bo1024

The pacers have to obey the same rules as every other runner. They start at the starting line, they can't run inside the lines, etc.

One of the reasons that Eliud Kipchoge's "Breaking 2" marathon wasn't a legal world record was that teams of pacers would rotate in, take a break, then rotate in again. This way he could have pacers with him almost the entire way. But in a real world record attempt, the winner is always running by themselves in the last part of the race because all of the pacers have been left behind or dropped out -- by definition of a world record, they can't keep up to the end.

What is the political goal behind this directive? I assume there is some completely non-subtle purpose, but I can't tell what it is.

Qwen3.5-122B is actually Qwen3.5-122B-A10B. The A10B means that this is a "mixture of experts" model where only 10B parameters are activated at a given time. Whereas Qwen3.6-27B is a "dense" model where all 27B parameters are activated all the time. So for many tasks, you'd expect the 27B dense model to be better than the 122B-A10B model.

Plus, this is more fuel for Anthropic's God complex.

Is it much different in other countries?

What is the policy and guidance you gave the students regarding LLM use?

There are basically two kinds of applications. One is where you want to correctly solve the problem at least 99 out of 100 times. LLMs generally don't (and not everybody realizes that) so there are a lot of debates and research around how useful and reliable they are or how to make them so.

The other kind of application is where you can try 100 times and you only need to be right once. Solving a mathematical research problem is like that.

There are about 2^64 more 64-bit integers than 32-bit integers.

What's wrong with a 20% tax? We who make a living from labor instead of capital pay more than that.

Paul tries to frame it as an increase of 20% in the tax rate, but in reality the increase is from 0% to 20%, and it's hard to see why that's unfair.

The reason I say it's currently 0% is of course that for the wealthy most of these 5% gains are unrealized (e.g. inflation in the value of their assets) and untaxed.

Your current understanding also seems a bit warped. The US government provides the source of a lot of research funding but historically exerts little "control". Generally grant applications are evaluated by panels of researchers who don't work for the federal government.

Also, this government funding supports fundamental innovations that private companies wouldn't fund because it's too general and too far from monetizable. But after those breakthroughs happen funded by public research, private industry benefits enormously. This includes most health and medical advances and the science underlying most technological advances. So government funding doesn't conflict with the work being necessary or important, on the contrary, it is possibly more important long-term.

Disclaimer? The government funds some of my research.

It's much easier than it seems.

* There are axioms, there are models, and there are theorems.

* A model is a particular structure with objects and relationships. The "standard model of arithmetic" is just the natural numbers 0, 1, 2, ... with normal rules of addition and subtraction and so on. A different model could be the real numbers, or one that includes infinitesimally small numbers, or so on.

* A set of axioms are rules about how a model can work. For example, we can have an axiom for any set of objects called "numbers" with an operation called "addition" that the operation must be commutative (x+y = y+x). The standard model above is one model that satisfies this axiom.

* A theorem is a fact that can be true or false about a given model. For example, "the model has infinitely many objects." If we can prove a theorem from a set of axioms, then that theorem is true for every model that satisfies the axioms. However, there can also be theorems that are true for one model that satisfies the axioms but false for a different model.

Godel's completeness theorem says that if a theorem is true for every model that satisfies a set of axioms, then one can prove that theorem from the axioms.

Godel's first incompleteness theorem says that in any axiomatic system (sufficiently complex) there are theorems that are neither always true nor always false. In other words, there is a theorem that is true for some model of the axioms but false for some other model of the axioms.

I'm not following your reasoning about the common denominator, not sure we're on the same wavelength about what I meant. I'm claiming that in order for an application to be "reclaimable", you have to be able to access and manipulate the data under the app. Some applications currently work that way now, lots of them don't.

For example, we can "reclaim" non-DRM ebook readers, audiobooks, and music players that play local files or use an open API. But a company-specific walled garden streaming DRM'd ecosystem will be almost impossible to build around.

The common denominator: the data needs to be owned by you, or at least made accessible. Companies love to create walled gardens where they own the content and control how you access it, making this kind of personalized interface impossible. Hopefully we can push back more now.

Thanks for the reply. Yeah, I think all of your filtering and categorizing makes these analyses really nice.

Nice article. One small comment, it's very hard to conclude anything about accuracy over time because the comparisons may not be apples to apples. For example if there used to be lots of questions about if it will rain in Boston and now there are lots of questions about if it will rain in Phoenix, it will look like predictions are getting more accurate, but the questions are just getting easier.

Sewing machine.

https://xkcd.com/2754/

Cool stuff. I think there have been projects recently that use LLMs to encode messages in plain text by manipulating the choices of output tokens. Someone with the same version of the LLM can decode. Note sure where to find these projects though.

If Claude code is written by Claude code, and AI outputs are not currently considered copyrightable, then how is Anthropic asserting copyright over the leak?

No, it's not enough. Maybe if the bias is 10% or more.

Not saying this gets through to people, but copyright is purely about the legal ability to restrict what other people do. Whereas property rights are about not allowing others to restrict what you do (e.g. by taking your stuff).

Interesting. I don't quite agree. It's one thing to predict what general topics will be hot and popular this year. But that's not the same as what particular research problem will be important and have lasting influence.

There are a few kinds of important research. One is solving a well-defined, well-known problem everyone wants to solve but nobody knows how. Another is proposing a new problem, or a new formulation of it, that people didn't realize was important.

There is also highly-cited research that isn't necessarily important, such as being the next paper to slightly lower a benchmark through some tweaks (you get cited by all the subsequent papers that slightly lower the benchmark even further).

I agree that (while the ethics of this are a different issue) the copyright question is not obviously clear-cut. Though IANAL.

As the LGPL says:

A "work based on the Library" means either the Library or any derivative work under copyright law: that is to say, a work containing the Library or a portion of it, either verbatim or with modifications and/or translated straightforwardly into another language. (Hereinafter, translation is included without limitation in the term "modification".)

Is v7.0.0 a [derivative work](https://en.wikipedia.org/wiki/Derivative_work)? It seems to depend on the details of the source code (implementing the same API is not copyright infringement).

Yes, and it opens by talking about STEM fields. I consider CS part of both STEM and science generally.

This is not how computer science publishing works, however. Post it on arxiv, submit to a conference, get 3 peer reviews, accepted, “published”. 99% of papers are effectively open access for free.

Logically and legally equivalent to "we will keep your data forever unless legally required to delete it.".

That's such a bummer and seems like poor design. It ought to be easy for a user to have multiple keys associated with their account.

I thought the point of passkey security is that you don't have to send the private key around, it can stay on your device. Different passkey per device. Lose or destroy a device, delete that passkey and move on.

I don't know, but arguably the OS version is better for privacy, as each app can just trust the signal sent by the OS instead of collecting a bunch of personal/biometric data.

The easiest way to read it is "there exists a function h in O(1) such that f(x) <= g(x) + h(x)."

I think first we should teach "f in O(g)" notation, then teach the above, then observe that a special case of the above is the "abuse of notation" f(x) = O(g(x)).

But people do and it is reportedly fairly easy so the majority of people are on Bluesky's layers while all is well.

The post discusses why, when all is not well, it will be too late.