Like high speed photography, small lenses require a lot of light, and are useless without it. I wonder how practical something like this can be? At the very least they're going to need to be paired with some pretty high tech image sensors. Regardless, it also brings into question how high the resolution could practically go.

Small, high res, low light, and fast shutter speed are all competing against each other, and you usually have to at least compromise on one (or two) to do well with the others. If you're starting out with a very small lens, there's going to be a lot of compromise going on, which will greatly limit the number of applications you can use it in.

I can see the argument that Nvidia has no obvious advantage in LLMs that would make them want to set up their own operation; it's basically everything else about the situation that would make me jumpy if I had bet on Nvidia.

"Investing $100 billion in OpenAI's spend $100 billion on Nvidia stuff initiative" sounds, at worst, like a slightly more legal version of the trick where you shuffle stock around between business units or stuff the channel and book that as sales because you suspect that your real sales numbers will disappoint; and, even if it's not quite that dire, Nvidia being willing to get paid in faith rather than in other people's money (or shift the stock to one of their customers that actually has money) looks very much like an indirect price cut, which gives the impression that either demand is outright softening, and Nvidia has units that it can't simply immediately shift to customers who are actually paying cash right now; or that Nvidia feels the need to help fill the gap between OpenAI's seemingly unlimited appetite for doubling-down money and the, sooner or later, limited supply of VC nose candy.

That said, it's not entirely novel; Nvidia's current holdings are something like 90% Coreweave(under 10% of Coreweave's total shares; but Coreweave shares are the bulk of other-company shares Nvidia holds); and they have an agreement with them obliging them to purchase any unused capacity through 2032; so they've been expressing confidence in AI-related companies and/or trying to keep the music going by paying some of their more fragile customers' bills even before this.

It could be that Nvidia isn't even trying to diversify; but the history of bad things happening when people underestimate correlated risks also doesn't make me feel great about the situation: Obviously it's going to be a bad day at Nvidia if 'AI' cools; stock price will take a hit and they will be left holding at least some inventory and TSMC and other vendor commitments; but it's going to be a worse day the more of their hardware they sold in exchange for stock in 'AI' Nvidia buyers, rather than in exchange for money, since the fortunes of those companies are going to be fairly closely correlated with Nvidia's own, albeit likely to swing harder and have further to fall.

loved that performance in Total Recall.

So this is where we're headed? Fail to pay them and they explode?

I get the unpleasant impression that we are doing a corporate reenactment of that period in European history where basically everyone was ruled by Habsburgs who were incestuous and incompetent in equal measure.

Perhaps more troublingly; they'll allow facebook to see what the people you see do.

My good-faith advice to anyone who is considering letting zuck into their refrigerator just to solve the crushing problem of what to cook with available ingredients or whatever would be "probably not worth it"; but that's ultimately a them problem one way or the other.

The trouble is that much of the pitch here is that you are supposed to provide footage as you wander around; merrily making the you problem everyone else's problem as you do so. And, yes, 'no expectation of privacy, etc, etc.' but there's a fairly obvious distinction between "in principle, it wouldn't be illegal to hire a PI to follow you around with a camera while you are in public", which involves a typically prohibitive cost in practice and "you paid them to upload geolocated footage, nice going asshole", where the economics of surveillance change pretty radically.

If people want to outource their thinking to facebook themselves I'd have to be feeling fairly paternalistic to intervene; but given that the normalization of these is, pretty explicitly, about facebook having eyes on everyone I can only hope that 'glasshole' continues to be a genuine social risk to any adopters.

Anyone who puts their money behind wildfire smoke as the leading public health threat of 2050 is just showing their abject lack of faith in the potential of malice and incompetence. Who are these faithless degenerates to tell us that we can't re-introduce enough trivially controllable infectious diseases or deregulate enough toxin smelters to outmatch some trees?

Sorry if I wasn't clear; that's the part I have deep concerns about getting done. My impression has been that(while, in theory, people are supposed to be averse to spending money) it's much easier to get funding for novel or sexy initiatives, especially if they promise to be magic-bullet solutions, than it is to push through money for boring stuff, even if it's low risk and abundantly proven; and the risk these recommendations address seems to sit firmly on the unfavorable side of that.

"We need to do a bunch of fiddly changes to eliminate quirks of build reproducibility, and generally have more eyes on important software" is not a terribly intimidating project in terms of novelty or risk; but "basically, just spend more on reasonably competent, reasonably diligent, software engineers than it seems like you strictly need to, in order to make improvements that outside observers could easily mistake for status quo, forever" is a deeply unsexy project. It's a much better project than "Agentic digital transformation" or something; but that's the sort of likely failure that someone looking to spend company money to look like a thought leader on linkedin will practically trample you in their eagerness to approve.

As best I can tell; most of the complaining about freeloaders is sideshow in the battle over who deserves subsidies, not objections in principle. I'm less clear on whether there's also a positive correlation between whining about the subsidies going to people who aren't you and actively seeking them yourself; or whether the cases of people who do both are disproportionately irksome and so appear more common than a dispassionate analysis of the numbers would reveal them to be.

It holds up in some cases, not in others, and calculating an average muddles that.

Personally, I use AI coding assists for two purposes quite successfully: a) more intelligent auto-complete and b) writing a piece of code using a common, well understood algorithm (i.e. lots of sources the AI could learn from) in the specific programming language or setup that I need.

