It seems almost meaningless to talk 'eventual' uptake when the relevant hardware has been added to more or less all current laptop CPUs, without there being NPU-less variants that cost less, so basically anyone who buys a new laptop in the future will, technically, be 'adopting an AI PC'.

I'm not sure that they'll end up going after Valve to get the results; but it's arguably the case that we've already seen the (really annoying) adaptation:

"Older games" certainly includes various well loved singleplayer titles or 'classic' multiplayer ones that don't kick back to the mothership anymore; but it also includes 'live service' and not-technically-live-service-but-look-at-all-those-skins-and-battle-passes titles that happen to have original release dates from some time ago.

Something like DOTA2 is a decade old; and based on ongoing cosmetics sales I'm doubting that anyone wants that 'old' game taken out and shot, at least not before DOTA3 is fully ready to go. This wouldn't show in the Steam numbers necessarily; but Fortnite is a mid 2017 release; and also still spinning money.

There has been a lot of publisher movement toward games that are expected to be long-running; but designed to continue paying out. That's not all of them; and there are probably a few back catalog titles whose longevity is cursed in the boardroom; but publisher strategies already seem to have substantially shifted to trying to cultivate and monetize ongoing play vs. steady purchase and relatively quick abandonment of retail releases.

I'm sure that Google can do whatever it does cheaper if it fires some people; but it feels weird for a company as erratic and rudderless as they are to be talking about 'efficiency'. Maybe don't develop and then cancel products apparently at random if you want to feel more efficient; rather than immediately moving to develop and then cancel products apparently at random but with 10-30% fewer staff?

This might just be a reflection of working with a limited toolset, or of it being more politically palatable to suggest that the PRC is up to something than to suggest that there's something glorious private sector is failing to do; but focusing on one bottom-feeding vendor seems like a really stupid choice if you actually care about router security.

At the level of consumer-tier routers there aren't actually just tons of SoC vendors (Broadcomm, Mediatek and Qualcomm seem to be the big ones that do fully integrated; a few more options if you are doing something wired only or using external wifi support; but mostly other ARM SoC slingers, I think Marvell comes up here, without the complication of worrying about GPUs, until you get up to Atom territory); and a lot of the vulnerabilities are in someone's lousy web GUI, an antique userspace component, or some dodgy TR-069 implementation. There are occasionally driver bugs that are security risks(most commonly to wifi devices, exploitable within RF range, less commonly on wired interfaces in a way that would be exploitable from the internet) but those are generally less common and higher skill than some shovelware web UI not validating an input.

I'm not sure you could make a TP-link really suitable for DoD work just by reflashing it, if you are really that valuable or prominent you can't rule out some sort of clever tricks implemented in a supposedly commodity eMMC part that's actually much higher capacity and rather smarter than advertised; or NIC firmware that listens for certain magic packets or something; but the bulk of routers-are-shit problems in lower performance/price sensitive areas could be addressed by treating it as unacceptable that you have to get your fairly basic embedded ARM linux package from the same vendor who designed the logo on the plastic box it came in.

Specifically for Fed and other high-assurance purposes they should probably just stump up and have some more heavily scrutinized and better trusted vendor slap SoCs from one of the vendors with enough domestic presence to be leaned on together: it'd be more expensive than the offshore low bidder; but probably cheaper(as well as helping to develop and maintain high-trust capacity) than taking the offshore low bidder's hardware and inspecting it to your satisfaction; but in general just demanding that routers be documented well enough to allow 3rd party OS builds from someone who is actually competent at that would go a long way to solve the problem. We've seen(with openWRT and others); that what's basically a hobbyist thing can outperform the average vendor shovelware; if you want a fancy enterprise contract there'd presumably be nothing preventing Cisco and Palo Alto and Fortinet and the like from either offering warrantied images for compatibility-baseline hardware targets or compatibility-baseline compatible hardware targets that meet the quality specs they are willing to warranty; and the feds can always pay BAE or General Dynamics Mission Systems to cook up an embedded linux image that meets all the NSA's recommendations for SELinux-ing correctly.

As long as you pretend that it's some sort of rule of the universe that you get your OS from your hardware OEM(or, more likely, a reskinned version of the chipset BSP that their ODM mutilated into shape for them); you are not going to solve the problem of software quality being fucking dire; while if you abstract the software even just a little bit you open the possibility of having much higher standards; because while good software requires a considerable number of hours from talented people to write and test, stamping out copies costs nothing thereafter.

