That blog is about training. For inference, the weights and kv cache are in SRAM. Having said that, the $100M number is inaccurate/meaningless. It's a niche product that doesn't have economies of scale yet.
The solution to that is to give more rights to people on visas, and yet pretty much nobody supports that. Not Ds, not Rs, not corporations, and certainly not the people who complain about H-1B.
Yes, that is still a problem, which is somewhat mitigated by the $200 fee to enter the lottery currently. The proposed change of giving higher odds to higher wages instead of a flat lottery is a step in the right direction.
Half the people complain about H-1B and the other half complain about outsourcing. Neither of them makes sense to me, but the outsourcing angle even less so. It's just labor cost arbitrage for knowledge work. And soon that arbitrage opportunity will perhaps shift from humans to llms. What is the issue with that?
That used to be a common abuse vector, but that loophole was fixed last year. Apps are deduped now (1 person, 1 app). It's still a lottery that is oversubscribed by something like 3x.
A friend of mine, originally from Iran but now a Canadian citizen, was stopped at the border during a recent visit by CBP. They interrogated him, scanned his phone, found some EE paper, insinuated that he makes bombs and is a terrorist, and sent him back to Canada. The US' treatment of Iranians is quite appalling.
This is a net positive action for the following reasons: The chuds have been clamoring for this for a long time. You can see every past thread on HN all the way back to the December blowup on twitter with Elon. At the same time, the economy is lagging and the admin's more direct measures to drum up support from the base such as chaining and deporting Koreans at the Hyundai factory are tanking future prospects for the economy and are causing diplomatic headaches. This current announcement gives the admin a way out by throwing some meat at the base before the midterms while knowing that this won't pass muster as they don't have the authority.
Instacart sent a cease and desist for trademark violation. You cannot become a middleman for random businesses/services. Kind of similar to how doordash and others got into trouble by hijacking restaurants' order flow without consent.
> different case by a different organization means this case is also political.
It's the same issue - the DOJ is going after companies and their ads. In SpaceX's case, the ads said citizen/LPR only due to export control, and DOJ got mad that it would exclude asylees and refugees for some of these positions which may not actually have export control requirements. Your complaint is also about ads and whether they are in print media or online or obfuscated etc. If you think that ads in print media violate the law, you need to prove that in a court of law. Note that the law explicitly requires ads in Sunday newspapers, whereas online ads are not mandatory. There is a check list of what is and is not required, and the lawyers are following the text (https://www.law.cornell.edu/cfr/text/20/656.17). The government doesn't think that there is positive EV in taking it to trial, and hence settles. You, as an individual, can still pursue a civil suit if you are injured.
>For all I know you might not even know what does "good faith" mean
There are thousands of bogus laws in the books and the government is not your friend. Good faith in this context means doing the minimal amount of work needed to comply with the law. Innocent until proven guilty, and the government has the burden of proof. This is how I view all interactions with the state.
> The government settles this kind of cases because of politics, not merit.
The government also files these cases in the first place because of politics, not merit. See my point about theater earlier.
>If there had been a modicum of will to go after lawbreakers, these cases would try themselves - tons of witnesses, tons of evidence zero traces of "good faith".
That's the official opinion of the government, the judiciary, and the defendants. A settlement is not admission of guilt - the opposite actually. What are we even debating here ?
> "good lawyers" cannot help here
A settlement for a pittance, as you said, is the mark of a good lawyer.
I don't see the contradiction here. The game is as follows: Company has to make a good faith recruitment effort. Not an exhaustive search, not beyond reasonable doubt. Just good faith which follows the preponderance of evidence standard. This is by design. The government doesn't believe that it can win on the merits, and hence they settle. The settlement gives everyone what they want.
larger pool + larger pie due to the growth of the economy. You are viewing it as a zero sum game. What's better ? Two jobs with a pool of 3 people, or 2 million jobs with a pool of 2.3 million ?
>The problem is not "good lawyers" but the fact that the only punishment for breaking the law is a pittance of a settlement.
That's how settlements go. The government gets to do its theater, the constituents believe that the government is fighting for them, and companies write this off as the cost of doing business.