It turns out that it is much faster and almost as reliable to have the AI do that then finding a few examples on github and stackoverflow, checking which ones are actually decent, and translating them myself.

Anything more complex than that and it starts being a coin toss. Sometimes it works, sometimes it's a waste of time. So I've stopped doing that because coding it myself is faster and the result better than babysitting an AI.

And when you need to optimize for a specific parameter - speed, memory, etc. - you can just about forget AI.

Hey, industry, I've got an idea: If you need specific, recent, skills (especially in the framework-of-the-month class), how about you train people in them?

That used to be the norm. Companies would hire apprentices, train them in the exact skills needed, then at the end hire them as proper employees. These days, though, the training part is outsourced to the education system. And that's just dumb in so many ways.

Universities should not train the flavour of the moment. Because by the time people graduate, that may have already shifted elsewhere. Universities train the basics and the thinking needed to grow into nearby fields. Yes, thinking is a skill that can be trained.

Case in point: When I was in university, there was one short course on cybersecurity. And yet that's been my profession for over two decades now. There were zero courses on AI. And yet there are whitepapers on AI with me as a co-author. And of the seven programming languages I learnt in university, I haven't used even one of them ever professionally and only one privately (C, of course. You can never go wrong learning C. If you have a university diploma in computer science and they didn't teach you C, demand your money back). Ok, if you count SQL as a programming language, it's eight and I did use that professionally a few times. But I consider none of them a waste of time. Ok, Haskell maybe. The actual skill acquired was "programming", not a particular language.

Should universities teach about AI? Yes, I think so. Should they teach how to prompt engineer for ChatGPT 4? Totally not. That'll be obsolete before they even graduate.

So if your company needs people who have a specific AI-related skill (like prompt engineering) and know a specific AI tool or model - find them or train them. Don't demand that other people train them for you.

FFS, we complain about freeloaders everywhere, but the industry has become a cesspool of freeloaders these days.

The number of people isn't the problem here.

The "started every" is.

How did they not catch that during development and found a solution? I mean, the meme's where a TV ad starts Alexa and orders 10 large pizzas are a decade old now.

This seems like the sort of advice that is going to be exceptionally hard to get followed because it's mostly so dull.

There can be some interesting futzing in principle to keep unnecessary sources of variation from getting folded into build artifacts, normally followed by less-interesting making of those change in practice across a zillion projects; and basically anything involving signing should at least be carefully copying the homework of proper heavyweight cryptographers; but most of the advice is of the "fix your shit" and "yes, actually, have 10 people, ideally across multiple orgs, despite the fact that you can get it for free by pretending that the random person in Nebraska won't make mistakes, get coopted by an intelligence agency, quit to find a hobby that doesn't involve getting yelled at on the internet for no money, or die" flavor; which is absolutely stuff you should do; but the sort of deeply unsexy spadework that doesn't have magic bullet vendors lobbying for it to get paid for.

What seems sort of damning is that the explanation is "our tech sucks".

The 'explanation' is that the demo triggered all the devices within earshot because apparently a device designed to perform possibly-sensitive actions on your behalf was assigned a model line wide, public audio trigger in order to make it feel more 'natural' or something; rather than some prosaic but functional solution like a trigger button/capacitive touch point/whatever; and that the device just silently fails stupid, no even informative feedback, in the even of server unresponsiveness or network issues. Both of these seem...less than totally fine...for something explicitly marketed for public use in crowded environments on what we euphemistically refer to as 'edge' network connectivity.

You obviously have limited control over the network in a situation like this; so nobody expects the goggles to fix the internet or facebook's server resource allocations for you; but having some sort of "can't reach remote system" error condition has been ubiquitous basic function since around the time that dirt was still in closed beta.

I suspect that this is symptomatic of the same phenomenon; but it seems especially weird that they'd be trotting the CTO out to give a, from context, apparently intended to be exculpatory postmortem when the problems with a device you are intended to wear on your face, in public, are 'sensitive to external trigger shared across entire product line' and 'silently fails stupid if network conditions are suboptimal'.

You definitely wouldn't come up with a fresh 10,000 liters of the stuff just lying around somewhere; at least not without resorting to nuclear chemistry or natural gas processing on a pretty heroic scale; but if you purely needed to ship something, anything, to be able to say that the amount provided wasn't zero; the terrestrial supply isn't zero either. I think the US is good for high single thousands of liters on a typical year, from nuclear warhead maintenance; Russia at least theoretically in the same ballpark in terms of warheads that would need their tritium checked, though no assurances either that that is happening or that they'll sell, they had formally stopped doing so at least for a while over a decade back; not sure what the mixture of reasons was between domestic users and not wanting inferences about their weapon maintenance.

Such a sale would be basically ceremonial if it has to come from the existing supply which is already spoken for every year; and there would be no point in Interlune as an intermediary; but if some finance construct wiggles one way if the sales are zero and another way if the sales are merely small, it presumably might be worth someone's time for Interlune to be listed as the supplier to Bluefors, even if it's just them slapping their label over whoever Bluefors normally buys from and doesn't actually change the allocation to different purposes or the total size of the market.

It's adjusting the allocation that would be at least difficult(potentially viable if the VCs doing 'quantum' are paying better than the people doing ultra low temperature MRIs or academic physics, or if you can out-lobby the 'national security' neutron detector market that doesn't get anyone excited but zOMG Dirty Bombs the Homeland!; but probably not cheap); and actually changing the supply that would be hardest, but possibly of actual interest.