dchftcs

It's possible to treat software close to a first-class citizen even technically as a cost center. Probably the key is to have a lean team that is effective and management trusts, in a business whose profit margin or stability can be greatly enhanced by good software. If you're relatively lean then nobody sane would be looking to cut fat in your team. On top of that, the attitude of management and the culture is really important, and this CEO's attitude is at least a good start.

I hear Meta Ads has one of the more toxic environments for SWEs, even though Ads make money.

You can learn to take good risks and handle hardship without taking stupid unnecessary risks. One important life lesson is that the only risks worth taking are those that offer corresponding upside - else the expected outcome is ruin. Education wise this means you give them necessary or low-impact risks to take - and let them endure outcomes such as failing a difficult but important project, failing to find love, losing a basketball match, or losing friendships. Of course, sometimes you may need to do something even if there is no upside for yourself directly, but that is outside the scope of this topic.

In some cases the retail store just doesn't list the price and the store owner may just come up with a highly marked-up price when asked. 10x would be very rare though indeed.

>True for all systems, but AV updates are exempt from such policies. When there is a 0day you want those updates landing everywhere asap.

This is irrational. The risk of waiting for a few hours to test in a small environment before deploying a 0-day fix is marginal. If we assume the AV companies already spent their sweet time testing, surely most of the world can wait a few more hours on top of that.

Given this incident, it should be clear the downsides of deploying immediately at a global scale outweigh the benefits. The damage this incident caused might even be more than all the ransomware attacks combined. How long to take to do extra testing will depend on the specific organization, but I hope nobody will allow CrowdStrike trying to unilaterally impose a standard again.

There's always a queue somewhere at some level of resilience. Sometimes it's as mundane as the TCP buffer, some other times the other end points may become unresponsive for a while, including the time series DB you seem to want to take as the source of truth, or whatever that tries to resend and reconcile the requests.

I think people have a unique edge here to counter your argument because most of the work that goes into fandom wikis are from unpaid volunteers, and they have every right to move somewhere esle.

This has been happening for a while. I've heard about it before (vaguely remember before sanctions), and I can find a news article mentioning this back in 2019: https://www.theguardian.com/technology/2019/apr/19/companies... In the limited time I tried I haven't found when exactly the Mossel brand was established but it probably is a very old practice given how aggressively they pursued markets under dysfunctional governments.

Total comp needs to be market rate, not just salary. And non-preferred shares should be valued lower than preferred stock. Lumping non-preferred shares with prefereed shares is one of the bigger lies startups tell employees.

This is so oddly relatable to scaling an internet business. At least it leaves some space for niche players.

There's a physical limit as to how well one can fight, but competition and growth is just as limitless in the business world as in anime. The top people in early Dragonball make six figures, late Dragonball if you don't have 100 billion you don't qualify for screentime.

230 may have helped Google grow big, but it also helps small operators from being liable for something they can't have resources to police. The fallacy in your argument is this: https://www.dhmo.org/

This is really twisting it, I don't find pairwise indepedence to be more advanced than applying a Chernoff bound. In this problem it'd mostly be the difference of using a Cherbyshev type bound or Chernoff bound.

Pairwise independence is to give the algorithm stronger guarantees and let it work with a simple hash function like ax+b, otherwise probably most existing algorithms can be simplified in the same way. The most similar algorithm is BJKST, which is almost identical except for specifyimg the sampling mechanism to require less randomness.

To someone who worked on this type of thing before, it's like seeing people familar with LLMs say "oh yet another X-billion parameter model on github doing more or less the same".

They manufactured things people wanted, but shouldn't have wanted...

Equity holders are probably forever, and the incentive problems are a hard thing to solve. But you won't be a founder or early joiner if you're not allowed to keep your equity anyway.

Fortunately it doesn't take that much to get top talent, because so many other companies underpay. Jane Street only has to pay out a small fraction of their PnL and doesn't even need to have a non-compete.

It probably includes the founding partners, so it might be that ~0.15% hold more than half.

To be fair, they are not mutually exclusive. Businesses are incentivized to fight penalties as long as they think the legal costs are small enough compared to the fines themselves, regardless whether the activity their fine on was still profitable after the fine.

In a relatively efficient market, you can reason about the average M&A profit. The typical PE/VC fund has 15% annual return (which is not really compounded) and a 5-10 year horizon. So let's say the average profit over 5 years is somewhere around 100% on the invested capital, or around 50% overhead.

But before you complain about that, R&D paid to universities also include overhead for the universities, and it is quite significant. In US it is over 50%. The universities partly act like investors in this way, the administration invest money (e.g. start-up grants, some professor salary) into research groups and expect them to bring external grants, and the overhead partially pays for that.

