GeneralMayhem

Accurately placing heavy, aerodynamic objects onto people when you start out directly above them is not very difficult. The hard parts are either placing the object on top of the person from a few hundred or thousand miles away, or - in this case - placing an object that tends to flutter rather than follow a ballistic trajectory.

Two big advantages.

First, if you want a declarative config with limited, domain-specific options, rolling your own DSL instead of using something as complex as Python is much, much easier to implement. You're not actually going to be running the code either way, at least not in the normal way, and the Python syntax tree is pretty complicated.

Second, having code that looks like Python can lead your users to believe that it is in, in fact, Python. When you're doing things like using your DSL as configuration that happens at setup time, but then actually "running" the resulting config later on, that can lead to people getting themselves into trouble - for instance, they might try to use `time.now()` and end up embedding the time of the config parser as a constant in their workflow definition.

If you want to use Python as your language, you probably want to define your "DSL" as a Python library, so that you can use a normal interpreter to work with it. Maybe you have library functions that return config objects, and a user's "configuration" is an arbitrary Python file with a standard function name as an entry point. But then when you want to introspect over types, you probably need to start playing games with decorators, which is tricky again, and you have to be very careful to have that evaluation step return meaningful errors.

Starlark (https://github.com/bazelbuild/starlark) is an example of using Python-ish as a "configuration" language. That took an absolutely massive amount of engineering to get to be well-defined, and was only arguably worth it because they wanted a language that's a loop construct away from being Turing-complete. If they had wanted a basic declarative relationship language, they probably would have used textprotos or GCL.

Past research ([1]) has shown that Twitter disproportionately boosts right-wing political content. Qualitatively, you can find whatever you want there, but Elon Musk's personal connection with and admiration for a number of right-wing individuals makes them stand out. In particular, Twitter under Musk has a clear pattern of not applying its moderation policies to high-profile accounts; in the current year, it is right-wing politicians who most frequently post hateful or blatantly false information, so failing to hold them to a normal standard amounts to a tacit approval of what they're posting.

[1] https://www.pnas.org/doi/full/10.1073/pnas.2025334119

The issue with normalization is that you lose a degree of freedom - which when you're visualizing, effectively means losing a dimension. Normalized 2d vectors are really just 1d vectors; if you want to show a 2d relationship, now you have to use 3d vectors (so that you have 2 degrees of freedom again).

Ideally integers, but at a large multiplier.

Google's money proto [1] has units and nanos. Any competent ad-tech system will use something similar: integer number of micro-dollars, nano-dollars, etc. You want a fair amount of precision, so just tracking whole cents isn't enough, but you want that precision to be (linearly) equally distributed across the value space so that you can make intuitive guarantees about how much error can accumulate.

[1] https://github.com/googleapis/googleapis/blob/master/google/...

Yes, the assumption, at least in the United States, is cannabis - unless you're in a social circle that routinely and exclusively uses a different drug. But the assumption is 100% cannabis when it's an article posted on April 20.

That is abjectly false. Synthetic pesticides have been used since before the second world war. Silent Spring was written in 1962 about the effects of the amounts of DDT that had already accumulated in the environment by that point.

I use https://github.com/spf13/cobra religiously for this kind of thing - it handles all the annoying corner cases of parsing flags, and also has an intuitive notion of subcommands (with basic usage/help text generated) for picking which task you want to run with positional arguments.

Mean GDP per capita, real or nominal, is a useless metric for almost all purposes. When we're talking about average people's lives, it's worse than useless.

This is true if you require the same amount of income to maintain your lifestyle, but if you've been working in tech for 20 years, you have savings. I'm not in my 40s yet, but it's much more viable for me to take a risk now than it was when I was just starting out because I can afford to fail for a couple years without, you know, becoming homeless.

Animals don't get a vote, but just as the government can forcibly set aside land as parks, the government can subsidize individuals to do so where it would be valuable to the common good.

Only if implemented naively.

The idea of LVT is to put an economic value (i.e. dollar amount) on the fact that you're preventing people from using land. It's a market correction. If you want to express that some values are actually good beyond what the market is equipped to reward, then you need to correct further. In theory, this means that the government - acting on behalf of the public good, or its citizens, or the future of humanity, or whatever your preferred vision for the social contract is - should offer its own contribution to the cost to be paid. In practice, this means there could be a significant tax credit for meaningful nature preserves, commensurate with the value of that preserve. Taken to its limit, this tax credit could mean a negative tax (i.e. subsidy) that's enough to pay for maintaining the land in its natural state, and oh look we just reinvented public parks.

