HNHacker News
TopNewBestAskShowJobs

curao_d_espanto

2 karma · joined March 10, 2026

submissionscomments
curao_d_espanto··on How Is Compression Prediction?
excellent article, I will happily follow it but the intro caught me well, it reminded me a bit of Shannon's Bandwagon https://reimbar.org/papers/bandwagon/ ; in computational neuroscience the problem setup is quite diverse around literature, with some "big" normative theories kind of accepted, but very far imho from that concepts and derivations like the channel capacity theorem become actually useful in terms of expalantory power or technology, they remain some niche work or some left away approaches in general. Reminded me also of last Tishby's work about explanations of deep learning training dynamics using the information bottleneck method. I ignore though and wonder if approaches like this are used as train targets on LLMs or something or they are just used to philosophize around (which I actually enjoy)
curao_d_espanto··on GitHub removed the old copilot multipliers on a pricing page
Copilot Pro+ having 7000 "AI credits" monthly is insanely low considering that a single gpt5.4 request would use 250 credits (28 request per month), and 5.5/opus using 500 (and even more depeding on the output size!), which is even 14 request per month. I dont know if I am missing something but the business model they are pursuiting is based on that i can make 1 request a day/every two days of the month for a full 2h run of one prompt? So far the best combo I've found is just claude code + deepseek v4, copilot cli became even more unusable than claude pro subscription around January/February when with a single question and getaround of code depleted the 5h session limit.
curao_d_espanto··on Billion-Parameter Theories
> The emerging field of mechanistic interpretability suggests otherwise. Researchers are developing tools to understand how neural networks do what they do, from network ablation and selective activation to feature visualization and circuit tracing. These techniques let you study a trained model the way a biologist studies an organism, through careful experimentation and observation.

honestly, when I read that part of the article I imagined that author never studied how computers were made and where the engineering ideas came from, all technology just "popped" and here we are talking about complexity and stuff like the LLM is truly alive