abhgh

15 ms·

121.
by abhgh 3y ago
If you're in the Bay Area and like visiting bookstores that have been around awhile, there are quite a few good ones here: * Bell's books, Palo Alto, ~89 yr. [1] * Moe's books, Berkeley, 65 yr. [2] * Green Apple Books, SF, 57
122.
by abhgh 3y ago
This reminded me of "Starter Villain"! [1]. Don't want to give away the plot, but I'll say its a fun read. [1] https://www.amazon.com/Starter-Villain-John-Scalzi/dp/076538...
123.
by abhgh 3y ago
Yeah, I agree. But I said what I did to make the larger point that empirical studies are hard to vouch for without some level of trust on both sides. For ex if someone selects favorable datasets, metrics, "random" seeds, p-hacking
124.
by abhgh 3y ago
Agree. With more compute-intensive experiments in ML, I am not sure if any reviewer re-runs experiments. The reviewing is already voluntary - who's going to pay out of their pocket for the compute? It's not feasible.
125.
by abhgh 3y ago
Unfortunately the state of meta-review is similar to that of reviews. Rarely delves deeper, mostly acts as a summarizer for the independent reviews.
126.
by abhgh 3y ago
This, and Manning and Schutze [1] were my hooks into NLP. [1] https://nlp.stanford.edu/fsnlp/
127.
by abhgh 3y ago
You're correct except for the use-cases where one of these come into play: A. Latency: for some systems, you need near real-time predictions. LLMs (today) are slow for that. B. Cost: when the low dev. effort (for building and deploying
128.
by abhgh 3y ago
If you're talking about Gradient Boosting, I've a somewhat popular answer on Quora from ages ago [1]. [1] https://qr.ae/pKJPbm
129.
by abhgh 3y ago
Just in case it's useful for anyone, Gemma 2B using ollama runs quite smoothly on raspberry pi 5 too. I have just used ollama's interactive shell so far.
130.
by abhgh 3y ago
Oh wow... I have had the Android port of his puzzles (your second reference links to it) on my phone for a while. Had no idea the developer of Putty had anything to do with it!
131.
by abhgh 3y ago
I got this [1] a few years ago deluding myself into believing that I'll learn a bit from it everyday. Alas, the only way it's useful to me on a daily basis now is as a monitor stand. I remain cautiously hopeful that I'll begi
132.
by abhgh 3y ago
This is interesting. For me it seems like me not reaching out for books is a consequence of lazy "rotation", i.e., these are the same books I'd eventually find time to read if I could see them around often enough, so when I h
133.
by abhgh 3y ago
That's probably the intent, but I don't know if this actually achieves this (I have another comment that's about the use of bayesopt here). But even if it did, bayesopt operates sequentially (it's a Sequential Model-base
134.
by abhgh 3y ago
Thanks for the explanation! I don't entirely understand what two models mean here, because typically the search strategy (or acquisition function) in bayesopt - which in your case seems to be some form of entropy search (ES) - decides
135.
by abhgh 3y ago
What's the BayesOpt maximizing? As in it identifies a subset based on what criteria?
136.
by abhgh 3y ago
I have written scripts in awk (what seems a lifetime ago!), bash, then perl, ruby and python - in that chronological order. I think awk scripting didn't take off for the masses because while it was good for its goal, (1) it was a bit n
137.
by abhgh 3y ago
Like many others here I'm going to recommend Jekyll too, and I recently started blogging with it [1]. Although the primary reason to suggest Jekyll would be that its non-nonsense and has been around a while (which is what the other com
138.
by abhgh 3y ago
Seconded. Great quality and the perfect gift for people who are into mathy things. I have bought two as gifts. Maybe I should have said thirded? The additional paperwork - commentary and notes from Clifford Stoll -that accompanies these bot
139.
by abhgh 3y ago
If you're looking for interesting derivative-adjacent ideas, I would also recommend Clarke derivatives [1]. They occasionally show up in ML papers, e.g., [2], [3]. Unrelated bu tangential, another place where you need derivatives but d
140.
by abhgh 3y ago
I am yet to read up on xcc so this is helpful, thanks. One thing that I found useful as an application from Knuth's lecture (I'm not sure if it's edited out because it was an audience discussion, I attended the 2023 talk in p
141.
by abhgh 3y ago
Like another comment says this is is not recent - these have been up on YouTube for a while. If you're in the bay area, these are free to attend (and recommended if you're a Knuth fan ofc), but reach early because the seats get fi
142.
by abhgh 3y ago
Beta-Bernoulli multi-armed bandits (BB-MAB) are definitely a good way to get started on a Bayesian version of A/B testing where you've the additional benefit that your population is dynamically allocated to the most performant op
143.
by abhgh 3y ago
You can make up by learning more parameters, albeit each parameter is of a lower resolution. The tradeoff evidently works favorably till upto 4-bits; I'm basing this off the results reported on the k-bit inference scaling laws paper by
144.
by abhgh 3y ago
Your comment prompted me to look up the Twitter/X thread. I had no idea that the Processing foundation was suffering from issues of mismanagement to this extent. It is sad to hear :(. I haven't used it for a long time, but I remem
145.
by abhgh 3y ago
"NLP is basically solved" --- somewhat but not the entirely yet. I work on a variety of specialized NLP use-cases for clients and there are different strengths these approaches have. The big thing with LLMs is the ability to deplo
146.
by abhgh 3y ago
As a kid discovering the puns over time was a delight. I still remember revisiting an issue after I had fractured my arm, and doing a happy double-take coming across a couple called Ulna and Radius (these are the bones in the forearm [1])!
147.
by abhgh 3y ago
Have you tried reaching out to dang (hn@ycombinator.com)? One of my submissions was blocked and I reached out to him - he responded a while later to tell me that the domain was blacklisted (it was an article I'd written while working w
148.
by abhgh 3y ago
Came here to comment about this book. Great read and focuses almost entirely on a somewhat uncommon theme within SF: challenges in communication between two differently-abled sentient species. I quite liked the book both for the core SF bit
149.
by abhgh 3y ago
As a followup to this comment, in Python you have the option of using `isclose()` which is present in both `numpy`[1] and the standard `math` [2] libraries. This has been quite helpful for me in comparing small probability values. [1] http
150.
by abhgh 3y ago
I would also recommend Pascal Poupart's talk [1] on Attention and Transformers. Doesn't seem to be cited often, but its quite well done and seeing him work out some of the details using chalk and board is very reassuring. [1] htt

More ›