nathan_compton

People should never use non-hygienic macro systems anyway, but even if they are using a non-hygienic macro, they should always use proper hygiene. Its kind of dumb to make the whole system weirder just to avoid issues which should never happen in the first place, in my opinion.

I am a strong proponent of Lisp 1, primarily because the distinction between functions and other types of values is artificial. Functions have first class semantics in Lisp 1 and Lisp 2, but Lisp 2 makes you denote them differently but in an inconsistent manner.

Lisp 2 advocates typically make a few arguments. One is that having a separate namespace for functions makes it clearer when you are using a function vs another value. The second is that the evaluator has less work to do when examining the head of a list - it needs only look in the function environment, not the full environment.

On the first subject I must disagree - you can bind a function to a regular variable and then use that variable everywhere (except in the car of a list representing a function call), so for most positions in a set of expressions you don't really get information about whether the object being denoted is a function or not.

I suppose the second point is somewhat valid, though I suspect if you benchmarked interpreters and compilers it would barely matter. As a person who favors functional programming with a lot of combinators, I find Lisp 2 introduces a lot of pointless noise in the syntax for no reason. And I fundamentally just don't see functions as significantly different sorts of values, so I find the syntactic distinction bizarre.

God, who the fuck are they even writing this slop for? Other machines?

Neat. Maybe even deeply interesting. Absolutely garbage write up.

Typically surveys are adjusted for sampling biases before reporting. That appears to be the case here. So there is usually some attempt to account for the biases in the sampled population.

The impulse to ask "what population was sampled?" is good but its not always a straight line from there to "these results directly reflect that sampling bias."

In fact, from the page you posted: "Data for the general U.S. population (including the High Net Worth oversample) were weighted to Census targets for education, age, gender, race/ethnicity, region and household income. A full methodology is available."

I would presume that the headline number attempts to account for sampling bias.

I know there are different contexts, but a tensor is not a collection of numbers, in a mathematical sense. A vector is not a list of numbers. Such collections of numbers are representations of objects with very specific kinds of properties under coordinate transformations.

I think it genuinely damages people's ability to digest the mathematics to tell them first and foremost that these objects are collections of numbers.

Run your agents in podman containers.

Well, sure, but if you want to accept those things you have to revoke the idea that distances and times are observer independent quantities. Even physicists fail to understand this after years of studying relativity. It is very difficult to really understand that relativity says: distances and durations that you measure in your frame of reference are not physically meaningful quantities.

Which are still aggressively trained to reproduce the orthodoxy. They have to, to be viable products, since most people want to know what the orthodoxy is when they pose a question to an LLM and because not even experts can consistently agree on what elements of the fringe are genuinely useful to consider and which are bullshit, so that doesn't get into the training data. This will get even more pronounced in later models, for which the training data is much more curated.

I presume you are an expert in some field. Think carefully about the boundary of the field and all the subtlety and complexity of that boundary and all the oversimplification you do to communicate that stuff to lay people. AI is, in some large sense, directed at all lay people, not experts, and even if we wanted to direct it at experts, at the edges of knowledge, there really isn't a lot of training data for that. Mathematics is a sort of exception because it has very clear validation criteria which makes RF particularly easy for it.

Einstein's discoveries were based (to a large degree) on negating very specific parts of scientific orthodoxy and then taking the steps forward to carefully derive results with those rejections in place.

LLMs are aggressively trained to reproduce facts and consequently struggle to reject orthodoxy. There isn't any reason they can't, in principal, make big new discoveries just by getting lucky, which is sort of also how humans do it, but its ok to acknowledge that current AIs aren't so good at certain things.

Maybe its late enough to say maybe we don't need to be devoting half the worlds capital to building data centers.

"Science advances one funeral at a time"

Well, these AI are never going to die in any real sense, so expect them to make orthodoxy more sticky, not less.

100k really isn't that much money in the grand scheme of things, especially because they will probably get caught anyway.

I've got a PhD in Physics (undergraduate in Math, fair amount of self study) and I have to caution against taking what the AI's say seriously.

This is definitely a why not both sort of situation. In fact, I think being dimwitted is often associated with being evil, since goodness is (often) just rational self-interest.

Genuinely, from the bottom of my heart, thank you for writing without an AI.

" A single document produces more then one embedding, depending on the complexity of the document it can produce hundreds or thousands of vectors."

That typo up there is kind of endearing in the AI slop era.

I think this is really the wrong framing. Science can't disprove religion. The question is whether it makes any sense to believe in the religion in the first place, given what observation and science say about the universe. Science can't prove that God didn't create the universe a nano-second ago in exactly the state to produce this temporal evolution, but no one believes that because its not explanatory and also fails a bunch of other, not-necessarily scientific, but rationally motivated epistemological tests.