What is frustrating is that, for want of some relatively lightweight standardization(lots of weird little boot and flash layout quirks between vendors even on the same SoC); that decoupling is only for nerds, on select models they can work with, or hyperscalers who buy in sufficient quantities that they can just tell OEMs how it is going to be(eg. Microsoft getting sufficiently fed up with assorted L3 switch management OSes that they just decided that you'd need to run linux to qualify.)

I'm not proposing some nonsense of the "everyone in the homeland must run JucheWRT for security!" flavor; just that it's lunacy to expect the software to get better when it's treated as an intrinsic part of the bundle with the hardware and mostly sold on hardware price to people wholly unequipped to evaluate its quality(at least until you get to the 'business/Enteprise' vendors; whose software quality ought not necessarily to be trusted, just look at basically every SSL VPN gateway provider over the last 18 months or so; but who at least acknowledge software as a major point of differentiation on which their customers are judging them).

No disagreement on Intel's apparent underuse; but(at least in theory) they've been open for business since they started talking up "Intel Foundry Services"(and possibly by less public special arrangement prior to that; but that was the branding for their big "we want to be a credible merchant foundry that also does a lot of work for Intel" push).

They do apparently have some foundry or advanced packaging customers, and they did some internal P&L rearrangement to allow for better accounting of the Intel design side having the intel foundry side do work for them(not sure how much they just mashed together and how much was held up by some nightmarish web of interdepartmental cross-charging prior); so, while I don't get the impression that they've set the world on fire, it's definitely not a case where spinning off would be required to do 3rd party work.

My understanding is that, for anyone playing Scrabble at a serious competitive level, the fact that the rules of valid tile placement were derived from natural language is essentially irrelevant. The list of valid words is chopped up according to its utility(point values, odds of particular letters appearing in your tile draws, likely size of openings encouraged by the special value squares on the board, etc.) and handled accordingly.

Using words is certainly the overwhelmingly sane choice: the game would pointlessly unapproachable if it just defined a set of arbitrary strings as valid when anyone vaguely literate in a language with an alphabet knows at least hundreds and can at least plausibly guess about whether that's a word or not for a couple thousand; but it's fundamentally a game unrelated to words or language in anything but a sense of convenience.

No idea what the breakdown is between high level players who use memorization techniques that would also work on arbitrary strings and ones that use natural language word construction rules plus memorization of exceptions to the rules; but they are just strings so far as Scrabble is concerned.

As best I can tell a lot of 'crypto'-adjacent activity basically boils down to drumming up a supply of greater fools because you can't cash out without one.

I assume that some of the people who bang on and on and on about 'onboarding to crypto' genuinely believe that building a UX on top of copy/pasting wallet addresses that isn't shit will lead to a more efficiently dystopian future or something; but the overriding incentive tends to boil down to 'onboarding' being someone exchanging money that's actually worth having for some scrip that would otherwise be highly liquid only among scrip enthusiasts looking to trade it for differently serialized scrip.

NFTs and 'web3', especially the stuff that most overtly and crassly seeks to emulate either real estate or awful pay to win mobile gaming, is the lowbrow consumer face of the effort; various securitization arrangements are the one for consumers who fancy themselves a bit upmarket to just gamble on their phone rather than their 401k; and 'strategic reserve' appears to be the same incentive repackaged for potential government marks.

I'm sure a CFO has some sort of elaborate M&A plan in mind; or at least an idea of who to call in to cook one up; but I'm having trouble seeing the endgame here:

If Intel's 18A process is unsuccessful then they'll be trying to sell a bunch of 10nm and older fabs that have "INTEL HAD TO GO TO TSMC TO STAY COMPETITIVE" spraypainted on them. Certainly not worthless; it's not like 10 and 14nm logic silicon is super old and busted(and there's still an awful lot done on considerably larger processes); but who is the buyer who would value them more than Intel does; and what customers are they going to be fabbing for that Intel's 'foundry services' attempts wouldn't be able to reach?

Is the hypothesis that the fabs are internally mismanaged in some way that would be decisively improved by whatever private equity outfit is feeling flush on a given day? Is it purely a matter of a finance guy seeing an opportunity to increase the amount of finance Intel does in the future, even if that's unrelated to how long that future is?

One should certainly be concerned about keyloggers and the various credential(and sometimes browser history) scraping and exfiltration tools; but Recall makes the situation worse in the sense that malware can only really work with either something that was captured by something else and stored prior to its introduction or with things that wouldn't ordinarily be captured during the time between being planted and being discovered.

Adding a vendor-blessed mechanism for fairly comprehensive history to be collected and stored(and in a format that's intended to be actively convenient to query if you have very limited exfiltration resources; or which can be plundered wholesale if you don't expect the network anomaly to be picked up) is raising the stakes pretty sharply on what losing control of the system, even briefly, at some point in the future will end up setting you back.