At the risk of feeding a troll... no, that is not the proposal. The proposal is to make it economically unviable to do socially suboptimal things with common resources.

The underlying principle is that land - including natural resources as a whole - is finite. It's a common good owned by humanity as a whole; or, failing that, at least in common by the nation-state in which it falls. Allowing individuals to hoard those finite resources is bad for society.

Under a land-value tax system, individuals can do whatever they want, but they're appropriately charged for their negative externalities. Just like we should have a carbon tax to disincentivize people using the shared natural resource of carbon fixing without a good reason, we should have a land tax to disincentivize people using the shared natural resource of flat, arable land without a good reason. You can choose to pay your taxes and do nothing with the land, if you have another source of wealth to make up for it - effectively, you need to express to society just how much it's worth it to you to sit on that land doing nothing, and the way that we express the strength of our convictions in a market economy is with dollars.

> There's also something very objectionable about attempting to use taxation as a tool to manipulate behavior and engineer outcomes.

The goal here is not to manipulate behavior for its own sake, it's to use the power of the state to correct for a mispricing in the market. Market forces alone are incapable of correctly pricing deals involving common goods, because most of the people who are implicitly participating in the transaction (i.e., the rest of humanity that is now deprived of that land/air/resource) are not compensated. It's a problem of negative externalities, for which a Pigouvian tax [1] is a perfectly appropriate remedy. In a democracy, the state is how the public as a whole imposes its collective will; having the state "bid" on behalf of the public interest is the most capitalist solution possible to the externalities problem.

[1] https://en.wikipedia.org/wiki/Pigouvian_tax

No. The land would be taxed whether it has improvements or not, and in fact would be taxed the same amount whether improved or not. That's how such a tax incentivizes improvement - if you're paying the tax either way, you want to do as much as possible with it to get your money's worth.

Sure. 100T/day * 1day/86400sec ~= 1B/sec. They're probably considering at least a few hundred candidates per impression, and every impression is going to go through _at least_ two models (relevance and pCTR/revenue), so you could get there just with online serving at 5Mqps, which is plausible. But they're also going to be doing a lot of stuff in batch - spam predictions, ad budget forecasts, etc - so that every candidate actually runs through four or five different models, and every actual impression could do more than that.

> or in non sunny California conditions at all

It does at least rain in San Francisco, and I've taken Waymo in the rain. It did more than fine. Plus, you know, night time.

Waymo is not public transport in the sense that most people mean it, which describes only state entities, as opposed to the private sector. Public transit exists to serve the public good, doesn't need to turn a profit, and is accountable to government (ideally elected) officials. Waymo is a for-profit entity that can do more or less whatever it wants within the bounds set by CPUC. I think it's a net good, but it's still private industry.

That's only true if there's competition with similar costs, though. Uber and Lyft can't reduce prices without collapsing, so Waymo has no incentive to charge anything less than what they charge minus one cent and they'll have all the demand they can handle.

This is pretty standard survivorship bias - which is to say, a well-studied problem with a famous pithy example (bomber plane armor), that many people and organizations still somehow don't learn.

It is literally true specifically for San Francisco. The street now named Main in San Francisco didn't exist for the first few decades after the city's founding and grid-laying (it was underwater, just off shore in San Francisco Bay!), and at no point has it been a street of any particular prominence.

That map also showcases another example of this - Townsend Street is named after one James Townsend, who was the alcalde [1] in 1848, not for its location at the end of town. (Geary, Leavenworth, and Bryant are also named for pre-statehood alcaldes, but their names are less amusingly coincidental.)

[1] https://en.wikipedia.org/wiki/Alcalde

Yes, but the Main St. in San Francisco hasn't been around nearly as long as SF has. When the downtown grids were laid out, what's now Main St. was underwater. It didn't exist as land, let alone a road that was either significant or central. By the time it was filled in, it was home to Charles Main's mining provision shop.

The other way you can tell that it's not "main" is that it's parallel to 1st Street. Most cities don't have both a Main and a 1st, they have one or the other (which is why "Second Street" is the second-most common street name in the United States behind "Main", with "First" coming in third place). If SF were using Main in the "normal" sense, it would be where 1st is, next to 2nd. Failing that, it would at least be the next block over from 1st, so that 1st was counting from it. Washington, DC, sort of works that way, where there's a "zero street" (North/South Capitol), and then 1st is the next block over. But... SF's Main St. is a full three blocks from 1st. It's just kind of in the middle of nowhere, logically speaking. And that's again because of the landfill situation - 1st St. is the first full block counting from the shoreline at the time the streets were built.

