SideQuark

Put all of these trends together, and I think that in 10-15 years, we are going to have consumer PCs (and phones!) running models doing pretty much anything that frontier models can do right now.

Not likely. The last 50 years had Moore’s law growth in compute. That’s over. Frontier models are roughly compressed all written text and a large part of images. Those don’t compress forever, and likely not a ton more than now.

Inference requires touching a significant of that per token.

All of these are up against fundamental limits, more or less.

That's not recursion. And reversing arrows on this category turns base cases into co-base, or start cases, (and these are still often called base cases in the literature), so it is still required.

Reversing arrows again, back to recursion, you have exactly the standard labeled bases case (well, cases in your code), and reversing arrows doesn't magically change algorithms or invent new structure, so it completely equivalent in the category.

For that code, the 1:1:... is exactly the terminating/starting point, reversing arrows in the category does nothing to change the requirements.

… which needs a terminating case, which cannot be not defined as the same recursion.

What is it doing to feed my family in burundi?

That it has not solved everyone's problems on the planet is no argument it against what it has provided more than any other system.

It's surprising how many of your comments place you in the US, and not one in Burundi. I'm guessing based on your past comments you're not too consistent on your claims.

Nobody said this, asshole

You claimed they provide little value against obvious reality. We're done.

It's because the underlying optimizer follows patterns in LLVM, which is used for a lot of languages, most applicable here is C/C++, and bit things like this are used a lot for performance.

The particular patterns were added to C2 just recently, being copied from the implementations in LLVM and GCC.

Also it's unlikely the compiler looks for this in particular, but this falls out of a set of optimizations that do matter and it collapses into the output assembly on relevant architectures.

Yeah, it only pays pensions, retirement accounts, provides liquidity for new industries, which then grow and provide jobs, and has been demonstrably one of the best inventions for pulling billions out of poverty and increasing standard of living generation after generation for hundreds of years over hundreds of countries on the planet.

We should outlaw all markets, right?

And then you summarized

4% of ~40B times 12 years = $19B.

as if this might be close to reasonable estimate of the profit off these farm repair contracts.

But you pulled that reasoning right out of nowhere. It's simply ignorant.

Multiplying 4% times revenue? Why pick revenue? Why not profit before taxes? Why not profit after taxes? Why not any of a ton of more relevant variables?

Because you have a belief in search of some numbers to make it sound smart.

Their net profit is around 10% of revenue, not at all out of line for companies in this sector. If you're going to magically claim that 40% of this profit is from farm contracts, that is massively ignorant of what they sell.

Since you seem unable to look at their financials, there's no need for me to show you how little of their overall revenue is farm equipment (which if course makes your number voodoo even more out of touch).

… so there's a physical process that's undecidable.

No, since you cannot physically build a Turing Machine. A Turing machine requires infinite tape. Any physically realizable machine doesn’t have that, so has finite states, so is decidable: enumerate the states in finite time - it halts or repeats, so all programs on a finite state machine are decidable.

Your example is not an undecudable physical process.

Godel things also don’t apply: Godel theorems are about proof of this or that from within the same system. In logic one can prove such things from an outside system, then construct towers, avoiding Godel theorems. Godel theorems also require a model of integers including multiplication (without multiplication, such systems were proven complete and decidable). However the universe does not contain a model of integers, as the physical universe is not unbounded: relativity places a finite limit in spacetime on what can interact.

Mixing math as reality fails at these requirements.

Go pull their financials, stop pulling nonsense out of thin air.

I looked over their financials. You conjectured out of feelings and ignorance. Calling my evidence naive, when you are simply making things up, is ludicrous.

Go do some legwork, get some actual data, not make believe feel good fairy dust.

The fed suit had exact numbers, which they blacked out for public. The fed suit settled for a dollar amount. If they thought they’d get a better amount, they’d go to trial. So likely the proceeds are on order of the settlement, else they’d likely win more by trial. The feds are not idiots.

But some people are.

Last year was 5b net profit on 44b revenue. Attributing more than a tiny fraction of profit to the right to repair stuff is wild dreams, given the amount of physical goods they sell.

Nothing in their SEC filings shows anything mentionable about such claims. It does break out actual profit by company sectors.

Ah, trying to move the goalposts huh?

What’s the number of Teslas sold per year? Declining year over year as others have grown significantly. Profit margin sans regulatory issues? Peaked 2021, declining since then. Same in about every company wide metric.

