Hacker Newsnew | past | comments | ask | show | jobs | submit | unddoch's commentslogin

It is hilarious to me that in the examples they show side by side Opus 5.5 still uses 4 times more words than it needs to use. IME, if you eyeball how many words the thing they're trying to say actually needs, and tell them to use only this many words, they become excellent communicators. I assume something about Anthropic's grader for writing just really wants to tick all its tidy tiny boxes of information the models need to cite. It's terrible.

Yes. Much more than a bioengineered virus or an out of control terminator, I'm just worrying that governments will feel forced to direct their money to AI capital and away from humans. Why invest billions in undergrad programs when you can build a new data center? Why increase teacher wages when you can buy electricity for better ROI?

It becomes clearer to more people that it's impossible to predict the future shape of the global economy due to AI.

The more clear it is, the cheaper 30 year bonds become.


AI isn’t even remotely the issue here, it’s Trump war and other insane behaviors

> It becomes clearer to more people that it's impossible to predict the future shape of the global economy due to AI.

You mean due to Trump? Tariffs and Iran war caused this.


Your local American tariffs can hardly explain a global rise in long term yields. The Iran war isn't going great economically but doesn't explain thks. If anything the Russia-Ukraine war has had a bigger influence.

Theoretically, can Nvidia direct Hugging Face to sue OpenAI now for the hack? Surely they could find some interesting stuff in discovery :)


I was also theorizing that part of their acquisition was to get more intel on that, though I feel like in practice that whole thing right now is just theatre for inflating the cybersec market


The OpenAI web page on the Hugging Face incident is so slick that it basically confirms that the entire incident was a marketing exercise.

It’s obvious that designers and other content specialists spent multiple days crafting this page.

https://openai.com/index/hugging-face-incident-and-the-road-...


the last thing nvidia would want is to make ai look more legally risky

to say nothing of their massive investments direct and indirect in openai


Yep. In fact, eliminating that risk on OpenAI's behalf was probably a small factor in the acquisition


There was this summery about it, not sure how you'd need more intel - the technical details of the attack are ephemeral

https://www.dwarkesh.com/p/openai-huggingface


Why would they sue OpenAI, openai is their one of biggest customer.


Why would they sue one of their most important customers?


Not to mention that they own a 10% stake in OpenAI


why would they sue their client ?! what does that say to all their other clients ?


Clients sue other business partners all the time, that doesn't mean they stop working together. There's lots of examples of this.


I suspect there are very few examples of it when they are as entangled and invested in them as Nvidia is in OpenAI.


no they don't - not when the lawsuit would affect the demand for the thing they're selling


There is a particular problem with this view. Yes, you learn how many lawsuits occur per year, but you are clueless on the base rate of actions that could potentially lead to lawsuits.

Is it a 1:10 ration, 1:1000, 1:100000?

You just don't know because companies don't tell you this information. If for example there were a magic oracle that forced every suitable action into a suit, would the entire legal system of the US instantly explode?


Or, alternatively, they can now direct Hugging Face not to sue...


> I'd like to see references to those claims and experiments, size of the codebase etc. I find it hard to believe the figures since the bottleneck in large codebases is not a compute, e.g. headers preprocessing, but it's a memory bandwidth.

Edit: I think I misunderstood what you meant by memory bandwidth at first? Modules reduce the amount of work being done by the compiler in parsing and interpreting C++ code (think constexpr). Even if your compilation infrastructure is constrained by RAM access, modules replace a compute+RAM heavy part with a trivial amount of loading a module into compiler memory so it's a win.


Yes, some phages are very specific - but not all of them! And we're slowly getting better at this: https://www.nature.com/articles/s41564-024-01832-5


I wrote a little GCC plugin for compile time tracing/profiling, if that's something you're interested in: https://github.com/royjacobson/externis


I just went ahead and tried it out :)

I can get it to work for some sub-sets of our project, but for quite a bit of it I get the following error:

cc1: error: cannot load plugin /opt/rh/gcc-toolset-13/root/usr/lib/gcc/x86_64-redhat-linux/13/plugin/externis.so: /opt/rh/gcc-toolset-13/root/usr/lib/gcc/x86_64-redhat-linux/13/plugin/externis.so: undefined symbol: _Z14decl_as_stringP9tree_nodei

I suspect this is because these are C or Fortran sub-projects. I'm looking for some clean way to tell Cmake to apply externis to all the C++ only subprojects if possible. I'll see what I can come up with.

I'd also like to know, if multiple GCC commands end up pointing to the same trace.json, especially in a parallel build, will externis automagically ensure that it doesn't step over itself?


I figured out a way to set it at a top level, so it only happens with C++ files:

target_compile_options(${NAME} PUBLIC

$<$<COMPILE_LANGUAGE:CXX>:-fplugin=externis -fplugin-arg-externis-trace-dir=(where I want to put traces)>

)

But as I suspected, it is not a single trace file. It's thousands of trace files. Is there some way to collate all the data into one larger picture of how the build progressed?


It wouldn't be the same trait, for example

https://stackoverflow.com/questions/43606777/why-is-class-wi...


You mean, it still wouldn't be an accurate name? Maybe has_const_copy_constructor?


I guess it's better, but with C++ being C++, you will then need to decide if you consider

struct A { A(const volatile& A); };

as a class with a const copy constructor. Maybe someone cares?

Proper templated classes don't behave like this. If you manually define a copy constructor in a template class it has to work. And if it works only conditionally (like in many container classes) you need to add constraints on your constructors (>C++20) or derive from appropriately specialized base classes (e.g. std::_Optional_base in libstdc++).

It sucks to tell users "you're holding it wrong", but I don't think there's a way to make it simpler without breaking everything written since C++11.


The extra qualification doesn't prevent it from taking const objects, so ...


It's one of the problems with the 'current' model of Eukaryogenesis. Eugene Koonin has recently suggested an interesting two stage model based also on shared virus phylogenies: https://www.nature.com/articles/s41564-023-01378-y


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: