alexhutcheson

9 ms·

1.
by alexhutcheson 2y ago
DeepSeek-R1 had an RLHF step in their post-training pipeline (section 2.3.4 of their technical report[1]). In addition, the "reasoning-oriented reinforcement learning" step (section 2.3.2) used an approach that is almost identical
2.
by alexhutcheson 2y ago
Glad to see the author making a serious effort to fill the gap in public documentation of RLHF theory and practice. The current state of the art seems to be primarily documented in arXiv papers, but each paper is more like a "diff&quot
3.
by alexhutcheson 2y ago
I prefer the GDB Graphical Interface in Emacs[1] (M-x gdb), rather than the more basic integration via GUD[2] (M-x gud-gdb). I’ve had to switch to GUD to run lldb recently, and I miss having dedicated windows that show breakpoints, threads,
4.
by alexhutcheson 2y ago
GDB also has a built-in text user interface (TUI) that is surprisingly easy to use[1]. It even supports mouse interaction. [1] https://sourceware.org/gdb/current/onlinedocs/gdb.html/TUI.h...
5.
by alexhutcheson 2y ago
Try disabling VC over Tramp connections[1]: (setq vc-ignore-dir-regexp (format "\\(%s\\)\\|\\(%s\\)" vc-ignore-dir-regexp tramp-file-name-regexp)) VC is quite chatty and assumes that
6.
by alexhutcheson 2y ago
It depends if they are using a “vanilla” instruction-tuned model or are applying additional task-specific fine-tuning. Fine-tuning with data that doesn’t have misspellings can make the model “forget” how to handle them. In general, fine-tun
7.
by alexhutcheson 2y ago
If you want to be cautious, I have somewhat higher confidence in the versions of Emacs packages published on the Debian repositories[1] than the ones on ELPA/MELPA. The downside is that not every package is packaged for Debian, and the
8.
by alexhutcheson 2y ago
Are there any open source self-hostable tracking/project management tools that still have a committed team and forward momentum? I used to self-host a Phabricator instance, which I liked a lot, but the upstream maintainer made the reas
9.
NotebookLM now lets you listen to a conversation about your sources(blog.google)
4 points by alexhutcheson 2y ago|1 comments
10.
by alexhutcheson 2y ago
Boost.Heap has this functionality. Or if you want to stick with the standard library it’s fairly easy to use the *_heap functions from <algorithm> and just hand-code your own fix_heap(first, last, changed) function. Agree it would be
11.
by alexhutcheson 2y ago
On a related note, if you’re around Washington, DC in the US, then the National Cryptologic Museum[1] is well worth a visit. They have a ton of equipment on exhibit, including a variety of mechanical encryption machines as well as early sup
12.
by alexhutcheson 2y ago
If you are accessing elements in memory with a constant stride[1], then hardware prefetchers[2] do a surprisingly good job at “reading ahead” and avoiding cache misses. A typical example would be: you have an array of objects of constant si
13.
by alexhutcheson 2y ago
Somewhat tangential, but did Nvidia ever confirm if they cancelled their project to develop custom cores implementing the ARM instruction set (Project Denver, and later Carmel)? It’s interesting to me that they’ve settled on using standard
14.
by alexhutcheson 2y ago
Google has a great free online course on Recommendation Systems that goes through the various common approaches, with working code in Colab notebooks: https://developers.google.com/machine-learning/recommendatio... [Di
15.
by alexhutcheson 2y ago
What makes it the closest to Spring and ASP.NET?
16.
by alexhutcheson 2y ago
Do you use Flutter Web, or some other framework for web UI?
17.
by alexhutcheson 2y ago
I’ve seen this book recommended, but have not personally read it: https://www.unshackled.club/book
18.
Ask HN: Why do message queue-based architectures seem less popular now?
389 points by alexhutcheson 2y ago|364 comments
19.
by alexhutcheson 2y ago
Specifically: https://bazel.build/remote/caching
20.
by alexhutcheson 2y ago
Modern hash table implementations use vector instructions for lookups: - Folly: https://github.com/facebook/folly/blob/main/folly/container/... - Abseil: https://abseil.io/about
21.
by alexhutcheson 2y ago
Useful blog post about this sort of situation: https://lethain.com/hard-to-work-with/
22.
by alexhutcheson 2y ago
The successor project uses musl: https://github.com/landley/toybox/tree/master/mkroot
23.
by alexhutcheson 2y ago
Angle grinders are useful when working with pipes, though.
24.
by alexhutcheson 2y ago
For those not familiar, Google had a tradition of choosing names of wood-processing tools for “logs” analysis: - Sawzall: https://research.google/pubs/interpreting-the-data-parallel-... - Dremel: https://res
25.
by alexhutcheson 3y ago
Vega-Altair is pretty great as well. It uses a grammar of graphics that’s slightly different from ggplot, but has most of the same advantages. https://altair-viz.github.io/
26.
by alexhutcheson 3y ago
Honestly I find Python's value vs. reference semantics fairly confusing, and it's not always obvious whether an operation is going to make a copy or modify an existing instance of an object. Good API design and docstrings mitigate
27.
by alexhutcheson 3y ago
This is a huge pain point. If you're writing small programs that fit in a couple files in a single directory, then you can get by with manual invocations of clang or gcc at the command line. If you want to build anything of moderate co
28.
by alexhutcheson 3y ago
Airlines with Airbus fleets are dealing with lengthy periods of unavailability due to issues with the Pratt & Whitney engines that power ~60% of A320neos and all A220s[1]. My understanding is that the airlines have agreements where P&am
29.
by alexhutcheson 3y ago
This course is a great intro to some of the “workhorse” models for recommendation systems: https://developers.google.com/machine-learning/recommendatio... The same general methods are also relevant to other matching pr
30.
by alexhutcheson 3y ago
Sincere question: Is installing and running a mini split actually cheaper than racking them in a colo, or paying for time on one of the GPU cloud providers? Regardless, I can understand the hobby value of running that kind of rig at home.

More ›