LG tv. WebRTC is awesome. Nice

Hackers by Steven Levy is an incredible story of the industry’s early years (60-80’s) and the characters that were in it for the “love of the game” vs what is more common now (“status and money”). A lot of heroes like woz, but who are less well known in this day and age (Gosper and Greenblatt!). If you are familiar with and a fan of Dealers of Lightning or Dream Machine, check out Hackers! (this is not a paid endorsement).

"The Last Lone Inventor: A Tale of Genius, Deceit, and the Birth of Television" is a great book detailing the Farnsworth journey.

If you have any details written up on your kit (in partic what solar you used) I'd appreciate a link. I'm looking to do similar

We do the token counting on our end literally just running tiktoken on the content chunks (although I think usually its one token per chunk). Its a bit annoying and I too expected they'd have the usage block but its one line of code if you already have tiktoken available. I've found the accounting on my side lines up well with what we see on our usage dashboard.

I struggled to get an intuition for this, but on another HN thread earlier this year saw the recommendation for Sebastian Raschka's series. Starting with this video: https://www.youtube.com/watch?v=mDZil99CtSU and maybe the next three or four. It was really helpful to get a sense of the original 2014 concept of attention which is easier to understand but less powerful (https://arxiv.org/abs/1409.0473), and then how it gets powerful with the more modern notion of attention. So if you have a reasonable intuition for "regular" ANNs I think this is a great place to start.

+1 you beat me to the punch! I think its helpful to start with simple RL and ignore the "deep" part to get the basics. The first several lectures in this series do that well. It helped me build a simple "cat and mouse" RL simulation https://github.com/gtoubassi/SimpleReinforcementLearning and ultimately a reproduction of the DQN atari game playing agent: https://github.com/gtoubassi/dqn-atari.

Token counting is importing when you are injecting fetched data into the prompt to make sure you don't overflow the prompt size (e.g. in retrieval augmented generation). You want to give the LLM as many facts as will fit in the prompt to improve the quality of its response. So even with billions of dollars... token counting is a thing.

+1 the book is entertaining (esp for engineers). Also was released under the title "The Newtonian Casino".

https://github.com/antimatter15/alpaca.cpp has links

I agree this is an incredibly interesting paper. I am not a practitioner but I interpreted the gradient article differently. They didn’t directly find 64 nodes (activations) that represented the board state as I think you imply. They trained “64 independent two-layer MLP classifiers to classify each of the 64 tiles”. I interpret this to mean all activations are fed into a 2 layer MLP with the goal of predicting a single tile (white, black, empty). Then do that 64 times once for each tile (64 separately trained networks).

As much as I want to be enthusiastic about this, it’s not entirely clear to me that it is surprising that such a feat can be achieved. For example it may be possible to train a 2 layer MLP to predict the state of a tile directly from the inputs. It may be that the most influential activations are closer to the inputs then the outputs, implying that Othello-GPT itself doesn’t have a world model, instead showing that you can predict board colors from the transcript. Again, not a practitioner but once you are indirecting internal state through a 2 layer MLP it gets less obvious to me that the world model is really there. I think it would be more impressive if they were only taking “later” activations (further from the input), and using a linear classifier to ensure the world model isn’t in the tile predictor instead of Othello-GPT. I would appreciate it if somebody could illuminate or set my admittedly naive intuitions straight!

That said, I am reminded of another OpenAI paper [1] from way back in 2017 that blew my mind. Unsupervised “predict the next character” training on 82 million amazon reviews, then use the activations to train a linear classifier to predict sentiment. And it turns out they find a single neuron activation is responsible for the bulk of the sentiment!

[1] https://openai.com/blog/unsupervised-sentiment-neuron/

+1. I've benefited greatly from your content, e.g. your CNN lecture was incredibly accessible [0]. I still find transformers stubbornly elude my intuitions despite reading many descriptions. I would very much appreciate your video lecture on this topic.

[0] I think https://www.youtube.com/watch?v=LxfUGhug-iQ