I have never considered which language a candidate knows, or had my set of languages considered, in hiring. That goes double for someone being hired right out of university. I want an interview to test whether you understand how a computer works, how networking/distributed systems/graphics/memory management work (pick the relevant ones). I don't care what the exact implementation language was. I _will_ test a bit more carefully if the applicant says they only know Python, because that can be a sign of fuzzy thinking around pointer/reference/value semantics, but if they can do the work, the language isn't interesting.

The only exceptions I can think of are where you're hiring someone specifically to work on language tooling - e.g., to work on the core C++ library team at Google, you probably need some experience with specifically C++ or C and compilers thereof. Those would be very specialized senior roles, though.

White Americans just after the second world war controlled a huge fraction of the world's wealth. It's incredibly easy to raise your family in relative comfort when you're part of the only industrialized society on Earth that hasn't been reduced to rubble, the ones being rebuilt owe you billions upon billions of dollars, and you have a reliable underclass of other races to do the most unpleasant bits of labor for you.

This strikes me as not dissimilar to religious people who are terrified of atheists because, without God, they must have no reason not to become mass murderers.

There are many people with kids who have revealed the value they place on the sustainability of the environment, democratic society, or the economy through the lifetime of their children, let alone their children's children, to be zero. Why would you assume that the childless are any more selfish, as a group? Why can one not care about the continuance and betterment of humanity without one's own direct descendents being involved?

You're thinking of Cruise. Waymo has not scaled back in any way, and in fact is in the process of expanding to LA with a limited pilot through the winter.

I don't think the fact that some of the first people on the team had worked together previously makes Waymo not "created at Google". The project they worked on before, the DARPA challenge, was not a commercial product, and at the time no company was seriously investing in self-driving cars as a viable technology. This isn't like YouTube, which was a well-known brand and viable business pre-acquisition. It was Google resources that made it possible to build the rest of the Waymo team, lobby governments to allow self-driving cars on the road, work with hardware manufacturers, and leverage the rest of Google's software stack, ML expertise, street view data, and datacenter capacity to build and train the driver.

Yep, that's the one I meant - sorry, should have linked.

His series on making a GPT from scratch is also great for building intuition specifically about text-based generative AI, with an audience of software developers.

The most important thing to learn for most practical purposes is what the thing can actually do. There's a lot of fuzzy thinking around ML - "throw AI at it and it'll magically get better!" Sources like Karpathy's recent video on what LLMs actually do are good anti-hype for the lay audience, but getting good practical working knowledge that's a level deeper is tough without working through it. You don't have to memorize all the math, but it's good to get a feel for the "interface" of the components. What is it that each model technique actually does - especially at inference time, where it needs to be well-integrated with the rest of the stack?

In terms of continued relevance - "deep learning", meaning, dense neural nets trained to optimize a particular function, haven't fundamentally changed in practice in ~15 years (and much longer than that in theory), and are still way more important and broadly used than the OpenAI stuff for most purposes. Anything that involves numerical estimation (e.g., ad optimization, financial modeling) is not going to use LLMs, it's going to use a purpose-built model as part of a larger system. The interface of "put numbers in, get number[s] out" is more explainable, easier to integrate with the rest of your software stack, and more measurable. It has error bars that are understandable and occasionally even consistent. It has a controllable interface that won't suddenly decide to blurt corporate secrets or forget how to serialize JSON. And it has much, much lower latency and cost - any time you're trying to render a web page in under 100ms or run an optimization over millions of options, generative AI just isn't a practical option (and is unlikely to become one, IMO).

I don't have a significant math or theoretical ML background, but I've spent most of the last 10 years working side by side with ML experts on infra, data pipelines, and monitoring. I'm not sure I could integrate the sigmoid off the top of my head, but that's not what's important - I've done it once, enough to have some idea how the function behaves, and I know how to reason about it as a black box component.

To be fair, there is such a thing as self-driving cars - just not from Cruise. Waymo seems to be doing much better across the board.

Every commercial operating system, and the driver for many consumer components (especially video cards), receives - and are expected by consumers to receive - regular updates for a number of years, and requires constant attention to detect and address security flaws. For the average user, Windows is not all that much less of a SaaS than any web app at this point.