BYD? Most all such metrics over the same period doubled to quadrupled. They outsell Tesla. They have better tech than Tesla. They’re clearly out innovating and outselling Tesla.

And, as the other poster made clear, your moving-the-goalposts let’s-ignore-the-rest claim doesn’t even stand slight scrutiny. Try harder indeed.

This feels like the “we should make an internet company that ships pet food to anyone, anywhere!” of the 21st century.

You’re claiming in 10 years they made something others won’t for another century? This has no historical precedent.

China is going to eat them, just like they ate Teslas insurmountable lead in a few short years.

Rent vs own is about price. If you want to own something, digital, physical, information, goods, then a vendor will charge more, sometimes a lot more, to sell you an infinite ownership vs you get limited rights.

Most everything I buy needs to fight for my money. Of course some vendor specific things, such as a specific movie, are not replaceable, but since very few things, if any, are so important to me that I won't spend elsewhere if a good is too much.

With that competition for $, vendors generally are competitive, and don't tend to have extremely high profit margins.

I prefer the limited rights, since I can try far more things that way, with finite budget. And I don't get stuck with higher prices on things I ended up not wanting/using after I got it. I can simply cancel.

I think the general disconnect is people too often assume they handed out money so should get infinite rights, while ignoring the nuances of economics/pricing.

If it's on the stack, it's almost free to do.

If it's held by a unique_ptr, then it auto frees, but there is a lot of programmer work to ensure things behave well, and there are nuances to allocating something into the unique_ptr.

If it's a std::shared_ptr, which is the closest to what people think of, it's then a reference counted object, which requires a little more storage, and also has the ability to be mishandled as a pointer, screwing it up, or making the wrong type of copy of the shared_ptr, so two there are two things, and also fails to release object in the case of circluar references.

Then there's a lot of other subtle things you can do, and none of this is as lightweight as a pointer, and none is as automatic as what most people think of as garbage collection.

And each has costs and tradeoffs, which is why C++ added them: making certain tradeoffs easier to use, but it requires solid understanding of how such things work.

It's interesting you cut out the "There’s plenty of papers showing exactly this." then pretended the rest is simply ignorant belief.

the businesses that won during the 2000-24 period (the so-called “tech” companies) were companies for which capex were well below average.

First, it's odd that when discussing the point about total US worker productivity and capital expenditures, you somehow think a sector consisting of less than 1% of workers is evidence. And even on that claim you are incorrect.

Here's the datasets for that period capex by industry [1]. You are demonstrably wrong by a large margin. I expect you to have some better sourced or more convincing data, otherwise it seems you're operating on belief, not evidence.

If you want to try again, try looking at all workers, not cherry picked examples.

That's not “well known”, you're just saying the gospel

Here's google scholar on the literature around this [2]. Go ahead and tell me again how this is not well known? Pretty much every paper and piece of empirical evidence on this points to the same reasons.

[1] https://pages.stern.nyu.edu/~adamodar/New_Home_Page/dataarch...

[2] https://scholar.google.com/scholar?hl=en&as_sdt=0,15&q=us+pr...

So you replace a loop that delivers last mile goods to a lot of cars going decently out of their way to the limited places just to pick one item? Surely it’s less miles driven when Amazon does a loop hitting several people right near me than each of us driving farther in total to get our goods.

Yes, there is always an issue upsampling since you have to make up new values. For downsampling by an integral factor, you have honest samples at those times, but any other rate will also need new samples to be constructed. So for these, in either direction, there will be the need to make things up.

No matter the process, whether naive fourier transform based, or any possible algorithm that is smarter and tries more things, it cannot make up the true values in all cases, since that data is simply not in the signal.

Ancient methods treated sound as a sum of band limited pure sinewaves, which it is not, and you can claim you reconstructed up to that model, but it's pretty bad, and fails even theoretically for the reasons I posted elsewhere.

So you then start doing all sorts of signal processing, but none can actually reproduce all sounds, and most (all?) methods make tradeoffs - what error do you accept for what gains?

And if you optimize a system for one rate, and don't do it to cover all other sample rates you want to handle well, then you will make some rates worse and some better.

Well designed signal processing has to be careful to handle all the ranges it cares about.

And no one who knows or has worked in these areas would simply drop in a new system that is untested for major uses for such a widespread tool as ffmpeg. Precisely because they understand the nuances in these things.

I take it you don't know much about frequency time tradeoffs.

It is not possible to convert 48 khz to 44 khz without approximation. And there is incredible room to design filters that have great response at frequency X but terrible at a nearby frequency, since most of filter design is about where to stash the noise. That noise can add perceptible beats to audio.

