It's definitely interesting that Nvidia thinks RISC-V is big enough to be worth the port; but describing the CPU as 'central' to Nvidia's preferred design is deeply overselling it. The recommended layout is basically a bunch of GPUs chatting with one another over NVLink within the chassis; and using GPUDirect RDMA on Nvidia infiniband cards located on the same PCIe switch that the GPUs are for scaleout; with Nvidia ethernet DPUs handling the remaining high speed networking; and the CPU doing housekeeping.

Given that porting and maintaining on another ISA isn't free the fact that Nvidia bothered is certainly a vote of confidence is at least middling RISC-V options actually being attractive to enough potential buyers to be worth it; but the CPU is not intended to be a major player in a CUDA-oriented system, especially one of the larger ones.

Even if we take the claims about the quality of 'AI' tools at face value; it seems fundamentally contradictory to talk about the situation as though they can just keep making engineers more productive.

The exact multiple of the baseline 1x engineer isn't entirely clear; but at some point the ability of a human to act on outputs and provide further prompts in response is reached. Any further improvements in productivity would then have to come from they system being reliable enough that much of its output simply doesn't need a close look and it can complete much larger chunks without being talked through it repeatedly. That's certainly not inconceivable; but if the tool is reaching those levels of semi-autonomy the need for a hotshot engineer is now declining because you can mostly get away with just letting the bot do its thing.

It seems sort of like saying that machine tools and heavy industrial equipment will produce "100x blacksmiths". If you just look at nail output per person per hour it's probably more like 100,000x blacksmiths; but the secret to producing more nails was not better power hammers; it was the fact that we don't really use blacksmiths to produce nails anymore.

I wanted to like GATTACA; but it really ended up being almost the opposite of what it professed to be about(admittedly, in a way that seems very much like something we would do): the premise is supposedly that genetic engineering has allowed all men to not be created equal, and there are some lightweight examples of that being true(none of the engineered characters require corrective lenses; some of the naturals are identifiable by being contact lens users); but in important areas it just never actually seems to matter; the main character is supposed to have a serious cardiac condition that, um, never actually stops him from out-swimming his modified brother through the power of the human spirit or conducting astronaut training as long as he plays back the cardiac data from the suicidal athlete whose identity he assumed(apparently genetic optimization doesn't extend to psych coverage?)

And the society itself seems to realize that at some level: theoretically the premise is deeply troubling because what of a society where people are profoundly unequal by birth; but in implementation there's precious little sign of 'meritocratic' squeeze-out (there's the one woman in the astronaut training program who isn't going to make the cut because she uses glasses; but is still in the training program for some reason?): just people getting hired based on a genetic test and the ongoing battle by the protagonist to carefully cover all traces of his DNA with replacements from the guy he is impersonating because his job performance won't actually matter if the geneticops enforcing the haves/have-nots distinctions aren't doing performance based evaluations.

As noted; using relatively weak scientific evidence to justify treating people with wealthy parents as though they are better than the rest is absolutely something we would do; we'd love a blood test for legacy admits; but as a movie about genetic engineering making people unequal, rather than a movie about people running a caste system loosely justified by genetic engineering, it really doesn't go much of anywhere.

I suspect that the "I made a lot of money selling enterprise resource data munging cloud solutions or something; therefore I am very smart" crowd is more susceptible than average to the specific "Now, I'm not going to say 'eugenics'; but clearly you want your children to be just plain better, like you..." sales pitch; but the genre of hope-and-hype 'this will make life better for your child' seems to have more or less universal appeal. The upmarket techie flavor leans a bit GATTACA, and probably has a lot more literature that is the layman's idea of what good scientific papers look like; but it's basically the same impulse that has people taking prenatal wellness supplements of deeply dubious efficacy and experimenting with prenatal classical music and stuff.