Depending on what you want to measure, you need to include the cost of people and resources involved in the production of the thing you want, and not just the marginal unit cost of the thing. It may not be easy to draw the line, but you have roughly two choices:

1. Deduct a reasonable percentage from both university grants and M&A

2. Accept that without profit, people would not have funded certain research (which carries risks and requires selection), so you include those as costs of doing business in your assessment

Another way to think about it is that the money you spend on research itself containts other people's profits too. If you buy equipment, the equipment company makes a profit. Staff has savings. The govt always takes a cut. It's profit all the way down.

I think the difference in cooking oil is in how the oil is used. What I hear abour Chinese households in US is that their kitchens tend to become more greasy, because their cooking style tends to produce more oil fumes, and fume extractors tend to be less common in US than in their home countries.

The subcontracting situation reminds me of a news article about someone in China hiring a hitman, the hitman then subcontracted to another hitman, who then subcontracted to another, then another, then another. Each hitman was offered half the money of the upstream. The last hitman defected and informed the target.

What you're trying to eliminate is just a small part of the infra arms race. But it's not even effective at that. You say one of the goals is to not have to squeeze many racks into a building, but your proposal assumes the servers placed in the racks are running at low capacity, which is completely not true. On the other hand, if you trade say GOOG, the matching engine for that is going to mainly sit in one place, how do you stuff everyone's program to run on the same server, and how do you make sure they each have fair access?

You mean enabling arbitrary code execution from a third party when you can lose billions of people's money in half a second? Also if two people want to make the same trade, who gets it?

Exchanges do provide very limited special conditional execution instructions such as peg orders or stop orders, but it seems like a hard problem for them to support anything more sophisticated and general.

It's trivial in an era where the concept of algorithms is established but less so back in the 1940s. You really had to be at the forefront to know that sorting algorithms are meaningful things to think about. But indeed mergesort is really not even a footnote of von Neuamnn's contributions to computing, which was perhaps not even 1/3 if his career. That's how much a giant he was.

Maybe you can also tell the business owner to like his damn job of managing the company more than the sweet profits coming from underpaying senior engineers?

It seems obvious when you don't need to write things down when you don't feel you have anything to say, but growing up I was always annoyed at how essay-type homework and exams were structured, which values form over substance, because there cannot be much substance when you throw a random topic at a teenage and ask them to care in 90 minutes.

It's much easier to write better essays when you care a lot about something, really want to make it convincing and spend effort polishing your technique based on actual reader feedback. Even the "form" taught at school doesn't really matter because different audiences and topics require different arguments and require you to focus on different details, so the best kinds of essays often look nothing like say full score GRE writing test essays.

Do you know of any shop that is running deep learning profitably?

I think the social norms are such that doing that while not specifically in a socializing event is low-class. Probably it's fine between peers in a blue-collar environment, or in high school.

The calculation is the power method, but the main gist is modelling internet browsing behaviour as a Markov chain, wandering through randomly clicking links. The equilibrium probabilities of being at any page is the vector "r". I'm a bit surprised how few sources offer this explanation upfront, I guess they want to maintain the mystique, and want to make the technique appear more profound that it was.

There's a lot of progress that can be made in cross-disciplinary areas, but if it's done by one person it tends to be someone who specializes in one thing and stumble upon a low hanging fruit elsewhere. Much more often, in case of deep results, is that two very smart people in their areas talk about their work and discover an opportunity to collaborate.

Most people who claim to be polymaths are dilettantes. Even if they are really smart. People who claim they're polymaths don't have the humility to recognize it's very difficult to be at the forefront of two disciplines. A dilettante wouldn't know this because they don't know the amount of effort and learning it takes to be a true expert.

A true expert spends a lot of time on a subject and may on the surface may appear to have no more knowledge than someone who just reads and follows their papers for less than half the overall effort. This is because much of extra work is spent on trying different ideas and learning for producing the next good idea, or trying to apply it effectively in various contexts, rather than just learning other's good ideas for the sake of it.

Whether the expert is more valuable in dollar terms or potential impact than the dilettante or a humble generalist is another question. People can indeed be good at wearing multiple hats and create value. But I think someone like that should recognize that they are trading depth for breadth and not claim they are better than others at everything they learn.

Thanks, I stand corrected. I guess at 35 times latest earnings it's more reasonably valued. Nvidia's growth in just the past year was stronger than I expected. At this level we need more nuanced arguments.

The next thing to explain the discrepancy between Nvidia's valuation and OpenAI's would be that Nvidia's monopoly position effective eats into the profits of the AI startups for the foreseeable future. Had OpenAI already been profitable, its valuation would have exceeded 86B.

The catch is that there are >10 startups and big-tech teams trying to do what OpenAIs is doing, and at least a few of them appear to be competitive. One or two of them may emerge as a winner. But there's only one Nvidia.

With that said, I do believe Nvidia is overvalued - if it triples profit its PE ratio would still be 30 (i.e. 30 years to return on investment), while there's a fairly good chance someone would catch up to them in 10-20 years.