Toots from 2024-11-20#
“So this entire process, in order to add “what is 2+2”, we do a non-deterministic a lookup …#
“So this entire process, in order to add “what is 2+2”, we do a non-deterministic a lookup from an enormous hashtable that contains the sum of public human knowledge we’ve seen fit to collect for our dataset, then we squeeze it through the tiny, nondeterministic funnels of decoding strategies and guided generation to get to an answer from a sampled probability distribution.”
https://newsletter.vickiboykis.com/archive/why-are-we-using-llms-as-calculators/
Mastodon Source 🐘#
“We hypothesize that this decline is due to the fact that current LLMs are not capable of genuine logical reasoning; instead, they attempt to replicate the reasoning steps observed in their training data. When we add a single clause that appears relevant to the question, we observe significant performance drops (up to 65%) across all state-of-the-art models, even though the added clause does not contribute to the reasoning chain needed to reach the final answer”
https://arxiv.org/abs/2410.05229
Mastodon Source 🐘#
Like coordination points, it’s typically more effective to tightly constrain where (and for who…#
Like coordination points, it’s typically more effective to tightly constrain where (and for whom) the complexity exists than to try and eliminate it.
“This is something I think about constantly. Once you get the feel for it, you can start thinking intentionally about how you want to push around complexity, where it ought to live.
That’s much more liberating than thinking only about eliminating it.”
https://notes.billmill.org/link_blog/2024/11/Complex_for_Whom_.html
Mastodon Source 🐘#
“We can’t fix software security in isolation without fixing the underlying software developme…#
“We can’t fix software security in isolation without fixing the underlying software development practices. And that in turn depends on many things but a powerful factor is the underlying management systems in place. Changing those management systems can’t happen without the existing leaders changing their minds, unlearning and relearning new ways of working.”
https://tldrsec.com/p/systems-thinking-for-cybersecurity-professionals
Mastodon Source 🐘#
Related excellent post: https://kellyshortridge.com/blog/posts/cybersecurity-isnt-special/
Mastodon Source 🐘#
“People want credit for the cleverness of their idea. And the line of thinking goes, more compl…#
“People want credit for the cleverness of their idea. And the line of thinking goes, more complex means more clever, which means they are smarter. Customers don’t care about that. They just want to understand how your product fits into their life and when they should use it. That’s it.”
https://andrewchen.substack.com/p/why-your-product-idea-sounds-too?publication_id=2401262
Mastodon Source 🐘#
“I really think organisational design has a lot in common with system design, you have to figur…#
“I really think organisational design has a lot in common with system design, you have to figure out your pieces and how they interoperate.”
https://makemeacto.substack.com/p/interview-with-cate-huston-author
Mastodon Source 🐘#
“This is an example of what I call “rigor theatre.” Appearing to be rigorous when you’re …#
“This is an example of what I call “rigor theatre.” Appearing to be rigorous when you’re really not. Model rigor is desirable when a precise, exact, definitive answer to a question should determine and constrain future actions. But when it is gratuitously added, it creates extra friction for users, extra work for software developers, and a false sense of having covered your bases.”
https://wirfs-brock.com/rebecca/blog/2024/11/12/
Mastodon Source 🐘#
“And even better, to design systems in a way that minimizes the chances of unintended behaviors…#
“And even better, to design systems in a way that minimizes the chances of unintended behaviors being relied upon in the first place.”
https://abenezerbelachew.com/blog/hyrum-law-in-golang
Mastodon Source 🐘#
“Let’s move on to something even more useful: OnceValue and OnceValues. These are cool becaus…#
“Let’s move on to something even more useful: OnceValue and OnceValues. These are cool because they remember the result of f after the first execution and just return the cached result on future calls.”
https://victoriametrics.com/blog/go-sync-once/
Mastodon Source 🐘#
Bomb cyclone: 1 Transformer: 0#
Bomb cyclone: 1
Transformer: 0
Mastodon Source 🐘#
Bomb cyclone: 2
Transformer: 0