I'm not particularly inclined to be trusting of someone squeezing PCR until it bleeds and then shoving what comes out into their proprietary risk model until they've shown some pretty solid results; but it's not like 'prospective parents who want healthy babies' are a weird niche audience. Especially if it's an IVF-related intervention; which presumably means that most of your customers are coming in the door with some combination of fertility issues, an atypically high number of miscarriages or stillbirths, or one of the relatively well characterized and dire heritable conditions that are closer to "25%, maybe 50-50 in males if it's X-linked" than to "well, meta-analysis suggests that these 853 genes can nudge the risk of the autism by an amount that was technically statistically significant".

"Right now, at $2,500 per embryo-screening on top of the average $20,000 for a single cycle of IVF, Siddiqui's social network in Silicon Valley and other tech hubs is an ideal target market..."

Selling upmarket sounds like it's more about the lack of validation than about the actual cost. If you actually knew that it worked a one time $2,500 to mitigate, or even moderately depress the incidence on average, of "bipolar disorder, cancer, Alzheimer's disease, obesity and schizophrenia" would be stupid cheap. If you spread appointments out and are mostly on older drugs with generics $2,500 might cover a year of outpatient psychiatry, ignoring the costs and unpleasantness of having a psych condition to deal with; and if there's a cancer you can deal with for at or under that amount it'd probably be a nice clean presentation of an early stage skin cancer that the dermatologist can deal with; definitely not one that gets punted to the oncologists.

It's just that if you are currently at the Theranos "we can totally do stuff with PCR that stuffy 'scientists' say is statistically troublesome"; then we grind it up with our proprietary models, you are basically left selling to the anxiety of the cost-insensitive. If the method were actually validated not paying ~10%more to have your IVF round QAed would look like reckless negligence; but that 'if' is very load bearing.

Allegedly this was a permitted practice; but the speed with which they said that they will be abandoning it once it became public knowledge; and the number of federal IT people ProPublica was able to find who had never heard of it, suggests that either the proposal that was approved was not entirely candid about what the plan was; or the approver was too low or obscure to actually approve.

This certainly wouldn't be the first time that something perfectly on the up and up was abandoned for PR reasons; but MS would probably be loathe to give up the ability to whitewash whoever into sensitive projects by having an $18/hr copy-paste pal in the loop; so they must see the exposure as potentially serious.

It would be interesting, and possibly useful, to know how these reports break down in terms of affiliation and motivation.

It's obviously a problem regardless; but, in terms of behavioral change, it seems likely that the well meaning but confused would have different incentives than someone taking advantage of the speed with which a bad bug report can be automated to spam everyone who has a bounty program of some kind in the hopes of getting lucky; someone in over their head and attempting to farm cred as a 'security researcher' would be sort of a hybrid of the previous two; and someone using OSS projects as guinea pigs for some 'AI securifies your code!' startup's training/hype process would be basically the worst case scenario; since they don't need to be motivated by the idea that they are being helpful and their financial incentives are separate from any bounty program.

So, TFA says "Today, CAISO engineers scan outage reports for keywords about maintenance that's planned or in the works, read through the notes, and then load each item into the grid software system to run calculations on how a downed line or transformer might affect power supply."

Sounds like there is already an actual piece of simulation software in place; but the people feeding it scenarios need to manually assemble them from outage reports and maintenance notes; because apparently you can perform, even schedule, maintenance without generating a record of which asset you'll be working on. And so someone proposes that the obvious answer is to bandaid that with a summarizer, and their notorious penchant for accuracy?

Does that strike anyone else as slightly insane? We're already recording outages and maintenance events and those records refer to various pieces of infrastructure; but not in a format that can be programmatically extracted from the report to load into the analysis system; so rather than doing the boring thing and slapping serial numbers on transformers or adding a "things we turned off for maintenance and when" field to the maintenance report; we'll just continue doing that as a big free text blob and hope that an LLM will save us?

Seems sort of like deciding that postal addresses and mailing labels would be a huge pain; so let's find a chatbot that can, hopefully, follow a rambling account of how mail for the Johnson place is 3rd right past the old rail bridge; but if you see the silo you've gone too far.

Even if we assume, for sake of argument, that trends are favorable and someone can actually figure out how to cover their costs; why do you say that "the train is leaving the station on this"?

