atq2119

The straightforward solution to this is to allow the tax to be paid in kind, to be paid into a sovereign wealth fund.

Owe 1M$ in taxes and have ownership stakes in a company worth 1B$? Fine, transfer a 0.1% share of ownership in the company into the SWF.

Of course that is a feature, but at this point I have to suspect that you're willfully ignoring the point.

The moral problems begin when writing that book gives you extraordinary power beyond what is healthy for society (which is extremely rare for authors, of course; the discussion isn't really about authors, you just invoked them in an attempt to conjure a moral shield for the people who are the real problem).

And again, even that in itself is still not a moral failing of the individual. It's primarily a failure of the system in which the individual operates.

It does become a moral failing of the individual if the individual uses that power to perpetuate the system.

Yes, the same is absolutely true for many other professions. And artists are probably more aware than most, at least on average, how much luck plays a role.

But note that I have been very careful not to call the fund managers individually immoral.

Yes, you got it!

It's because while for some people there may be jealousy involved, many more have realized or are now realizing how much power the ultra rich have, and how bad that is for our societies.

Once a company becomes large enough, it becomes so influential in society that control of it must be made more democratic.

I don't think anywhere has figured out quite the right way to go about it yet, but it's clearly the right goal. Some countries require employee representation on the boards of large companies.

Part of it is work, yes. But consider. How much is your take home as a fund manager with 100M AUM? How much is your take home with 10B AUM? The work is the same, yet if the take home is different, you've proven that your income is not in fact earned through work in a moral sense.

The sad part about it, and one that has become a bit of a theme with his postings, is that pg stopped being intellectually honest in his online writings at some point over the last two decades.

His post here in particular violates the fundamental principles of HN in that he does not engage with the argument at all.

The argument isn't that it's impossible to become a billionaire legally, the argument is that it's impossible to become a billionaire in a moral way, though that's more of a problem of the system than it is necessarily one at the individual level. A just and moral system would assign the value being created in such a way that becoming a billionaire would be essentially impossible.

Yet pg never even acknowledged the possibility that that might have been the argument.

Agreed. To put this in perspective, batch 1 token decode is bandwidth limited in theory.

Memory bandwidth of RTX 3090 is listed as 936GB/s. The post isn't fully clear on which model they used and how big it is, but even assuming it perfectly filled the 24GB of that GPU, 30tok/s means the achieved bandwidth is only 720GB/s. There's a bunch of room for improvement here even without MTP, and those improvements should largely stack with MTP.

Interesting to see somebody argue out in the wild for what I have been subjected to in the past and have long considered absolute worst practice.

It may be that my perspective is different because my work tends to be in "hard" foundational software (think OS components and programming language tool chains). It's not that customer requirements aren't a thing at all in that kind of work, but they tend to be far removed from the day-to-day and instead, mechanical sympathy rules supreme.

Commit messages need to focus on the software, not on the trappings of the process by which it evolves. Links to tickets can provide helpful context especially for bug fixes, but they belong in the commit message footer.

Your last paragraph is absolutely an anti-pattern at least in this work. If the implementation of a new feature is split over multiple changes, then surely there is something different and important to say about each of those changes. Does your split even make sense otherwise?

That is definitely not my experience using Claude Code with Opus. I work in a very sparsely commented code base, and the agent produces substantially more comments than the surrounding code.

Of course it does, but that's beside the point.

As a software developer, you must never subject your team mates to a PR that you yourself believe to be low quality. The point of code review by others is to catch things that you missed.

There are multiple lines of defense for quality. Yes, automation can and should be one of them, but your own self-review always has to come before review by your team mates.

That quote shows an utter disregard for basic human decency.

It is the responsibility of the person running the coding agent to make sure the resulting PRs are high quality. Putting that on your team mates, or worse, random open source project maintainers on the internet, is the definition of an extractive contribution.

You may be confusing some abstract unachievable ideal with the reality of the world we live in.

In reality, being superwealthy absolutely comes with a tremendous amount of power.

It is also pretty much the opposite of anarchic, given effects like regulatory capture and politicians pandering to the desires of the wealthy.

This is a genuine problem indeed and part of the appeal of an UBI. The idea being that if the rules of redistribution are dead simple, then that helps minimize the potential for grift, which in turn minimizes the potential danger of a redistributor class.

That said, it is fundamentally important that nobody has too much power, and that power changes hands on a regular basis.

At a global scale, this necessitates taking power away from the capitalist class.

