That's less than the federal grants a single university (UNC) got in 2022. [begin sarcasm] I'm assuming the salaries offered to relocating researchers are going to be in the €200K range at a 32% income tax rate with a regulatory environment that is going to allow them to do research unimpeded by overbearing regulations and abundant 2800 sqft family homes near major research centers in the €700K range? [end sarcasm]

There is more to attracting talent than a token contribution to a communal pool. Europe has lower wages, higher income taxes, higher property prices, and onerous regulations. There was a time when the ancillary benefits of Europe exceeded those downsides, but those times are in the past. Regardless of what misguided nonsense our tariff-in-chief pushes through in his imaginary state-of-emergency, I have no concern about a wholesale brain drain from the US.

I legitimately feel for researchers whose funding got cut. It’s undoubtedly a very unpleasant situation for them. I'm sure they have families to feed and this is going to be hard on them. However, the majority of the grants that were cut were not for STEM or medical research, contrary to the narrative pushed by detractors. Researchers whose funding got cut have options, they can seek out private backers to fund their research assuming there's value to be created from it. If Europe wants to bring in swaths of displaced social science researchers, that's absolutely their prerogative, but the cold hard truth is that the net impact on the US economy from this will be imperceptible.

It's a virtual paradise, barely 5000 crimes a month and only 10% involve physical violence!!!

https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fwww.civichub.us%2Fca%2Fsan...

Last time I was in the city, a guy took a dump in front of the window of the coffee shop I was in, but that was performance art, I just misunderstood it at the time.

That makes a whole lot more sense, so it's not actually 1-bit model after all.

I want to know how a single bit, which, by design, can ***ONLY*** be 0 or 1, manages to add a sign bit without actually adding a sign bit. Did bits grow a 3rd state when I wasn't paying attention?

Automattic will have a hard time hiring competent engineers going forward. I'd bet they're on a whole bunch of "do not apply" lists. Although, if you want to have some fun, respond to a recruiter for them with a watermarked response along the lines of "no way in hell".

That show was so good. If you haven't seen it, see it. Back on topic, after watching that clip, read this:

https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fpublichealth.jhu.edu%2F2...

The show set Patient Zero in Indonesia. If they had set it 5 miles north in Singapore, it would have been eerily prescient.

We currently don't have good anti-fungals for these adapted fungi, but I don't see a path to us becoming "clickers". That said, interesting fact, we're actually closely related to fungi, both being eukaryotes, rather than bacteria (prokaryotes) and viruses (acellular). The primary weapon we have against fungal cells hinges on the fact that fungal cellular walls are comprised of chitin rather than proteins. Currently, fungal infections have a ~30% fatality rate under ideal conditions, but if that line of defense falls due to a mutation or mass infections, the Last of Us scenario would move out of the realm of fiction, minus the "clickers", as 80-90% of humanity becomes mushroom fodder.

The C++ macros mess is why I swore off it. I like the Google C++ style of "better C", but templates + macros make code nearly impossible to introspect from source. I've actually moved back over the past few years to languages with fewer abstractions like Go, C, and Zig. I used to like Swift, but cargo-culters have started adding debug / performance-hindering abstractions to it and I see the opacity only getting worse.

I really don't get the "let AI write code for you" and humans can debug, improve, and maintain it. It's completely backwards from what it should be. AI should first be used to write unit tests and run automated tests. So long as the tests are mostly right, that's a big win and there's little potential downside in production. Next, AI should move on to identifying and fixing bugs. That's the natural progression from writing tests. Once AI is good at writing tests and fixing bugs, then you let it take a crack at refactoring and improving existing code. Finally, when AI can write tests, fix bugs, and improve existing code, you let it write Greenfield code, maybe. Slapping together a prototype, generally speaking, is the easiest part of the coding. It's absolutely not the most time-consuming (read: expensive) part of software engineering. The expensive parts are maintenance, debugging, tuning, hardening, and making code testable. Right now, the focus is on "look AI can spit out a prototype it saw on StackOverflow or GitHub!!!". That's the wrong place to start. If you remove the programmers from the initial coding process, it will take orders of magnitude more time and effort to figure out how the code is structured, where the code paths are, and how to improve performance characteristics.

> All of these jobs require more brawn then brain.