The way I see this is that science cannot disprove any particular religion, but it can probably offer more compelling explanations for the state of the world than religion can offer. People haven't flocked away from religion because explanations for the state of the world aren't really what people want from religion. They want a sop for their anxieties. they want community, etc. I think believing in nonsense is a real shitty way to get these things, but I'm not most people.

This book is really fascinating because it contains a surprising amount of Soviet ideology. The authors repeatedly state that mathematics is posterior to the material world, not prior to it. That is, mathematics is just the observation of regularity in the world, particularly those discovered by people working to create things. Contrast this with the still heavily idealistic world of western mathematics, where mathematicians are more likely to sympathize with the notion that numbers are real things somewhere out there whose structures the real world supervenes upon in some way.

Interesting stuff!

Even though I favor the Soviet view of mathematics personally (I do not think numbers "exist" out there independent of the material world), I think this approach hampers the didactic goals of the text and probably hurt Soviet mathematics as well. The examples in the text are all highly concrete (literally things like rubber mats when discussing curvature). This very down to earth style makes the abstract notions of curvature in other contexts (for example, general relativity) more difficult to grasp, in my opinion.

On the other hand, some people prefer strong, material, examples of mathematical ideas. This book definitely provides that. The section on affine maps in terms of fixing the plane of a surveillance airplane photograph is beautifully concrete.

There are plenty of high intensity activities with low risk of injury. Rowing and swimming come to mind.

I think the bigger problem is that, as far as I can tell, very few people have the appropriate personality type for high intensity exercise. Most people seem to experience it just as pointless discomfort.

Do you resent it when babies are born?

Delegation does not have to be because of a lack of knowledge. In fact, it seems like if one delegates for this reason its probably a sign of trouble to come. We delegate because of lack of time.

I guess its impossible for an executive to know ALL the details of the work they delegate, but I'd be willing to wager that executives who understand the details function better in the long run.

It certainly isn't tautological that executives be imbeciles about the businesses they run.

I think this is a rosy estimate. The vast majority of what people do with these models is just the same old shit, I would be surprised if 1% of it were genuinely novel stuff worth folding back into the training data.

I'm pretty sure I didn't provide an address or an id when I got my library card.

Yes, it will show up on your bill as a cost drastically smaller than purchasing the equivalent amount of gas. If you put solar panels on your house you can get that cost even lower.

One under-appreciated value of having an EV is that you don't have to buy gas. You literally do not have to buy gas. I cannot emphasize this enough: you do not buy gasoline for these cars. Not only that, but many places let you charge them for free. That is like someone giving you free gas.

I think there is a difference here. Signal qua signal isn't dangerous. Its purely what one does with it that might constitute a criminal act. In particular, using signal is unlikely to cause problems accidentally.

Driving a giant truck for no good reason entails a certain risk because of both the inevitability of human error (on the part of the driver but also others) and basic physics.

Furthermore, one might argue that the right to private communication is more important than the marginal improvement in life that you get from driving a big car for no particular reason.

Like consider a person who wants to drive a car filled with TNT. By its very nature this is a danger to others regardless of the intent of the driver or other drivers. Society might be argued to have an interest in regulating such behavior. I think there is a good case to be made that unnecessarily large vehicles differ in degree only from the TNT case, but differ categorically from using signal.

I'm not making the case that the differences justify banning big cars, but there are differences between the two situations.

I've felt for a long time that laws should have an attached "intent" section and one function of the courts should be to invalidate laws which, after a period of being in place and some analysis, fail to achieve their stated intent.

I think it would be good for democracy if lawmakers had to put in writing what the law is supposed to accomplish AND if that were something legally binding.

"The uptick in rental delinquency isn’t new. It started six years ago, when the pandemic flung the city’s economy into chaos and plunged low-income New Yorkers into dire financial straits. But even as the city has rebounded, rent collection rates in affordable housing remain short of pre-pandemic levels. As costs balloon, landlords say insufficient rental income is threatening their ability to stay afloat."

It started 6 years ago, before he was mayor.

Are the smartest researchers in the world out there saying there isn't a wall? I don't know of any people doing the actual R&D who frequently make outrageous claims.

Synthesizing a bunch of stuff I've read here lately, it seems like if OpenAI and Claude have actually found product market fit (generating code) then the question of hallucination is going to get less attention in the future. If the real money is in code generation (where there is a relatively clear acceptance criteria of at least "it runs and does what I wanted as far as I can tell") then there doesn't seem to be a lot of juice in pulling ones hair out on hallucination of facts.

It seems like for agentic coding, just making sure the AI can find the relevant documentation to establish a ground truth is probably sufficient.

Note that I'm distinguishing here between hallucination of what you might call "free facts" and hallucination of material which deviates from what is in the context itself. The latter seems both a tractable problem and one which will improve coding agent functionality. But the former seems like its no longer on the critical path, probably because its hard.