Viewing profile — alexhutcheson

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 (sectio…

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…

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 m…

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.htm…

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 th…

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…

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…

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…

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 and just hand-code your own fix_heap(first, l…

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 va…

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 e…

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…

[flagged]

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…

What makes it the closest to Spring and ASP.NET?

Do you use Flutter Web, or some other framework for web UI?

I’ve seen this book recommended, but have not personally read it: https://www.unshackled.club/book

In the late 2000s and early 2010s, I remember seeing lots of hype around building distributed systems using message queues (e.g. Amazon SQS, RabbitMQ, ZeroMQ, etc.) A lot of compan…

Specifically: https://bazel.build/remote/caching

Modern hash table implementations use vector instructions for lookups: - Folly: https://github.com/facebook/folly/blob/main/folly/container/... - Abseil: https://abseil.io/about/de…

Useful blog post about this sort of situation: https://lethain.com/hard-to-work-with/

The successor project uses musl: https://github.com/landley/toybox/tree/master/mkroot

Angle grinders are useful when working with pipes, though.

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-paralle…