The lack is established by a good faith recruitment process, not an exhaustive search. This is intentionally vague because it's a non-sensical requirement that is hard to prove one way or the other and was only added as a political compromise. The company is free to tailor the minimum requirements to its liking. Recall that this is a free capitalistic country. So you can establish that you can't fill the req. locally and hence are hiring a foreigner. The reason I'm pointing this out is because you have picked some type of textual or literal interpretation of things ("this is what EB is for"), and companies have lawyers who are good at following the text.
The ways to build the system you desire are simple; the political challenges though are insurmountable. This isn't rocket science. However, there has been no substantial legislative change in this area in over 30 years. The current morass of H-1B, PERM etc. is a carefully engineered compromise to keep all demanding factions - the restrictionists, the capitalists, the left and the right, acceptably (un)happy.
You're reading too much into it. It's a case of bad UX. The jobs do not exist. The actual job application/interview etc. happened years ago, when it did exist, and everyone, including locals, had the same shot at it. When the job existed, someone was hired for it, and it happened to be someone on a visa. In order to keep that person employed in this job and get them a green card, the government requires that the job be advertised again afresh. It's a non-sensical requirement that was added because some politican or lobbyist asked for it. The natural way to add protectionism to this model would have been to add it at the outset, but that clearly wouldn't work for the economy. So a compromise was engineered. Companies can hire anyone generally as long as they are, in principle, temporary. When it comes to keeping them permanently, the government requires that they do this charade of posting ads again and doing a market test etc.
Companies are always going to lay people off, because they can and it is in their financial interest to do so when the shareholders or investors demand cutbacks. AFAIK, Amazon even has a formalized stack ranking system where a certain percentage will be laid off every year to make room for new talent. This has nothing to do with visas or immigrants. If you want to stop that, you need better worker protections, but that's a separate discussion.
But they would still be higher than a chemist's salary. This has nothing to do with open borders. If you use money as a proxy, some professions will come out ahead. That's just market dynamics. The only way to avoid that is to create carve outs or normalization by profession.
That is what everyone is asking for though. As you can see from the complaints in the thread, people are angry about jobs going to immigrants in the US. The alternate is for jobs to go to immigrants abroad, i.e. offshoring. It appears that people are generally happier about the latter.
They are all roundabout ways of getting to IQ. It used to be Leetcode but since that has become a known quantity and quite gamified, the interviews have evolved. Most of all, I have seen that there is crazy focus on speed and volume. Do a lot, read a lot of code, in little time. They've also adapted for AI coding tools. So the interviews have a debugging component, or take homes are open web, open tools, but again really compressed timelines.
It is entirely relevant to the explanation. PERM and H1B are entirely different processes handled by different departments of the government.
>One way to 'ensure' this is to only advertise jobs in newspapers or in radio advertisements
That's not some nefarious plot. The government (Dept. of Labor) requires them to post these ads in newspapers. If the government didn't, they wouldn't post these asinine ads.
>That way, the American never knows the job exists
The job does not exist for all practical purposes because there is already someone working in that job. Perhaps your goal is to effect some change so that more Americans get these jobs. This is not the vector of attack though. You want to focus on the root of the problem, namely that corporations are able to hire people at will in the first place. Once they've hired someone, applying to same job is not going to make them fire that person and hire you instead. You want to take the agency away from corporations at inception, so that you can be hired.
The issue is that people opposing it don't have a coherent view of what they want. They don't want a free market solution; they don't like the socialist position where the government picks the people; they don't like the lottery either. They want their preferred type of people to get the visa in a manner not too different from quotas/dei/affirmative action etc., which is impossible to engineer. It ultimately ends up creating a compromise brokered by lobbyists and politicians, kind of like how we do everything else in the country.
Because the executive doesn't have the authority to do this. Trump also ultimately didn't do it. I don't recall if he lost the suit or withdrew it voluntarily. The new court is different and submissive, so it'll likely get through this time.
Really hard to do training hardware as the software stack is quite complex and needs to be general and flexible. Inference on the other hand is easier, which is what a lot of new hardware will be. I doubt if it makes sense for Tesla to build hardware for datacenter inference. Xai on the other hand could benefit from it.