I'd be a lot more sympathetic to the question of how to accurately filter out sensitive numbers; except that Microsoft is quite eager to upsell you on glorious "Microsoft Purview" data loss prevention technology that allegedly:

"DLP detects sensitive items by using deep content analysis, not by just a simple text scan. Content is analyzed:

For primary data matches to keywords
By the evaluation of regular expressions
By internal function validation
By secondary data matches that are in proximity to the primary data match
DLP also uses machine learning algorithms and other methods to detect content that matches your DLP policies"

Either they are fairly shamelessly selling something that is largely unfit for purpose to Purview customers, or they specifically neglected to bring actually-functional heuristics over to 'Recall', or the actually-functional heuristics didn't fit in the resource constraints of the Recall environment and they are overselling whatever simplification remains.

None of those options seem like good looks.

Remember when tech people talking about the future were able to make it sound like a promise rather than a threat?

The question of whether Intel has 'anything' seems like a testament either to how much slack TSMC had or how hard to work with the Intel foundry guys are.

Logic silicon isn't the only game in town; it's a somewhat different list for analog and mixed signal, DRAM, NAND, and optoelectronics(especially non-networking ones, like camera sensors); but for most of thing things Intel actually does their competitors are fabless outfits.

Apparently Intel can't actually avoid losing money while AMD and Nvidia are busy eating most of the highest margin datacenter sockets; but I'd be curious about how much actual marketshare Intel can lose before life really starts to suck for everyone who isn't a particularly high margin not-intel entity. Nvidia can presumably pay more or less whatever it takes to get an H200 out the door; while Apple and AMD are paying less per die but mostly buying smaller dies; but what about all the guys licensing some random A72 cores from ARM to build totally undistinguished application processors?

I'm not sure if Intel will be well-positioned to take advantage of this, they seem to have been talking about, rather than doing, 'foundry services' for something like a decade now; and they tend to get bored with things that aren't a suitable combination of high margin and high volume; but unless there's a lot more slack in the foundry market than one would expect it seems like Intel having a bad day will end up going poorly for basically all their less-prestigious competitors; since Nvidia, AMD, Apple, and even cetain intel parts(Lunar Lake and the Arc GPUs) are all reliant on the most desirable TSMC nodes; while most of Intel's historical bread and butter parts were manufactured internally on fabs that didn't do a lot of outside work.

The trouble in comparison to the Cintiq pro is that is rapidly goes from "you still need to buy a PC to plug it in to" to "you can plug it into a different PC".

The MS unit had a stupidly nice screen for the money; but it's permanently tied to a quad core mobile CPU from 2021, a 6GB 3060(the weaker laptop variant), and a maximum of 32 GB of RAM. There is no provision for running the system in some sort of 'monitor mode' that takes video input and pipes touch and stylus events over USB to something else. Wouldn't surprise me if someone with steady hands and steady nerves could do some surgery; it's probably just eDP and either USB or i2c in there; but not something they bothered supporting.

That's not yet an unusable computer by any means(thought the 16GB base model with a 2017-era i7-7820HQ is likely getting to the point where you actively have to baby it to avoid noticing slowdowns for even fairly lightweight purposes and a mobile GTX 1060 is getting a lot closer to iGPU class than it used to be); but it is not aging gracefully. Good news for the receptionist who will end up getting a really nice screen because they've got a highly visible desk and relatively lightweight computing requirements; but whoever is supplying the graphics department will be starting over from zero rather than just swapping out whatever is currently attached to that cintiq(if they even need to do so at this point; since they probably didn't choose fairly lightweight laptop parts for the job).

I assume that ifixit has relatively little leverage when it comes to getting manufacturer-blessed parts; but it seems awfully pointless to sell $450 motherboard replacements for a $400(with controller) console.

Some of their other offerings are actually worthwhile; but this looks like nothing more than PR work for Microsoft.

Sorry, those are model numbers that correspond to 'series' of Epyc server CPUs.

The 7001 series was "Epyc Rome", based on the zen microarchitecture; started releasing mid-2017. It was followed by the 7002 series (Rome) based on Zen2, in mid 2019, then the 7003 (Milan) Zen3 parts. 4th and 5th gen got a little more complicated; with 4th gen including Genoa, Bergamo, and Siena; with both Zen4 and Zen4c higher core count parts and Sienna's smaller socket and reduced memory controller for smaller systems. 5th gen is the 9005 series; with 'Turin' based on Zen5 cores; and 'Turin Dense' based on the smaller Zen5c cores for higher core counts per socket.

The time period covered is roughly the same as that of Skylake through Emerald Rapids on the Xeon side of things.