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.
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.
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.
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.
* 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.
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.
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?
Thanks for making sure I’m secure enough to read your three-paragraph news article…