So far the LLM guys have appeared to be relatively low moat; with significantly cheaper and only modestly worse competition not far behind the big names and ongoing contention between the ones you've heard of.

If Google, whose main obstacle to dominance of the search market is its own self destructive tendencies; with microsoft a distant second, is paying Apple 20 billion a year to stay default and everyone who isn't them is basically begging you to switch to them, why would we expect that the LLM guys would be in a position to squeeze Apple on terms?

Where is the point where Apple can no longer board the metaphorical train for some reason? There may come a point where they can no longer simply ignore the matter; if Siri's ongoing mediocrity becomes an actual issue or the like; but I'm not seeing why they'd be in a notably worse position tomorrow than they would be today to either get satisfactory terms from one or more of the largely interchangeable and margin-challenged competitors; or just snap up one of the more competent fast-followers(or an open weights model and some appropriate hires) and deal with that when it comes up.

Where's the urgency? What marks the end of the station platform in this increasingly tenuous metaphor?

Even if we, for sake of argument, accept the theory that Apple products are suffering for want of 'AI'; what's the case for Apple to pay a fairly stiff premium to acquire vs. just taking advantage of the current state of the market; where there are multiple people who will fight for the opportunity to lose money on every sale and attempt to make it up in volume?

Absent a fairly specific argument why Apple needs to own one; rather than just snapping up some of the useful people or commanding the influence that being a customer with actual money tends to provide over suppliers that are in the process of bleeding out, the idea that Apple needs to buy into 'AI' seems sort of like the idea that Apple needs to buy into DRAM, except that it's vastly more evident that Apple's products actually need DRAM; and Apple still doesn't go there because why bother with a capital intensive business whose margins are constantly buffeted by spot prices and on the thin side when you are buying enough that you don't need to worry about your place in line?

It's not clear that that's an either-or. A financial sector metastasizing into the real economy is going to have many of the same symptoms.

This seems more like a "Zuckerberg threatens hundreds of billions for AI datacenters" situation.

In all seriousness; even if you are an 'AI' optimist(perhaps especially so, since you presumably think that this isn't just Zuck pissing away more money after his metaverse successes); would you want Facebook to have a commanding position in the area? It's not literally the worst possible company to potentially have to deal with; but it tries.

Given the level of commitment it implies; basically the most lightweight of expendable pilot programs even if you are saying that you 'plan to' in a legally binding context; is seems at best exceptionally dubious to treat the answers to "do you plan to adopt generative AI?" as straightforwardly meaningful.

The differences mean something; it's just not obvious to what degree they reflect actual company strategy, vs. personal fascination with the new shiny thing, vs. people saying what they think the audience wishes to hear.

It's sort of an interesting mix of goofy hype and actual(but relatively boring) worth-looking-into.

Not so much because of 'quantum' necessarily; it's entirely possible that someone will get an at least somewhat worrisome classical efficiency improvement worked out before the quantum computing types reach anything of useful size; and it's probably worth betting money that particular cryptographic implementations will turn out to be flawed; but because it takes a fair amount of awareness to even have a complete idea of what you are running; and more than that to know the implications of needing to swap it out in some or all locations.

The people selling 'quantum' and 'post-quantum security' are mostly in the business of "forget your boring arduous problems by focusing on our exciting ones!"(good business; bad way to do security); but it's a pretty solid idea to be aware of the boring arduous problem of exactly what ciphers you use, and what implementations, and whether there are any places where you've inadvertently left a compatibility toggle that allows something to be downgraded to some 90s 'export grade' cipher; and have an idea of how hard it would be to change ciphers or update implementations if you needed to for one reason or another.

Shockingly enough, the people with the biggest marketing blitzes and best 'executive whitepapers' with stock photos of shadowed hoodie hackers and chinese quantum AI owning your cyber are not the ones mostly advising that you should do some really boring systems administration and SBoM stuff while waiting for mature industry-standard implementations to become available; so the people selling immature proprietary implementations and dubious silver bullets tend to out-shout the more sensible ones.

Hopefully the NTSB's tier 2 people are ready; since they did in fact try turning it off and turning it on again and there's still more to the situation.