And don't try some silly Nyquist limit stuff - none of that is actually correct in actual filtering. The initial simple reasons are that that stuff assumes infinitely precise, mathematically perfect sample values - impossible since these are PCM. They assume infinitely periodic inputs, also false (and there's an entire industry of trying to find window functions to smooth lack of periodicity, also with flaws and tradeoffs). For perfect reconstruction, they use sinc, which has infinite support, which is not computable. So people think they'll use some windowed sinc, which again then fails to do reconstruction fully.

After these ancient, and in practice lacking, ideas, modern filter theory and practice has invented generalized sampling, which has provably (from math and in practice) better signal reconstruction. And even that is only the tip of an iceberg of knowledge.

These codes don't simply "work in the frequency" domain as if this is old school DFT stuff. Even in that domain, your claims don't work. These codecs also do psycho-acoustic modeling, to be able to decide which data to drop, alter, increase, or decrease, to get following bitratres to encode at target. Again, these type of filters and models can be terrible at 44, but awesome at 48.

So yes, they do need to test before putting this into the world's most used by far audio/video transcoder.

So yes, is is cynical, apparently driven by ignorance.

It’s not that cynical. The author didn’t test on the most common rate in use, so it would be ludicrous for any serious project to wholesale replace a decades old working pipeline with it. It makes perfect sense to wait till due diligence is done.

There’s plenty of papers showing exactly this. What do you think has driven productivity? People simply bring smarter?

The fact is capital expenditure from company or investors has bought machinery, compute, pipelines, transport, and massive investment to make those workers more productive for decades. As such, the returns to capital as a share has increased. Those places able to deploy capital to add productivity win over those that don’t.

And real total remuneration across all quintiles has increased significantly. BLS among others has all historical data to check.

If/when there’s a period where there isn’t more gains to be had by more investment per worker, and workers become more productive via their own skill (education, diet, genetic implants,…), then more returns will flow that direction.

This is all well known, and easily checked.

True. From that page RAM companies have been sued and found innocent before:

“ The district court ruled in favor of Samsung, Hynix, and Micron and dismissed the lawsuit. This dismissal was affirmed on appeal by the United States Court of Appeals for the Ninth Circuit, which ruled in March 2022 that the plaintiffs did not offer sufficient plausible evidence for their allegations to make a case under the Sherman Antitrust Act and that the district court properly dismissed the lawsuit.”

RAII has nothing to do with garbage collection. It only provides a way to get a resource somewhat atomically. If the creator doesn’t delete the object, or if the object has a malformed destructor, it fails. If the creator calls delete twice, it fails. There’s really no garbage collection at all here.

where the programmer is responsible for invoking "free", was a serious mistake and it was an obsolete technique already at the date of its introduction.

It wasn’t obsolete then nor now. Garbage collected languages, to this day, still use on average about 2x-5x the working memory of carefully written manual memory programs. A significant amount of devices then and even today cannot support such sloppy use of resources.

That's not a reason. That's an anecdote, a post selected and cherry picked example. Is this truly how you'd allocate resources?

A more correct way to use reason is to look at all possible things that have been proposed that cost $500m, and see how those turn out in general. Then possibly weight using expert knowledge about this particular problem and the possible benefits. Then consider what the opportunity cost is, i.e., what you give up by spending $500m here instead of something else.

Trite armchair rebuttals are not convincing.

It’s 500m spend without guaranteed success.

The Englert paper did not have the equations for the the hypothetical field that would become known as the Higgs field, and it did not have the Higgs boson. The Higgs paper had the details to make these more aptly named for him.

Englert (and his coauthor) had some of the mechanisms, but not enough details to truly have the field or boson, hence the phrase "Brout-Engelert-Higgs-Guralnik-Hagen-Kibble mechanism".

The photo is subtitled “ Representative image of a vintage notebook, not Marie Curie's. Photo via Pexels.” so it’s clear it’s not trying to be deceptive. It’s entirely common, even among the best human authors, to put in a photo to give some idea of a topic when actual photos are hard to obtain.

Because it costs a massive amount to get standards with this technical quality. They go through meetings, have a decent amount of staff to run, organize, have conference costs (locations...), take years to get done.

Someone has to pay for this. Making companies (and often, many of the individual members do this out of their own pocket) pay it all means worse standards, as some people stop going.

Sharing the cost to make the standard makes it a better mix of getting good standards and having low costs for final users.