Ideally that power just doesn't go to anybody, but to the extent that it has to go somewhere, it almost doesn't matter where. Or perhaps it's better to say that there are many options that are acceptable and better than allowing power to continue to accumulate unchecked.

Unless you're systematically repeating the exact same task, the most parsimonious explanation is that you're seeing natural variation based on different tasks, random sampling of tokens, etc.

It's been a while since I saw a detailed paper on a high end training run, but extrapolating from what I remember, it seems those training runs are in the 10s of trillions of tokens. This already accounts for potentially sampling tokens multiple times during the training run.

That seems like a large number, until you realize that OpenAI claims to have almost a billion weekly users. And OpenRouter shows many models at over a trillion tokens per week.

So in pure token terms, I'd say it is in fact extremely plausible that inference dominates, at least for the popular models.

If the inputs are constant, then all the multiplies are constant and the only thing that toggles is the accumulation. Which explains the pi situation.

Normal vs uniform is less clear, but also not as much of a difference. The arguments about signs isn't just about a signs bit, though. The way you negate during accumulation is that you flip all the bits. Only the final float representation is sign+magnitude, the accumulation itself has two's complement steps. I don't actually know the analysis here, just pointing out that it's not that simple.

Think about it from the other end. Why would any bits flip at all in the data path of your matrix multiplier when all the matrices are 0?

You may be talking past each other. You're talking mostly about quality of life. That is different from being (de)humanized.

Dehumanizing is treating other humans as not, or not fully, human. For example, the term "human resources" is dehumanizing because it puts humans on the same level as other resources. If you're treating humans like you'd treat, I don't know, lithium or the ocean, you're dehumanizing them.

The more humans are treated as numbers on spreadsheets and other forms of computation, the more humans are dehumanized.

So both can be true: we're more dehumanized than in 1900, but while that does impact quality of life negatively, the overall quality of life may still be better than back then.

The question should be whether and how we can have both: overall quality of life without being dehumanized.

And, since there's a pipeline from students and hobbyists to professional use, it's risking the future.

That sounds more like the 80286/80386.

Sure, that's the main challenge with building good sandboxing systems. But it's not actually that hard to do when the will to do it is there.

For example, Android already allows you to give apps restricted access to your media. My understanding of the way it works is that the resulting interface for picking photos etc. is not under the control of the app. The app only receives whatever file you picked.

Internet access. An editor extension does not need it.

Fully agree with the first half of your comment. The second half goes off the rails, though.

I rarely see people complain about sandboxing.

What people complain about is when devices are locked down in a way where you are only allowed to install software that is approved by a central gatekeeper, even though sandboxing is in place that should make it far safer to run arbitrary safer than on traditional desktop systems.

I don't think this is factually accurate. What it really boils down is a question of scale of societies.

Most of us humans inherently value each other. There are exceptions, and small communities can get nasty. But for the most part, small human communities tend to be supportive and valuing each other.

This really only stops being the case when you get large-scale societies that allow humans to view others through an overly abstract lens. Combine that with an unchecked accumulation of power, and you have the potential for those in power to view the rest as without value.

It really depends on how often you use git bisect and blame. This varies greatly across projects.

That said, if/when stacked PRs become a first-class citizen in GitHub, more projects will see the benefit of this approach (though they'll probably mostly get there through squash-merges).

Thank you. I vaguely remember being put off by Mercurial at the time and feeling much more comfortable with git, but couldn't have put it into words as well as you did.

People always say you have to know git internals in order to use it, but that's just not true. Git has the right data model and has always been about empowering users to edit their data. Which makes the data model be "in your face" compared to the alternatives (and I think that's what people latch on to when they talk about "internals"), but it ultimately makes for a better tool.

I agree that training is a forever task, and the current rate of training is probably not sustainable. But all that means is that once the current investment mania ends, the market will most likely find a new equilibrium where continuous training still happens, but at a slower rate that can be sustained by inference revenue.

> Without continuous training there is no inference product.

This claim deserves teasing apart.

Clearly, training is a Red Queen's race today. If a model provider were to unilaterally decide to stop training, they would very quickly lose market share to competitors with better models.

On the other hand, what if market and investment conditions change such that everybody has to stop training?

In that case, the models are still there and still as useful as they were the day before. So why wouldn't there still be an inference product?

Let's not get ahead of ourselves. Millions, really? I can believe there are a lot of enthusiasts doing this, but "millions" needs a citation.

Inference at per-token pricing is profitable.

The article's point is that if you're relying on flat fee subscriptions, a rude awakening may be coming. That seems plausible to me. Issues around token quotas are a frequent topic on HN.