I can't tell if you're joking, but Electrician, Plumber, and Nurse Assistant all require on-the-job training, and you have to pass license certification in the US and Canada. I'd hardly call those "brainless" jobs. Do you want an unlicensed electrician to wire your home? How about an unlicensed nurse assistant monitoring your vitals after a heart attack? I've helped a plumber replumb a house I owned; it looks simple, but if you don't get the slopes and runs right, very bad things occur after a little while.

Wish I had mod points, this is definitely informative. I knew private clinics existed in Quebec, but I had no idea Ontario forbade it.

RAM is a minor part of the interference performance equation, no matter how optimal the engine implementation, memory I/O bandwidth is the primary limiting factor for inference performance. Nvidia SoCs don't have the fastest CPUs, but their I/O bandwidth is off the charts with 3.35TB/s for H100, 8TB/s for the B100, 16TB/s for the GB200. For comparison, the fastest Mac M-series SoC is the M3 Ultra with 800GB/s. Inference is about walking memory, I'm sure you can run a lobotomized variant of R1 on a Pi, but you pay in breadth, accuracy and hallucinations. The G4 has memory bandwidth of 2.7GB/s, so with a small enough model, it will "run", a 1000 parameter model would run fine, it would just be useless. Call me when they get a non-quantized 7B parameter model running on a G4. Hint, they won't because the hardware memory bandwidth makes it impossible, even with a hand tuned assembly inference engine, it's just not possible to walk memory fast enough.

They can legally work here if they’re among the "best and brightest" engineers globally; that's the stated purpose of O1 visas. O1 visas, unlike H1B visas, have no caps or lotteries. An officer reviews applications and determines eligibility based on documentation of the applicant’s achievements from their peers.

If they're not currently considered among the best and brightest, then it's trickier. The best bet is to obtain enough credentials to qualify for an O1. Keep working and upskilling until they’ve obtained documented peer recognition. Participate in popular open source projects, publish papers, write books, win coding/math olympiads, present at conferences, apply for awards, get their name known.

If none of that is possible, then they likely only have the H1B option. That's when things turn less copacetic; at least in 2025, you will be displacing similarly qualified Americans, and people tend to get upset when that happens.

Cost-plus bidding is basically a license to waste money. When the customer (ie DoD) is on the hook for all costs, what incentive is there to control costs? Even worse, the contractor determines the costs, so almost anything can get nicely disguised as a "cost". Government contracts should be fixed price. You agree on a set of deliverables, and you get paid when you deliver, or you get half up-front, and half on delivery. Otherwise, I could bid $1 on cost-plus contracts and bill for "expenses" like mad to make my profit. Simply form a new company and subcontract to them with a nice 2-3x margin.

$326M dollars for a for an HR system? Are they delivering hardware, an operating system and compiler for it too? I'm sure the systems at DoD are decrepit, but there's no logical way this is a $316M undertaking.

(Puts on sarcasm cap)

Of course, they could have Cursor AI code gen the whole thing in 45 seconds for $3.26 in tokens and get a same day 100,000,000x cost savings.

(Takes off sarcasm cap)

Now, I'm sober enough to know you don't build defense systems on Node and npms, but even the original $36M seems high to me. This should be a 10 person team over 3 years. If each person got paid $1M/yr, there'd still be money left over for a 20% margin. The original RFP was a 1 year deliverable. I don't know that I'd commit to a 1 year timeline, even with a gun to my head, that's basically guaranteed to fail. Ignoring nonsense of people knocking proof-of-concepts in a weekend, this feels like a 3 year project. I would expect 6 months to define and document, 24 months of development time, and a 6 month integration period. Once the RFP is signed off at the end of 6-months, that's when development starts. I don't know if COTs was an option here, and if it was, then $36M was an even bigger rip off.

Like it or not, Intel is the closest thing to a TSMC we have in the US. Having been on the board of Intel previously is a concern, it would have been more reassuring to get a clean start from someone outside Intel's inner circle. Hopefully he understands that Intel can't count on the Intel + Microsoft duopoly anymore. Tan needs to push Intel firmly into the ARM game, no matter how many internal egos or deniers get hurt from putting x86 on life support.

To all those who think they're being extra cautious by merging onto a 70MPH highway at 40MPH, I think he's talking to you.