randomness

 i guess the post frequency is gradually becoming once a week T_T anyways i am EXPECTED to post smth on substack soon so 👀lol what are these blogger emojis here is most recent quanta article i'm reading: new pyramid shape that always lands the same side up  - neat application of this thing -- space exploration! and it's interesting how ppl can just visualize these things, as someone who sucks at geometry T_T, esp in higher dimensions - conway was brilliant.  i'll just paste some literature review i did on some neural network stuff, some of the stuff is like copy-pasted from the paper abstracts (sorry!) Activation Anomaly Analysis (Mar 2020)  a novel approach for anomaly detection based on the hidden activation patterns of NNs, semi-supervised, purely data-driven anomaly detection solution, transferability of algorithm comprised of two parts:  a target network unrelated to the anomaly detection task  an alarm network analyzing the target’s activations Expe...

shap model (SHapley Additive exPlanations, great acronym btw) - interpreting machine learning model predictions based on game theory, model agnostic/ad hoc approach, shows how much a certain feature pushed the output up or down speculative decoding  - make ai more efficient (algorithmic details here ), calculate tokens autoregressively but in parallel to be more efficient, fast approximative function for decoding i tried to watch this video on sheafification but i didn't understand anything other than the fact that i've basically forgotten everything from topology um really cool article about how scientists are using ai for manipulating the brain, takes the analogy between ai and neuroscience a bit further --> developments in ai can help us improve our understanding of the brain and develop applications like readers to help ppl with dyslexia. i think the interviewee's response to the question about the ethics of having a digital copy of a brain to be very insightful, ...

yes im very late this article argues about how mech interp is not that useful and it raises a lot of good points (i.e. in biology / other complex systems, we don't do the bottom up approach, and a lot of the buzz from mech interp comes from cherrypicked results, the compression involved leads to the loss of edge cases, so much has been invested but not much has come out, this post talks about how google stopped prioritizing SAEs because they weren't performing as well). i think i should definitely keep this in mind as i get more interested in interpretability; i've heard friends make such comments too. meanwhile this article by dario amodei  urges for the importance of mech interp, tho he is from anthropic which may influence his viewpoints. he advocates for the govt to also get involved, citing all the bad disasters that may happen because ai is too "opaque". i get his points too; the race between the development of ai models and our understanding of them is pret...

 hiiii here's an over due post (will maintain daily posts !) on why we need ai safety - specification gaming article by google deepmind  here's a quote that sums up the whole thing - reward function / intent engineering is a big thing How do we faithfully capture the human concept of a given task in a reward function? How do we avoid making mistakes in our implicit assumptions about the domain, or design agents that correct mistaken assumptions instead of gaming them? How do we avoid reward tampering? vanishing gradient problem - occurs potentially when u use sigmoid activation, slower learning in beginning --> use relu sum-free sets  and the beauty of a grad student expanding on other's work, smth about littlewood norm and fourier analysis being powerful reverse computing  -- potential benefits for energy conservation, has been theorized and engineers are trying to make it happen! this feels like a strange but very interesting idea and in the end its sorta a tra...