seeknotfind

Joined 4 years ago

Karma

832

Submissions

448

Member since

Mar 20

View on Hacker News

Recent submissions

Comment · 2 days ago

Ubsan is a misnomer here from a technical perspective. However overflow unsigned or not is unexpected in many cases. As such, sanitizing it (and many other hard to reason about behaviors) by default is great practice. For an example, when unsigned values are used to count, overflows from summing over 2**N often result in incorrect behavior. This may be incrementing or even counting beans.

One particularly famous example of a bug: https://projectzero.google/2023/01/exploiting-null-dereferen...

We usually get away with ignoring unsigned rollover because most application counts don't exceed 2**32 or 2**64. However, many can exceed these values, and in those cases, explicitly handling this case in critical systems is often better than ignoring potentially erroneous behavior.

Another concern, some people might have about this is the performance implications or what to do when you want to allow overflow. In that case you can locally disable overflow protections.

Comment · 2 days ago

Yes, this is the best answer. This also works with `-fsanitize=unsigned-integer-overflow`.

Comment · 22 days ago

Wow, that's pretty bad, but imagine if 50% of software allowed this to happen at any time, and it was discovered on December 1st, 2026. What would happen?

Comment · 1 month ago

I would worry more about 5D chess or 6D and even 7D chess. This article talks about teaching models to be good and methods that could train goodness, but it sounds like there are some unsolved problems to graduate to 8D+ chess.

Comment · 2 months ago

Is this supposed to be scary or amazing?

Comment · 2 months ago

Wow. This is incredible. Turning the entire field of mathematics into a formalized and connected system. An index of mathematical understanding. All fields will undergo this change!!!

My man Terrance Tao, I hope to contribute to your symphony of progress.

If the interrelationships of this are also exposed and searchable, if it can build many bridges inside itself, then this is truly the cipher key to all that can be known.

Comment · 2 months ago

100%. Context is big for AI, but it's nothing compared to everything a human can learn. If you efficiently represent everything in context, it may be many papers, but if AI is actively working through proofs, it will quickly fill up. They're no denying AI is making strides, but pinning it to memory is an oversimplification.

Comment · 3 months ago

We are?

Comment · 4 months ago

If you're calling this across translation units, the calling convention will come with a performance penalty, but boy have we come full circle since pre-ANSI C required you to pass args as a struct. Much love - wish the language required struct and arg list to be the same thing. You can send a list of em and it'll work with algebraic data types for batching calls. The dream. CPU doesn't play nice though since structs aren't register shaped, but maybe they could be in a future calling convention.

Comment · 5 months ago

You can click on that to see more detail :)

Comment · 6 months ago

"I'm sorry". What do HN readers think about how this inclusion changes the meaning of the rest of the email?

Comment · 7 months ago

Did you ever try multiple copies?

Comment · 8 months ago

They're are plenty of mods that inject AI into games, like Skyrim. Though in order for a game to have LLMs and scale to a shipping audience without a subscription, you would want to run them locally.

The real technical blocker is performance and voice synthesis. If voice synthesis was at parity with human actors, it would be worth it to battle the performance aspect for major studios. In text based games especially, taking time to generate enough text is just too slow to be convenient

Comment · 9 months ago

Very cool. Can't wait to see what's next with this project! Congrats on the huge scale of tests/examples as well.